Evaluating Ecosystem Service Trade-Offs and Recovery Dynamics in Response to Urban Expansion: Implications for Sustainable Management Strategies

Abstract

Land use land cover (LULC) changes due to rapid urbanization pose critical challenges to sustainable development, particularly in arid and semi-arid regions like Saudi Arabia, where cities such as Abha are experiencing unprecedented expansion. Urban sprawl is accelerating environmental degradation, affecting key natural resources such as vegetation, water bodies, and barren land. This study introduces an advanced machine learning (ML) and deep learning (DL)-based framework for high-accuracy LULC classification, urban sprawl quantification, and ecosystem service assessment, providing a more precise and scalable approach compared to traditional remote sensing techniques. A hybrid methodology combining ML models—Random Forest, Artificial Neural Networks, Gradient Boosting Machine, and LightGBM—with a 1D Convolutional Neural Network (CNN) was fine-tuned using grid search optimization to enhance classification accuracy. The integration of deep learning improves feature extraction and classification consistency, achieving an AUC of 0.93 for Dense Vegetation and 0.82 for Cropland, outperforming conventional classification methods. The study also applies the Markov transition model to project land cover changes, offering a probabilistic understanding of urban expansion trends and ecosystem dynamics, providing a significant improvement over static LULC assessments by quantifying transition probabilities and predicting future land cover transformations. The results reveal that urban areas in Abha expanded by 120.74 km² between 2014 and 2023, with barren land decreasing by 557.09 km² and cropland increasing by 205.14 km². The peak ecosystem service value (ESV) loss was recorded at USD 125,662.7 between 2017 and 2020, but subsequent land management efforts improved ESV to USD 96,769.5 by 2023. The resilience and recovery of natural land cover types, particularly barren land (44,163 km² recovered by 2023), indicate the potential for targeted restoration strategies. This study advances urban sustainability research by integrating state-of-the-art deep learning models with Markov-based land change predictions, enhancing the accuracy and predictive capability of LULC assessments. The findings highlight the need for proactive land management policies to mitigate the adverse effects of urban sprawl and promote sustainable ecosystem service recovery. The methodological advancements presented in this study provide a scalable and adaptable framework for future urbanization impact assessments, particularly in rapidly developing regions.

1. Introduction

Urban expansion is increasingly recognized as a critical issue in the contemporary world, shaped by rapid population growth, industrialization, and the global shift toward urban living. This ongoing process involves the conversion of natural and agricultural lands into urban areas, resulting in significant environmental, social, and economic consequences. Among these, ecological disruptions are particularly alarming, as urban growth encroaches upon forests, wetlands, and croplands, fragmenting habitats and threatening biodiversity. The effects of urbanization on ecosystem services (ESs) are especially pronounced in arid regions, where fragile ecological balances are particularly susceptible to disruption. For instance, rapid urban expansion in the Yangtze River Delta has significantly reduced terrestrial carbon storage due to the large-scale conversion of fertile croplands and natural ecosystems into built-up areas, impairing carbon sequestration and exacerbating climate change [1]. Similarly, urban sprawl in Wuhan, China, has resulted in substantial vegetation loss, diminishing the region’s net primary productivity and contributing to further ecological degradation [2]. In Ethiopia, urban expansion has encroached upon forested regions and croplands, posing serious threats to biodiversity and food security [3].

Beyond biodiversity loss, urbanization amplifies other ecological challenges, including soil degradation, water pollution, and disruptions to hydrological cycles. The proliferation of impervious surfaces such as roads and buildings diminishes groundwater recharge by obstructing rainwater infiltration, leading to increased surface runoff and heightening the risks of urban flooding. Additionally, urban growth contributes significantly to air pollution due to anthropogenic emissions from industrial and residential sources, increasing concentrations of airborne pollutants such as particulate matter (PM2.5), which pose severe risks to both environmental and public health [4]. These changes also intensify the urban heat island effect, a phenomenon observed in cities such as Hefei, China, where dense built-up areas trap and retain heat, intensifying local temperatures and imposing additional stress on both human populations and ecosystems while reducing biodiversity [3].

Urban expansion in developing countries, including Ethiopia and parts of Asia, has further exacerbated food security concerns by converting vast areas of agricultural land into urban settlements, thereby undermining ecological balance and agricultural productivity. The broader implications of urban expansion extend beyond localized environmental changes, influencing critical ecosystem services such as carbon sequestration, water filtration, and soil fertility—essential functions for maintaining ecological stability and human well-being. Studies focusing on urban agglomerations, such as the Shanghai–Hangzhou Bay Urban Agglomeration, have demonstrated that urban sprawl disrupts key ecosystem services, resulting in substantial declines in carbon sequestration, food production, and habitat quality [5]. Given these challenges, it is essential to assess Saudi Arabia’s urban expansion within a comparative global framework, drawing insights from similarly arid regions to develop informed and sustainable urban development policies.

The nature and trend of rapid urban expansion are shaped by various interrelated factors, primarily driven by population growth, economic development, and changes in land use policies. Population increases in urban areas, often fueled by rural-to-urban migration and natural population growth, have accelerated the demand for housing, infrastructure, and services. Economic activities, particularly in industries and services, tend to concentrate in urban centers, further promoting expansion as individuals and businesses seek opportunities in these areas. For instance, in cities like Mangalore, India, rapid urbanization has led to a 400% increase in built-up land between 1972 and 2018, primarily driven by population growth and economic activity [6]. Similarly, studies in North America indicate that urban areas are expanding at an unprecedented rate, driven by a combination of economic growth and population density [7]. Land use policies, including zoning regulations and urban planning strategies, play a pivotal role in directing the growth of cities. In Europe, the spatial delocalization of economic activities and discontinuous settlement expansion have reshaped urban areas, with a significant decline in peri-urban agricultural lands, as observed in the metropolitan region of Rome, Italy [8].

To better understand and manage urban expansion, various methods have been developed to measure trends and delineate urban areas. Remote sensing and geographic information systems (GISs) are widely used to track changes in land use and land cover over time, providing high-resolution spatial data that enable the detection of sprawl patterns and the quantification of expansion rates [9,10,11]. For example, a study in Lagos, Nigeria, used remote sensing techniques alongside the Cellular Automata–Markov model to analyze urban growth trends, showing a 25% increase in artificial terrain between 2000 and 2020 [12]. Population density thresholds, often based on census data, are also used to classify urban versus non-urban areas. In Sylhet Sadar, Bangladesh, a Cellular Automata model predicted a 220% increase in urban sprawl from 2020 to 2035 by analyzing population data alongside land-use patterns [13]. Nighttime light data from satellites, such as those used in China, also offer a proxy for human settlement patterns by capturing the intensity of artificial lighting, showing significant expansion in urban areas, especially in rapidly developing regions [14].

A key methodological approach in this study is the integration of multiple machine learning (ML) techniques to analyze land-use change. While ML applications in remote sensing and urban studies have expanded, the theoretical basis for combining multiple models remains underexplored. This study integrates algorithms such as Random Forest, artificial neural networks, gradient boosting machine, and convolutional neural networks to enhance classification accuracy, particularly in distinguishing vegetation cover and built-up areas [15]. Moreover, the study incorporates the Markov transition model to analyze land cover changes probabilistically, enabling the prediction of future land use transformations and assessing their implications for ecosystem sustainability [16]. Compared to previous studies, which often use static classification techniques, this study offers a novel methodological advancement by integrating machine learning models with urban growth metrics—including Shannon’s entropy, fragmentation indices, and edge growth analysis—to provide a more detailed understanding of the spatial structure and expansion patterns of urban areas over the past decade.

By addressing these theoretical and methodological gaps, this research provides a comprehensive analysis of urban expansion and its impacts on ecosystem services in Saudi Arabia. The primary objectives of this study are to quantitatively measure urban expansion and apply models to analyze the trends of urban growth. Specifically, this research aims to develop and employ accurate quantification methods that capture the extent and rate of urban sprawl, addressing the current literature’s focus on temporal patterns over exact measurements. In addition, the study applies the Markov transition model to analyze land cover changes probabilistically, offering insights into future urbanization trends. This approach enables a more accurate assessment of urban expansion dynamics compared to static LULC classification techniques. By integrating these models with quantitative data, the research aims to offer a better understanding of the current dynamics of urban expansion and its potential future trajectories, contributing to more informed urban planning and sustainable development strategies. The key contributions of this study include: (i) an improved LULC classification framework using machine learning models, which enhances classification accuracy beyond traditional remote sensing approaches; (ii) a predictive Markov-based model for urban expansion analysis, offering a dynamic perspective on future land transitions; (iii) a novel quantification of urban sprawl through advanced spatial metrics, providing a more detailed understanding of the spatial heterogeneity of urbanization; and (iv) a robust assessment of ecosystem service trade-offs, highlighting the interplay between urban expansion and natural resource resilience. By integrating these models with quantitative data, the research aims to offer a more nuanced understanding of the current dynamics of urban expansion and its potential future trajectories, contributing to more informed urban planning and sustainable development strategies.

2. Materials and Methods

2.1. Study Area

The study area Abha, the capital city of the Asir region, Kingdom of Saudi Arabia (KSA), extends between latitudes 17°59′21.452″ N and 18°30′33.812″ N and longitudes 42°18′56.269″ E and 42°56′25.909″ E, covering a total area of 2286.59 km² (Figure 1). The region features undulating terrain with elevations ranging from 1038 to 2990 m, averaging 2180 m above sea level. It boasts extensive natural vegetation, making it one of the most floristically diverse areas in Saudi Arabia. Notably, Jabal Al-Sooda, located in the northwest, is the highest mountain in the region at 2990 m and is known for its rich flora [17]. The varied climate and topography have fostered the growth of a wide range of plant species in the Asir Province [18]. However, anthropogenic activities, combined with steep slopes, fragile geology, and rainfall, have resulted in significant land degradation and ecological imbalance. Among the cities in this region, Abha serves as the capital. According to the General Statistics Authority’s 2019 census, the population of Abha, Khamis Mushayt, Al Wadi’ain, and Ahad Rufaida totals 2.2 million, with one-third of the population being Saudi nationals [19]. The urban layout of these cities, situated at various topographical levels, is a key factor driving urban development. The mountainous landscape is also a major tourist attraction. As these four cities anticipate substantial urban growth, sustainable urban planning is essential to improve the quality of life for both residents and tourists, addressing the challenges of urban sprawl and development in the region.

2.2. Materials

In this study, Landsat 8 Operational Land Imager (OLI) imagery from 28 February 2014, 8 March 2017, 13 February 2020, and 21 February 2023 with 0% cloud cover was used to derive essential land cover indices and analyze land use changes over time. The preprocessing steps included atmospheric correction, cloud masking, and radiometric calibration to enhance data quality. Atmospheric correction was performed using the surface reflectance (SR) product, which applies the Landsat ecosystem disturbance adaptive processing system (LEDAPS) algorithm. This process removes atmospheric distortions caused by aerosols, water vapor, and scattering effects, ensuring that the retrieved spectral reflectance values accurately represent ground conditions. The correction was performed using the Landsat Collection 2 Level-2 product in Google Earth Engine (GEE), which applies per-pixel adjustments based on the MODIS-derived atmospheric conditions. Cloud and shadow contamination were removed using the ‘QA_PIXEL’ band in the Landsat dataset. This approach employs the Function of Mask (FMask) algorithm to classify and eliminate pixels obscured by clouds, cirrus, and cloud shadows. Additionally, a threshold-based normalized difference snow index (NDSI) filter was applied to exclude any potential residual cloud pixels. Raw digital numbers (DNs) were converted into top-of-atmosphere (TOA) reflectance using the calibration coefficients provided in the metadata file. This conversion ensures consistency across images captured under different illumination conditions, aiding in accurate multi-temporal analysis.

The selection of training data is a critical step in land use classification, directly impacting model accuracy and generalizability. Training samples were derived from Google Earth images, field surveys, and expert knowledge. The process followed these key steps. A stratified random sampling technique was employed to ensure a balanced representation of all land cover classes. Each class—built-up, vegetation, barren land, cropland, and water bodies—was assigned a proportional number of training points. High-resolution satellite imagery (e.g., Sentinel-2 and Google Earth historical imagery) was cross-referenced to validate training samples. To maintain temporal consistency, training samples from different years (2014, 2017, 2020, and 2023) were manually inspected and updated based on historical land cover maps and ancillary datasets. Key spectral bands and vegetation indices (NDVI, NDBI, EVI) were extracted from the training samples to enhance classification accuracy.

2.3. Methods for Grid Search-Tuned Machine and Deep Learning Models for Land Use Land Cover Mapping

2.3.1. Random Forest

In this study, a Random Forest (RF) classifier was utilized, which is recognized for its ability to handle complex classification tasks, particularly in land use land cover (LULC) mapping. The RF algorithm, an ensemble-based method, uses multiple decision trees, combining their outputs to enhance prediction accuracy and reduce overfitting. To optimize the performance of the RF model, grid search was employed to tune the hyperparameters. This method involves an exhaustive search across different combinations of hyperparameters to identify the most suitable configuration. For this process, parameters such as the number of decision trees (n_estimators), the maximum depth of each tree (max_depth), and the minimum number of samples required to split a node (min_samples_split) were adjusted. For example, grid search-based hyperparameter tuning significantly improved the performance of RF models in urban land cover classification by reducing overfitting and enhancing accuracy [12]. The grid search was carried out within a cross-validation framework (cv = 3) to ensure model accuracy and prevent overfitting. Once the optimal parameters were identified, the model was trained on the training data and subsequently evaluated using test data. The classification performance was measured through a classification report, which provided precision, recall, and F1-score for each class. In other studies, such as those optimizing RF for the classification of land cover in São Paulo, the grid search method similarly improved classification results, achieving a global accuracy of over 89% [20]. This approach of tuning the RF classifier through grid search was found to enhance the accuracy and reliability of LULC mapping, contributing to more correct landscape classification outcomes, as seen in multi-temporal analysis projects across large geographic areas [21]. This highlights the importance of hyperparameter tuning in achieving high-performing LULC models.

2.3.2. Artificial Neural Network

This study used an artificial neural network (ANN) model applied for LULC mapping, utilizing grid search to fine-tune its hyperparameters. ANNs, known for their deep learning capabilities, are inspired by the human brain’s neural networks and are particularly effective in capturing complex, non-linear relationships in data. For this task, a multi-layer perceptron (MLP) classifier was employed, consisting of multiple layers of neurons that work together to generate predictions based on the input data. Grid search was employed to identify the best combination of hyperparameters, enhancing model performance. Parameters explored during this process included the number and size of hidden layers (hidden_layer_sizes), the activation function (activation) to introduce non-linearity, and the solver (solver) for weight optimization. In a recent study, grid search significantly improved ANN-based machine learning models for land use classification, ensuring more accurate and robust outcomes [21]. The grid search was executed with cross-validation (cv = 3) to ensure the model’s accuracy and to prevent overfitting. After selecting the optimal parameters, the model was trained using the training dataset (X_train, y_train) and subsequently evaluated on the test data (X_test). A classification report detailing precision, recall, and F1-score for each class was generated to assess the model’s performance. Hyperparameter optimization through grid search, combined with cross-validation, has been demonstrated in various studies to improve the prediction power of ANN models across different applications, including land cover and health-related predictions [22]. Grid search-tuned ANN models have been shown to achieve high accuracy in LULC mapping, allowing for optimal parameter selection that ensures reliable classification of land cover types [23].

2.3.3. Gradient Boosting Machine

The gradient boosting machine (GBM) was used for LULC mapping, a powerful ensemble machine learning algorithm that builds a series of decision trees, where each successive tree corrects the errors of its predecessors. This iterative approach allows GBM to effectively handle complex classification tasks by improving prediction accuracy with each iteration. To further enhance the model’s performance, a grid search was employed to identify the optimal set of hyperparameters, ensuring the best configuration for the task at hand. The hyperparameters tuned in the grid search process included the number of boosting stages (n_estimators), the learning rate (learning_rate), which controls the contribution of each tree to the final model, and the maximum depth (max_depth) of the individual trees. Recent research applied GBM for LULC and other prediction tasks, showing that hyperparameter tuning using grid search significantly enhances model performance, as observed in mineral prospectivity mapping in the Xiong’ershan Area [24]. Cross-validation (cv = 3) was utilized to avoid overfitting and assess the model’s generalization performance. After selecting the optimal hyperparameters, the GBM was trained on the training dataset (X_train, y_train) and evaluated using the test dataset (X_test). A classification report was generated, providing key performance metrics such as precision, recall, and F1-score for each class [25]. Additionally, comparative studies in chronic kidney failure prediction tasks found that grid search-tuned GBM models outperformed baseline models, achieving superior precision and F1-scores [26]. This demonstrates the versatility and power of GBM across various domains.

2.3.4. LightGBM

LightGBM (light gradient boosting machine) was employed for LULC mapping, which is an advanced machine learning algorithm known for its efficiency in handling large datasets and its ability to perform fast and accurate classification tasks. It builds multiple decision trees and uses a leaf-wise growth strategy, which allows it to focus on the most important features, making it particularly suitable for high-dimensional data in LULC mapping. To enhance the performance of LightGBM, a grid search approach was used to identify the best combination of hyperparameters. The hyperparameters tuned in this process included the number of boosting iterations (n_estimators), the number of leaves (num_leaves), which controls the complexity of each tree, and the learning rate (learning_rate), which governs the contribution of each tree to the final model. Grid search, combined with cross-validation (cv = 3), was employed to prevent overfitting and ensure that the model generalizes well on unseen data. A recent study demonstrated significant improvements in hyperparameter optimization using a Bayesian hyperband for machine learning tasks, showing superior computational time and accuracy [27]. After identifying the optimal parameters, the model was trained on the training dataset (X_train, y_train) and evaluated on the test dataset (X_test). A classification report was generated, providing a detailed breakdown of performance metrics like precision, recall, and F1-score. The use of LightGBM in similar applications, such as remote sensing image scene classification, has been shown to enhance model accuracy while optimizing computational efficiency [28]. These studies highlight the importance of hyperparameter tuning to achieve robust and reliable outcomes in LULC mapping tasks.

2.3.5. Convolutional Neural Network

A 1D convolutional neural network (1D CNN) was implemented for LULC classification. CNNs are highly effective deep learning models, especially suited for capturing spatial hierarchies and patterns in data. While typically applied to image data, 1D CNNs can be utilized for sequential or tabular data, making them applicable for LULC classification tasks where high-dimensional feature sets are available. The 1D CNN processes the data through a series of convolutional layers, enabling the model to learn intricate spatial dependencies between the features, which is critical for accurate classification in land cover studies. This approach has been particularly useful in applications such as wetland mapping and hyperspectral image classification, where CNN models significantly improved classification accuracy by leveraging spatial features [29].

The input data were pre-processed using standard scaling to ensure uniformity across feature values, followed by reshaping the data to fit the input format required for a 1D CNN. The architecture of the model consisted of a convolutional layer with 64 filters, followed by a fully connected layer and an output layer with softmax activation, designed to predict the land cover class probabilities. The model was compiled with the Adam optimizer and trained using categorical cross-entropy as the loss function. Training was performed for 10 epochs with a batch size of 32, and the performance was evaluated on the test data using accuracy as the key metric. The 1D CNN demonstrated strong performance, providing reliable classification results [30]. The use of deep learning models like 1D CNNs in LULC mapping is particularly important because of their ability to automatically extract relevant features from complex datasets, significantly improving classification accuracy compared to traditional models.

2.4. Methods for Accuracy Assessment of Developed Models

In this study, an extensive accuracy assessment was conducted for the developed models used in LULC classification, including RF, ANN, GBM, LightGBM, and 1D CNN. To evaluate model performance, key metrics such as precision, recall, F1-score, overall accuracy, producer’s accuracy, and user’s accuracy were calculated. Additionally, receiver operating characteristic (ROC) curves were plotted for each model to analyze their classification effectiveness across different LULC classes, using false positive rates (FPRs) and true positive rates (TPRs) to compute the area under the curve (AUC) values for each land cover type. Higher AUC scores indicate better discrimination ability between land cover classes [31].

To ensure a comprehensive classification performance assessment, confusion matrices were generated for each model. These matrices provide detailed insights into correct and incorrect classifications, allowing for the identification of systematic classification errors and model biases. Additional key metrics—producer’s accuracy (class-wise recall) and user’s accuracy (class-wise precision)—were included to measure classification consistency, particularly in challenging LULC categories such as Sparse Vegetation and Cropland, which often exhibit spectral confusion. The Kappa coefficient was computed to measure agreement between the classification results and reference data while accounting for random chance. To statistically validate model differences, paired t-tests and Wilcoxon signed-rank tests were conducted. The paired t-test, a parametric test, was applied to assess whether performance differences between models were statistically significant, assuming normally distributed classification errors. Meanwhile, the Wilcoxon signed-rank test, a non-parametric alternative, was used to compare models without assuming normality, ensuring robustness in model ranking and significance analysis. These statistical tests help confirm whether the observed accuracy differences between models are meaningful or occur by chance, improving the reliability of model selection.

Studies have shown that comparing multiple models, such as GBM and Random Forest, provides insights into model-specific advantages, such as better recall for certain classes or higher AUC scores for certain classifiers [32]. By integrating multiple evaluation metrics, including AUC, precision, recall, F1-scores, Kappa coefficient, producer’s and user’s accuracy, alongside paired t-tests and Wilcoxon tests, this study offers a comprehensive, statistically validated evaluation of LULC classification models. These assessments ensure that the models are robust, accurate, and suitable for large-scale land cover mapping applications [33,34]. The accuracy of classified land use maps was validated using a separate validation dataset (30% of the total labeled data), ensuring unbiased model performance evaluation. Confusion matrices were constructed for each classification year (2014, 2017, 2020, and 2023) to compare predicted versus actual land cover labels. Overall accuracy was computed as the proportion of correctly classified pixels relative to the total validation pixels. Additionally, the Kappa coefficient was used to quantify classification agreement beyond chance. ROC curves and AUC values were calculated separately for each land cover class to further refine model performance analysis. By combining statistical validation with detailed classification accuracy assessments, this study ensures a rigorous and well-rounded evaluation of LULC classification models, leading to improved classification reliability and interpretability.

2.5. Methods for Land Use Dynamics Using Markov Transition Model

The dynamics of LULC changes over time were analyzed using the Markov transition model, which allows for the examination of class transitions between different time periods. The Markov transition model is a widely used tool for understanding land cover dynamics, as it captures the probabilities of changes from one land cover class to another over consecutive time steps, providing insights into how land use patterns evolve. This method is particularly valuable in identifying dominant trends, such as urban expansion, vegetation loss, or land degradation [35]. To implement the Markov model, transition matrices were computed for four time periods: 2014 to 2017, 2017 to 2020, 2020 to 2023, and 2014 to 2023. These matrices quantify the likelihood of transitions among LULC classes (e.g., Built-up, Dense Vegetation, Sparse Vegetation, Water Bodies, Barren Land, Cropland). The LULC data for each year were first flattened to facilitate pairwise comparison, and for each class transition, the number of occurrences was tallied. The resulting counts were then normalized to produce a probability-based transition matrix, which provides the probability of each LULC class transitioning to another. This approach aligns with recent studies on Markov-based models, such as applications in environmental systems where time-varying transition rates were used to predict ecosystem changes [36]. The transition matrices were visualized using heatmaps, where the values indicate the probability of change from one class to another across each time period. This approach allows for an intuitive interpretation of land cover dynamics. For instance, the transition probabilities between classes, such as cropland to built-up areas, can help identify urbanization patterns, while transitions from vegetation classes to barren land can indicate environmental degradation. The generated heatmaps were plotted in a 2 × 2 grid, providing a comprehensive visualization of land use changes over the four periods. The use of transition models to predict complex spatial patterns has been widely validated in diverse fields, including studies on air pollution in urban areas [37].

2.6. Methods for Quantifying Urban Sprawl

In the present work, urban sprawl was quantified using a combination of urban growth analysis, Shannon’s entropy, fragmentation index, edge growth metrics, urban compactness index (UCI), aggregation index (AI), and leapfrog development index (LDI) to evaluate the spatial and temporal dynamics of urban expansion [38]. This integrated approach provides a comprehensive assessment of urbanization trends, capturing both the extent of land conversion and the structural changes in urban form. Such a multifaceted methodology is crucial for sustainable land management, urban planning, and environmental conservation, as it offers deeper insights into spatial organization, clustering patterns, and urban expansion tendencies.

First, urban growth was quantified by calculating the percentage of land that transitioned to built-up areas over consecutive years. This metric identifies the pace and magnitude of urban expansion, highlighting periods of rapid growth and stabilization. Shannon’s entropy was then computed to measure the spatial dispersion of urban and non-urban areas, with higher entropy values indicating increased spatial spread and reduced compactness [39,40]. A study conducted in West Bengal, India, utilized Shannon’s entropy to track built-up area dispersion trends over time, emphasizing its utility in urban sprawl analysis [39]. Next, the fragmentation index was calculated by identifying the number of discrete urban patches, a critical indicator of scattered urban growth. Higher fragmentation values signify more dispersed development patterns, which can influence infrastructure costs and urban connectivity. Edge growth was analyzed through the perimeter-to-area ratio, capturing the complexity and irregularity of urban expansion at boundaries. In Bahir Dar, Ethiopia, these metrics revealed increasing fragmentation due to rapid urbanization, demonstrating their applicability in assessing urban sprawl dynamics [41].

To further refine the spatial analysis, urban compactness (UCI) and aggregation index (AI) were introduced. UCI quantifies the degree of spatial cohesion, distinguishing between compact and sprawling urban structures, with lower values indicating dispersed development and higher values reflecting more consolidated urban areas. AI measures the clustering tendency of built-up areas, offering a statistical assessment of whether urbanization follows a contiguous or scattered pattern. The leapfrog development index (LDI) was also computed to evaluate discontinuous expansion, where values above 0.6 indicate significant leapfrogging, suggesting suburbanization or uncontrolled peripheral growth. These spatial metrics were calculated for LULC data across four time periods (2014, 2017, 2020, and 2023), ensuring consistency in temporal analysis. The results were saved as CSV files and visualized through a 2 × 2 grid format, providing a clear comparative interpretation of urban expansion trends over time. The importance of combining Shannon’s entropy with spatial metrics to assess urban sprawl patterns has been widely emphasized in studies across India and Ethiopia, highlighting their effectiveness in understanding complex urbanization dynamics [37].

2.7. Methods for Quantifying Natural Resource Loss Due to Urban Sprawl

In this study, the quantification of natural resource loss due to urban sprawl was carried out by analyzing transitions from natural resource land cover classes to built-up areas over time. Urban sprawl often results in the conversion of ecologically valuable areas, such as forests, water bodies, and agricultural land, into urbanized regions, leading to the degradation of natural ecosystems. To assess this impact, a methodology was developed to track the loss of natural resources in response to the expansion of built-up areas. This method has been widely applied in urbanization studies, such as in the Indian Himalayan region, where urban expansion caused significant loss of forested and agricultural land [42]. The analysis involved calculating the transition of pixels from natural resource classes (such as dense and sparse vegetation, water bodies, barren land, and cropland) to built-up areas. For each time period, the LULC images were compared, and the number of pixels transitioning from each natural class to the built-up class was computed. These transitions represent the loss of natural resources due to urbanization. The method was applied across multiple time periods, including 2014–2017, 2017–2020, and 2020–2023, to capture the temporal dynamics of resource loss, following methodologies similar to those used in studies of urban sprawl in Baghdad and Shanghai [43,44]. The resulting loss statistics were recorded for each time period, with a detailed breakdown of the transitions from each natural resource class to built-up land. This approach provides valuable insights into how urban expansion directly impacts various natural resources, highlighting specific areas where significant losses occur. By quantifying these losses, the study enables a better understanding of the ecological costs of urban sprawl, such as the loss of urban forests in Nairobi, which contributed to environmental degradation [45]. The ecological impacts of these transitions are significant, as shown by the transformation of farmland to urban areas in Budapest and other global cities [46]. Monitoring these changes is essential for policymakers to mitigate the negative effects of urbanization on the environment and to ensure that land use policies account for the preservation of critical natural resources. This has been demonstrated in the Beijing–Tianjin–Hebei region, where sprawl metrics were used to guide sustainable land-use strategies [2].

A linear regression analysis was conducted to quantify the relationship between built-up expansion and the loss of various natural resources, including dense vegetation, sparse vegetation, and water bodies. The primary goal was to understand the extent to which urban sprawl contributes to the degradation of these natural resources. Linear regression was employed to model the linear dependency between the independent variable, built-up growth, and each of the dependent variables representing natural resource depletion, offering a detailed and accurate statistical framework for quantifying the magnitude and direction of these impacts. Linear regression was selected for its ability to model the linear relationship between an independent variable (built-up growth) and dependent variables (loss of natural resources), providing a straightforward yet effective way to quantify these impacts. Similar methods have been successfully employed in land use change studies, demonstrating the efficacy of linear models in explaining resource loss due to urban growth [47]. To prepare the data, built-up growth percentages were used as the independent variable (X), and the losses in dense vegetation, sparse vegetation, and water bodies were used as the dependent variables (Y). Three separate linear regression models were fitted: one for dense vegetation loss, one for sparse vegetation loss, and one for water bodies loss. These approaches are consistent with studies that predict water resource trends and environmental sustainability using regression models [48]. The regression coefficients, which indicate the contribution of built-up expansion to the loss of each natural resource, were extracted from the models. The coefficients derived from this analysis reveal how much natural resource loss can be attributed to urban expansion. For instance, in regions like West Bengal, linear regression models have been effectively applied to quantify built-up expansion impacts on forest and water body losses [49]. This quantitative approach is critical for understanding the environmental consequences of urbanization, enabling urban planners and policymakers to develop strategies that mitigate the negative impacts of urban sprawl on natural ecosystems and promote sustainable development. Similar studies in various regions, including Africa and Southeast Asia, have employed regression models to establish causal relationships between urbanization and resource loss, providing essential insights for sustainable land management [50,51].

2.8. Methods for Analyzing Trade-Off Between Ecosystem Loss and Recovery Due to Urban Sprawl

Urban expansion significantly impacts ecosystem service dynamics, leading to both ecosystem service loss due to land conversion and ecosystem service recovery when urban areas transition back to natural land cover. This study employs a resilience index-based approach, combined with ecosystem service valuation (ESV) analysis, to quantify these trade-offs and assess their long-term implications for sustainable land-use planning. Similar approaches to measuring ecosystem resilience in urban environments have been applied in studies, such as one that examined the resilience of China’s ecosystems over time [52]. It is conducted through the following steps.

2.8.1. Quantification of Ecosystem Loss Due to Urban Sprawl

To assess the economic and ecological implications of natural resource loss due to urban sprawl, a transition matrix-based approach was applied. LULC maps were generated for multiple time periods, and changes in land use were quantified by identifying the number of pixels transitioning from natural land cover classes (dense vegetation, sparse vegetation, water bodies, barren land, and cropland) to built-up areas. The number of converted pixels was then multiplied by the area per pixel to calculate the total land area (in hectares) lost to urbanization in each period.

The economic loss associated with these transitions was quantified using the benefit transfer method, which applies standardized ecosystem service coefficients (USD/ha/year) from Costanza et al. [53]. These coefficients provide an estimate of the annual monetary value of ecosystem services per hectare for each LULC class. These coefficients were originally expressed in USD in 1994, and no inflation adjustment was applied, as the study focuses on relative comparisons of ESV over time rather than absolute monetary values. To ensure regional relevance, the coefficient values were selected from global and regional studies applicable to arid and semi-arid landscapes, aligning with the ecological characteristics of the study area. The total ESV for each land cover class was computed using the formula:

ESV = A × VC

where A is the area of the LULC class (in hectares), and VC is the value coefficient (USD/ha/year) from Costanza et al. (1997) [53]. Since 1 km² equals 100 hectares, all LULC areas were converted accordingly. The valuation coefficients used in this study include 969 USD/ha/year for dense vegetation (forests), 232 USD/ha/year for sparse vegetation (grasslands), 14,785 USD/ha/year for water bodies (wetlands, rivers, lakes), and 92 USD/ha/year for cropland. Built-up and barren lands were assigned a coefficient of 0 USD/ha/year due to their minimal ecosystem service contributions. Recent studies in the Yangtze River Basin [54] and Italy [55] have demonstrated similar approaches to estimating ecosystem service losses associated with urbanization. By applying this framework, the study provides an economic valuation of ecosystem degradation caused by urban sprawl in the study area.

2.8.2. Quantification of Ecosystem Service Recovery Following Urban Sprawl

While urban expansion results in ecosystem degradation, certain regions experience ecosystem recovery, where built-up land transitions back to natural land cover classes over time. To quantify this recovery, a resilience index was computed by analyzing the total area of built-up pixels that reverted to vegetation, cropland, water bodies, or barren land during successive time periods. This analysis was performed by comparing LULC images across years to track post-urbanization land transitions.

To estimate the monetary value of ecosystem service recovery, the ecosystem service coefficients (USD/ha/year) from Costanza et al. [53] were applied to the recovered land using the same formula:

ESV_recovery = A × VC

where A represents the area of built-up land that transitioned back to natural land cover, and VC corresponds to the respective ecosystem service coefficient. Studies in Shanghai [44] and the Yangtze River Basin [54] emphasize the importance of such recovery analyses in urban resilience research, as they provide insights into how natural systems regain ecological functionality post-urbanization.

2.8.3. Trade-Off Analysis Between Ecosystem Loss and Recovery

A comparative analysis of ecosystem loss and recovery was conducted to determine whether urban sprawl leads to net degradation or a sustainable balance between land-use transitions. The total ESV loss due to urbanization was compared against the ESV gained through natural land recovery, allowing for an assessment of whether ecosystem recovery compensates for the services lost due to built-up expansion [56].

To evaluate the statistical significance of the differences between ecosystem loss and recovery, a paired t-test was performed. This test examines whether the mean difference between two paired datasets (ESV loss and ESV recovery) is significantly different from zero. The test statistic (t-value) and p-value were computed, with p < 0.05 indicating a statistically significant difference. If normality assumptions were not met, a Wilcoxon signed-rank test was applied as a non-parametric alternative. Similar statistical approaches have been validated in environmental impact studies [57,58,59] to assess ecological trade-offs in rapidly urbanizing regions.

3. Results

3.1. Analysis of ML and DL Model Development for LULC Modeling

3.1.1. Model Development

In this study, several machine learning models were developed and fine-tuned for LULC classification using grid search optimization to identify the best-performing hyperparameters. The RF model was optimized with 200 estimators, a minimum sample split of 5, and no maximum depth constraint. The ANN was configured with two hidden layers of 50 neurons each, using the ReLU activation function and the Adam solver. For GBM, the optimal parameters included a learning rate of 0.1, a maximum depth of 5, and 200 estimators. Lastly, the LightGBM model was tuned with 100 estimators, 31 leaves, and a learning rate of 0.1. These optimized models ensured accurate predictions for LULC mapping, enhancing their robustness and reliability in classification tasks.

Figure 2 shows the training and validation loss curves for a 1D convolutional neural network (CNN) model over 10 epochs. Initially, both training and validation loss start relatively high, with training loss at approximately 0.92 and validation loss at around 0.83. By epoch 3, a sharp decrease in both losses is observed, with training loss reducing to around 0.81 and validation loss to 0.80, demonstrating rapid convergence in the early stages of training. From epochs 4 to 7, the loss continues to decrease steadily, with training loss dipping to approximately 0.78 by epoch 7, indicating further improvement in model fitting. The validation loss also follows a similar decreasing trend, reaching around 0.79, suggesting that the model generalizes well to unseen data up to this point.

However, after epoch 7, a slight divergence between the training and validation loss curves begins to emerge. The training loss continues to decrease, reaching a final value of approximately 0.76 by epoch 10, while the validation loss exhibits minor fluctuations, increasing slightly to around 0.80 after epoch 8. This could indicate the onset of slight overfitting, as the model fits the training data more closely, but its performance on the validation data plateaus. The model’s training and validation loss curves exhibit strong convergence, with minimal overfitting observed within the first 10 epochs. The difference between the final training and validation losses (0.04) suggests the model achieves a good balance between learning and generalization.

3.1.2. Accuracy Assessment of the Models and Comparisons

Figure 3 presents confusion matrices for five different machine learning models—RF, ANN, GBM, LightGBM, and 1D CNN—used to classify six land cover classes: Built-up, Dense Vegetation, Sparse Vegetation, Water Bodies, Barren Land, and Cropland. Each matrix highlights both the correct classifications (diagonal values) and misclassifications (off-diagonal values), allowing for a detailed comparison of model performance across these classes. The RF model demonstrates strong classification accuracy, particularly for Dense Vegetation (1009 correctly classified out of 1203 instances) and Cropland (906 out of 1098). However, it struggles somewhat with Water Bodies, correctly classifying only 70 out of 120 instances, and also has notable misclassifications in the Built-up class, where 903 out of 1004 are correctly classified. Misclassifications are prominent, with Water Bodies being confused with Barren Land and Sparse Vegetation. The ANN model performs similarly, excelling in Dense Vegetation (1008 correctly classified) and Cropland (952), but it shows more confusion between Sparse Vegetation and Barren Land, misclassifying 110 Sparse Vegetation instances as Barren Land. Like Random Forest, ANN struggles with correctly classifying Water Bodies. The GBM model also demonstrates strong performance with Dense Vegetation (1009 correctly classified) and Cropland (905), but there are significant misclassifications in Sparse Vegetation, where 170 instances are misclassified as Barren Land, showing that distinguishing between these classes is a challenge for GBM. The LightGBM model maintains similar classification performance, with strong results in Dense Vegetation (1073 correctly classified) and Cropland (921). Finally, the 1D CNN model exhibits the highest classification accuracy for Dense Vegetation (1073 correctly classified out of 1203) and Barren Land (844 correctly classified). However, it also shows a significant number of misclassifications in Sparse Vegetation (583 correctly classified out of 919), with confusion primarily occurring between Barren Land and Cropland.

The evaluation of machine learning models for LULC classification shows that CNN achieved the highest overall accuracy of 0.805, followed by Random Forest (0.797), LightGBM (0.790), GBM (0.785), and ANN (0.778). The Kappa coefficient, which measures agreement beyond chance, was also highest for CNN (0.782), indicating strong classification reliability, while Random Forest (0.765), LightGBM (0.758), GBM (0.751), and ANN (0.742) followed. Precision values ranged from 0.776 (ANN) to 0.812 (CNN), with Random Forest achieving 0.798, confirming CNN’s higher ability to avoid false positives. Similarly, recall values varied between 0.768 (ANN) and 0.809 (CNN), with GBM and LightGBM closely following at 0.784 and 0.790, respectively, indicating CNN’s superior ability to correctly classify different land cover types. The F1-score, a balance between precision and recall, was highest for CNN (0.810), with Random Forest (0.792), LightGBM (0.786), and GBM (0.780) performing similarly, while ANN lagged at 0.771.

Producer’s accuracy (recall per class) and user’s accuracy (precision per class) revealed variations across LULC categories. Built-up areas were classified with high producer accuracy (CNN: 92.1%, RF: 90.3%) but had moderate user accuracy (CNN: 86.5%, RF: 85.8%) due to spectral confusion with barren land. Sparse vegetation had the lowest classification reliability, with producer accuracy of 76.3% (CNN) and 74.2% (RF), showing difficulties in separating it from cropland. Dense vegetation and water bodies were consistently well classified, achieving producer accuracy above 85% for all models. These findings suggest that CNN outperforms other models in overall accuracy and reliability, particularly in urban areas, but all models struggle with spectral similarities in vegetation classes.

Figure 4 illustrates the receiver operating characteristic (ROC) curves for five machine learning models—Random Forest, ANN, GBM, LightGBM, and 1D CNN—for six land cover classes: Built-up, Dense Vegetation, Sparse Vegetation, Water Bodies, Barren Land, and Cropland. The area under the curve (AUC) values for each class are presented alongside their ROC curves, providing a quantitative measure of classification performance. The Random Forest model achieves high AUC values for Dense Vegetation (0.93), Cropland (0.82), and Built-up (0.85), indicating excellent classification performance for these classes. However, performance is significantly lower for Sparse Vegetation (AUC = 0.69) and Water Bodies (AUC = 0.70), suggesting these classes are more challenging to differentiate. The AUC for Barren Land stands at 0.78, demonstrating moderate classification ability. The ANN model exhibits similar trends, with the highest AUC for Dense Vegetation (0.91) and Cropland (0.83). Built-up areas are classified with an AUC of 0.84, while Sparse Vegetation and Water Bodies show weaker classification performance, with AUCs of 0.69 and 0.77, respectively. The Barren Land class has an AUC of 0.80, indicating reasonable classification accuracy. For GBM, the model performs well with Dense Vegetation (AUC = 0.90) and Built-up (AUC = 0.84). Cropland is also classified accurately, with an AUC of 0.82. Similar to the other models, Sparse Vegetation (AUC = 0.69) and Water Bodies (AUC = 0.72) are the most challenging classes, while Barren Land is moderately classified with an AUC of 0.78. The LightGBM model shows the best AUC for Dense Vegetation (0.91), followed by Cropland (0.83) and Built-up (0.85) land. Sparse Vegetation and Water Bodies have relatively low AUCs of 0.70 and 0.73, respectively. The classification of Barren Land achieves an AUC of 0.80, which is consistent with the other models. Finally, the 1D CNN model performs similarly to the other models, with Dense Vegetation achieving the highest AUC (0.93) and Built-up areas reaching 0.84. Cropland and Barren Land also show reasonable performance with AUCs of 0.82 and 0.81, respectively. Sparse Vegetation (AUC = 0.67) and Water Bodies (AUC = 0.72) remain the most difficult classes for accurate classification. Therefore, it can be concluded that dense vegetation and cropland consistently achieve the highest AUC values across all models, indicating strong classification performance for these classes. Conversely, Sparse Vegetation and Water Bodies have consistently lower AUCs, highlighting the difficulties in accurately classifying these land cover types. The overall performance of the models is comparable, with minor variations in AUC values across the different classes. The overall accuracies for the models are Random Forest (0.797), ANN (0.778), GBM (0.785), LightGBM (0.790), and CNN (0.805). Paired t-tests indicate no statistically significant difference between Random Forest and ANN (p = 0.287), Random Forest and GBM (p = 0.412), Random Forest and LightGBM (p = 0.362), and Random Forest and CNN (p = 0.198). Similarly, Wilcoxon signed-rank tests confirm that the differences are not statistically significant, with p-values of 0.294 (RF vs. ANN), 0.415 (RF vs. GBM), 0.374 (RF vs. LightGBM), and 0.205 (RF vs. CNN). These results suggest that while minor variations exist in accuracy, the models perform comparably, with no significant statistical difference in classification performance.

3.2. Assessment of LULC Dynamics

Figure 5 and Figure 6 present the temporal dynamics of various land cover classes—Built-up, Dense Vegetation, Sparse Vegetation, Water Bodies, Barren Land, and Cropland—over the years 2014, 2017, 2020, and 2023. Quantitative values for each land cover class are expressed in square kilometers (km²), revealing significant trends in the expansion and contraction of different land types across these time periods. From 2014 to 2023, Built-up areas exhibited a steady increase, growing from 370.05 km² in 2014 to 490.79 km² by 2023. This increase reflects ongoing urban expansion and infrastructure development. In parallel, Dense Vegetation also shows a slight but continuous increase, expanding from 99.62 km² in 2014 to 150.93 km² in 2023, indicating some improvement in vegetation coverage, possibly due to reforestation or improved land management practices. The class of Sparse Vegetation follows a similar trend, starting at 298.07 km² in 2014 and growing significantly to 477.26 km² in 2023. This increase in sparse vegetation could indicate land degradation or shifts in vegetation density due to factors such as land use changes or environmental pressures. The area of Water Bodies, on the other hand, remains relatively stable, with only minor changes from 3.88 km² in 2014 to 4.58 km² in 2023, reflecting minimal fluctuations in water resources over this period. A notable trend is observed in the Barren Land category, which shows a sharp decline from 1435.79 km² in 2014 to 878.70 km² in 2023. This decrease in barren land suggests that large portions of previously unused or degraded land have been converted into other land uses, potentially for agriculture, vegetation recovery, or urban development. Cropland exhibits a prominent increase, growing from 170.16 km² in 2014 to 375.30 km² in 2023, indicating an expansion in agricultural activities over the observed period.

From the discussion, it can be seen that significant trends exist in land cover dynamics, particularly the steady increase in urban areas and cropland, the recovery or reforestation of vegetation, and the considerable reduction in barren land. These trends reflect land use changes likely driven by human activity, urban expansion, and environmental management initiatives over the past decade.

Figure 7 presents transition matrices illustrating land cover changes for four time periods: 2014–2017, 2017–2020, 2020–2023, and 2014–2023. The matrices quantify transitions between six land cover classes—Built-up, Dense Vegetation, Sparse Vegetation, Water Bodies, Barren Land, and Cropland—by showing the proportion of land that shifted from one class to another. For the period 2014 to 2017, the matrix reveals that 81% of Built-up areas remained unchanged, while 6% of Sparse Vegetation transitioned to Cropland, and 16% of Water Bodies converted to Sparse Vegetation. Significant stability is observed in Dense Vegetation (91% retention), while Sparse Vegetation experienced notable transitions to Barren Land (16%).

Between 2017 and 2020, Built-up areas retained 84% of their previous coverage, with 10% transitioning to Cropland. Dense Vegetation showed a slight decrease in retention (83%), with small portions converting to Sparse Vegetation (5%) and Cropland (2%). Water Bodies exhibited a marked reduction in stability, retaining only 55%, with notable transitions to Sparse Vegetation (9%) and Barren Land (21%). During the phase 2020 to 2023, Built-up areas remained relatively stable (83% retention), while Sparse Vegetation retained 69% of its area. Cropland expanded, absorbing 13% of Sparse Vegetation. Water Bodies showed improved retention (65%), though Barren Land exhibited significant transitions to Cropland (13%). Over the longer period from 2014 to 2023, 80% of Built-up areas persisted, while 19% of Sparse Vegetation converted to Cropland. Dense Vegetation showed the highest retention (93%), with minimal transitions. Water Bodies retained 67% of their area, while Barren Land experienced notable transitions to Cropland (32%) and Built-up areas (11%).

The matrices, therefore, indicate substantial land cover transitions, with Built-up areas, Cropland, and Dense Vegetation showing the highest stability across the periods. In contrast, Sparse Vegetation and Water Bodies experienced greater flux, with notable transitions into Barren Land, Cropland, and other categories. The continuous expansion of Cropland, particularly at the expense of Sparse Vegetation and Barren Land, reflects shifts toward agricultural use, while urbanization (Built-up areas) continues to expand steadily across all periods.

3.3. Quantification of Urban Sprawl

Figure 8 presents a detailed quantitative analysis of urban expansion metrics over three distinct time periods: 2014–2017, 2017–2020, and 2020–2023. Six key urbanization indicators—urban growth, Shannon’s entropy, urban fragmentation, urban edge growth, UCI, and AI—are analyzed to capture both temporal trends and spatial distribution shifts in urban sprawl. The urban growth over time graph (top left) indicates a decline in urban growth percentage from 3.05% (2014–2017) to 2.76% (2020–2023), suggesting a slowdown in land conversion, likely due to land-use restrictions, geographical constraints, or increased urban densification. In contrast, Shannon’s entropy over time (top right) shows an increase from 0.32 in 2014 to 0.38 in 2023, highlighting a shift from compact urban growth toward a more dispersed expansion pattern. This trend aligns with a rise in the leapfrog development index (LDI), which peaked at 0.63 in 2017 before declining to 0.55 by 2023, indicating a period of scattered development followed by more controlled urban expansion. The urban fragmentation index confirms this trend, with the number of urban patches increasing from 22,200 in 2014 to a peak of 23,000 in 2017, then declining to 22,600 in 2023, reflecting a shift from highly fragmented growth toward urban consolidation. The AI increased from 78% in 2014 to 82% in 2023, further supporting the trend of built-up areas becoming more clustered rather than dispersed over time.

The urban edge growth over time graph (bottom right) reveals a steady decline in the perimeter-to-area ratio from 0.54 in 2014 to 0.51 in 2023, indicating that urban patches are becoming more compact and regular in shape, likely due to improved urban planning strategies. This is further corroborated by the urban compactness index (UCI), which increased from 0.65 in 2014 to 0.71 in 2023, demonstrating a transition toward less fragmented, more contiguous urban development. Additionally, the built-up density (BUD) increased from 4250 persons/km² in 2014 to 4820 persons/km² in 2023, reflecting a shift toward higher-density urban cores rather than uncontrolled peripheral expansion. These trends suggest that urbanization in Abha City has evolved from a phase of rapid, fragmented sprawl (2014–2017) to a more structured and consolidated growth pattern (2020–2023). The findings highlight the impact of urban planning policies in reducing uncontrolled land conversion and promoting more sustainable urban expansion, with increasing densification and clustering of built-up areas. This shift toward compact growth patterns has significant implications for infrastructure efficiency, transportation planning, and ecosystem conservation, as it may reduce resource consumption, enhance land-use efficiency, and minimize the environmental footprint of urban expansion.

3.4. Assessing Trade-Off Between Ecosystem Loss and Recovery Due to Urban Sprawl

Urban expansion and land-use changes significantly impact ecosystem services, leading to both economic and ecological consequences. The following analysis provides a quantitative assessment of ecosystem service loss and recovery, examining the monetary valuation of ecosystem services over time, the trade-off between loss and recovery, and the statistical significance of these trends. The results highlight both the challenges of urban sprawl and the effectiveness of environmental restoration programs in Saudi Arabia, particularly in alignment with the nation’s Vision 2030 initiative.

3.4.1. Monetary Valuation of Ecosystem Services over Time

The estimated economic value of ecosystem services provides insights into how land-use changes affect environmental contributions in monetary terms. As shown in Table 1, the total ESV of dense vegetation increased from USD 9.65 million in 2014 to USD 14.62 million in 2023, representing a 51.4% growth in ecosystem services from forests. This positive trend suggests the implementation of afforestation, reforestation, and conservation programs that have expanded vegetative cover over the years. Similarly, sparse vegetation increased its ESV from USD 6.91 million in 2014 to USD 11.08 million in 2023, reflecting an overall improvement of 60.4%. This change may be attributed to desert greening projects, sustainable rangeland management, and ecological restoration efforts. However, fluctuations in 2020 indicate minor losses due to urban expansion, which required further interventions to restore vegetative areas. Water bodies, representing wetlands and reservoirs, experienced a moderate increase in ESV from USD 5.73 million in 2014 to USD 6.76 million in 2023. This suggests enhanced water resource management strategies, including the construction of artificial reservoirs, wetland protection policies, and desalination initiatives to support water availability in arid environments. The ESV of cropland also increased significantly, from USD 1.56 million in 2014 to USD 3.45 million in 2023, marking a 120.7% improvement. This growth aligns with the expansion of agricultural zones, sustainable farming practices, and food security policies promoted under Saudi Arabia’s Vision 2030. Despite these positive trends, urban expansion has led to substantial losses of ecosystem services, particularly in regions where natural land cover was converted into built-up areas. While some ecosystem services showed resilience and recovery, urban sprawl continued to exert pressure on natural landscapes, necessitating further mitigation measures.

3.4.2. Trade-Off Between Ecosystem Loss and Recovery Due to Urban Sprawl

Urbanization results in significant ecosystem service losses, particularly when natural landscapes are converted to built-up areas. Table 2 quantifies the monetary loss of ecosystem services due to land conversion, highlighting that the most substantial loss occurred during 2014–2017, with a total ESV loss of USD 79.59 million per year. This period coincides with rapid infrastructure development, leading to the conversion of dense vegetation (USD −15.88 million), sparse vegetation (USD −41.81 million), cropland (USD −16.64 million), and water bodies (USD −5.24 million) into urban land. The scale of these losses underscores the environmental costs of urban expansion, emphasizing the need for effective land-use planning. Between 2017–2020, the total loss of ecosystem services declined to USD 35.07 million per year, reflecting a 56% reduction compared to the previous period. This decline suggests that policy interventions, conservation projects, and land management efforts were partially successful in reducing urban-driven ecosystem degradation. However, water bodies experienced a significant loss of USD 14.49 million, indicating potential issues related to water scarcity, reduced wetland areas, or increased groundwater extraction. In 2020–2023, total ESV loss increased again to USD 52.41 million per year, primarily due to large-scale urban expansion initiatives. The loss of dense vegetation (USD −29.26 million) and sparse vegetation (USD −12.77 million) accounted for 80% of the total ESV decline, indicating the continued impact of development on vegetative landscapes. While urban growth remains a necessary component of economic expansion, these figures highlight the persistent challenge of balancing urbanization with ecological sustainability.

Despite these losses, ecosystem service recovery played a crucial role in offsetting urban-driven degradation, as shown in Table 3. Between 2014–2017, ecosystem service recovery exceeded losses, with an ESV gain of USD 85.2 million per year, compared to a loss of USD 79.59 million per year. This net positive gain of USD 5.61 million suggests that targeted afforestation, land restoration, and ecological rehabilitation efforts were highly effective in compensating for urban sprawl. Between 2017–2020, ecosystem service recovery closely matched losses, with USD 34.82 million gained versus USD 35.07 million lost, leading to a net deficit of USD 253,860 per year. This balance suggests that while urban expansion continued, land restoration programs successfully mitigated most of the ecological damage, highlighting strong policy enforcement in this period. During 2020–2023, ecosystem service recovery exceeded losses again, with USD 53.78 million gained versus USD 52.41 million lost, resulting in a net gain of USD 1.36 million per year. This renewed positive balance indicates that urban expansion was counteracted by stronger conservation efforts, particularly in vegetation and cropland areas. These findings suggest that Saudi Arabia’s Vision 2030 is actively contributing to sustainable urban growth, ensuring that land degradation is systematically offset by restoration initiatives.

3.4.3. Statistical Validation of the Ecosystem Service Trade-Off

To statistically assess whether ecosystem recovery effectively compensates for urban-driven ecosystem losses, a paired t-test and Wilcoxon signed-rank test were conducted. The paired t-test yielded a test statistic of −1.776 with a p-value of 0.220, while the Wilcoxon signed-rank test produced a test statistic of −0.674 with a p-value of 0.500. These statistical results indicate that there is no statistically significant difference between ecosystem service loss and recovery (p > 0.05). In practical terms, this means that land restoration efforts are increasingly balancing ecosystem degradation, although some fluctuations remain. The non-significance of the test results suggests that Saudi Arabia’s land-use policies are progressively improving ecological resilience, as evidenced by the narrowing gap between loss and recovery over time. However, while the statistical results indicate overall balance, they do not account for localized variations in urban expansion and land recovery rates. Some regions may still experience net ecosystem losses, necessitating continuous monitoring, targeted restoration, and stricter land-use regulations.

The quantitative analysis of ecosystem service valuation, loss, and recovery provides compelling evidence that urbanization-induced degradation has been progressively mitigated by conservation policies in Saudi Arabia. The economic value of dense and sparse vegetation, cropland, and water bodies has improved significantly over the years, reinforcing the positive impact of afforestation and water management initiatives. While periods of rapid urban expansion have led to temporary declines in ecosystem services, targeted restoration efforts have ensured a strong recovery, leading to net ecosystem gains in recent years. The statistical validation of these trends confirms that Saudi Arabia’s conservation strategies are successfully balancing urban growth with environmental sustainability.

4. Discussion

The study’s findings provide a comprehensive analysis of LULC dynamics, urban sprawl, and the resulting ecosystem service trade-offs. The development and fine-tuning of machine learning ML and deep learning DL models, RF, ANN, GBM, LightGBM, and 1D CNN have resulted in highly precise classification accuracies for the six land cover classes. The models consistently achieved high classification accuracy for Dense Vegetation and Cropland, as indicated by confusion matrices and ROC curves. The highest-performing models, including RF and 1D CNN, showed strong classification capabilities for these classes, with AUC values as high as 0.93 for Dense Vegetation and 0.82 for Cropland. However, lower accuracy was observed for classes such as Sparse Vegetation and Water Bodies, which displayed greater classification challenges across all models, highlighting the complexities of distinguishing these land types due to their overlapping spectral characteristics. The analysis of LULC dynamics over time demonstrated significant changes, including the steady expansion of built-up areas, cropland growth, and a sharp decline in barren land from 2014 to 2023. The transition matrices captured significant transitions between land cover classes, particularly the expansion of cropland and urban areas at the expense of sparse vegetation and barren land, revealing the intense land cover transformations driven by urbanization and agricultural expansion. The quantification of urban sprawl provided key insights into the spatial distribution and fragmentation of urban growth. The temporal analysis showed a deceleration in urban growth, declining from 3.05% between 2014 and 2017 to 2.76% between 2020 and 2023, while Shannon’s entropy exhibited an increasing trend, indicating more spatially dispersed urban expansion. Urban fragmentation also saw a peak in 2017, with the number of urban patches increasing, but this was followed by a decrease in fragmentation in 2020–2023, suggesting a shift toward more consolidated urban forms. The steady decline in urban edge growth further supported the idea of more compact urban growth in recent years. Natural resource loss due to urban expansion was particularly notable for barren land, followed by sparse vegetation. However, using linear regression analysis, it was found that built-up expansion contributed positively to dense vegetation loss while indicating potential improvements in sparse vegetation and water bodies. This highlights the complex interplay between urban expansion and resource dynamics, where both losses and recoveries occur simultaneously.

The significance of these findings lies in their implications for land use planning and environmental management in rapidly urbanizing regions. The study reveals critical insights into the trade-off between urban expansion and ecosystem service dynamics, showing that urban growth is not only consuming land but also exerting significant pressure on natural resources. While dense vegetation and cropland show high classification accuracy and stability over time, the challenges of classifying sparse vegetation and water bodies emphasize the need for improved techniques to accurately monitor more vulnerable land types. The increasing spatial dispersion of urban areas, as reflected by the rise in Shannon’s entropy, indicates that sprawling urban growth is a persistent issue, which can lead to more inefficient land use, increased infrastructure costs, and greater environmental impacts. The reduction in fragmentation and edge growth, however, suggests that recent urban development efforts may be moving toward more sustainable, compact urban forms, which could mitigate some of the negative effects of sprawl. The findings related to ecosystem service loss and recovery highlight the importance of effective land management strategies in mitigating the adverse effects of urban expansion. Although the paired t-test revealed no statistically significant difference between ecosystem service loss and recovery, the recovery values observed in 2023 indicate tangible efforts to restore ecosystem functions post-conversion. This suggests that while urbanization results in significant ecosystem service losses, appropriate recovery efforts—such as reforestation, water management, and land reclamation—can play a pivotal role in balancing this loss over time. These insights emphasize the need for integrated urban planning approaches that account for both immediate land cover transformations and long-term ecosystem resilience. Policymakers and urban planners should use these findings to guide the development of urban areas in ways that minimize environmental degradation while enhancing ecosystem services, ensuring a more sustainable balance between urban growth and environmental conservation.

The results of this study align with contemporary research on LULC changes and ES evaluations, although there are notable distinctions in methodologies and findings. Previous studies conducted in Saudi Arabia have also implemented machine learning techniques, including SVM and ANN, to model urban expansion and assess its impact on ecosystem services. These investigations have shown that rapid urbanization, particularly the expansion of built-up areas, leads to significant declines in ecosystem functions such as hydrological balance and biodiversity [4]. Similarly, the present study, utilizing machine learning approaches like RF and one-dimensional convolutional neural network (1D CNN), has demonstrated high classification accuracies for land cover types, such as dense vegetation and cropland. Despite this consistency, certain regional differences in urban expansion patterns have emerged. This study identifies a slowdown in urban sprawl and a transition toward more compact urban growth in recent years. However, contrasting trends are observed in other regions, such as Ethiopia, where urban expansion continues at an accelerated rate, leading to the conversion of agricultural land into urban areas. Such studies highlight the persistent challenge of balancing urban growth with sustainable land management, especially in developing nations where conservation efforts struggle to keep pace with rapid urbanization [60]. Furthermore, this study’s difficulty in distinguishing between sparse vegetation and water bodies is reflective of classification challenges encountered in other studies. For instance, the application of Cellular Automata–Markov models in ecosystem service valuation has revealed similar spectral ambiguity in satellite imagery-based land classification [61].

While this study shows that ecosystem service recovery efforts, such as reforestation, help to mitigate the negative effects of urbanization, others argue that ecosystem recovery lags behind the rates of degradation, particularly in high-urbanization regions like China. Research in Shanghai demonstrated a sharp decline in ecosystem services, even as efforts to restore certain areas improved their hydrological regulation functions [44]. This discrepancy in recovery rates may be attributed to differences in regional policies and environmental conditions, emphasizing the need for tailored restoration strategies across different urbanizing regions. These comparative insights illustrate the global complexity of LULC dynamics, where urban growth, land transformation, and ecosystem service trade-offs require region-specific approaches for sustainable land management.

To ensure more eco-friendly and sustainable urban growth in the future, both short-term and long-term mitigation measures must be implemented, keeping in mind the uniqueness of the region. In the short term, cities can adopt ecosystem service zoning and strategic land use planning. For example, the zoning approach, which integrates ecosystem service value (ESV) and ecological risk assessments, has been shown to improve urban ecosystem management by delineating areas for priority development and ecological improvement. This approach has been successfully applied in Shizuishan City, China, and can be adapted to other regions experiencing rapid urbanization [62]. Long-term measures should focus on incorporating nature-based solutions (NbSs) into urban planning. The restoration of green infrastructure, such as urban forests and wetlands, can enhance biodiversity and improve resilience to climate change. Studies in the Guangdong–Hong Kong–Macao Greater Bay Area (GBA) highlight the benefits of carbon-neutral pathways, where increasing forest cover led to significant improvements in habitat quality and carbon storage [63]. Moreover, the adoption of sustainable cropland management practices, such as conservation tillage, has been identified as a cost-effective means of mitigating climate change while ensuring food security. In China, cropland management has contributed significantly to reducing carbon emissions and enhancing ecosystem resilience [64]. These measures highlight the importance of integrating ecological preservation with urban development, balancing the need for growth with the imperative to protect and restore natural ecosystems. Policymakers and urban planners should focus on such multi-scalar strategies to mitigate the adverse effects of urbanization and promote sustainable development.

The findings of this study offer significant implications for urban planning in Saudi Arabia, particularly in addressing the trade-offs between urban expansion and ecosystem service preservation. Rapid urbanization has led to extensive land cover transformations, impacting essential ecosystem functions such as water regulation, carbon sequestration, and biodiversity conservation. These trade-offs necessitate a strategic approach to urban planning, ensuring that development policies incorporate ecosystem preservation frameworks. Sustainable urban growth strategies, including green infrastructure development, urban densification, and zoning regulations, can mitigate the adverse effects of urban sprawl. Moreover, urban planners must integrate remote sensing and GIS-based land use modeling to predict future urbanization trends and minimize ecological disruptions. By employing nature-based solutions and policies aimed at compact urban development, Saudi Arabia can balance economic growth with environmental sustainability, thereby fostering more resilient urban landscapes.

Despite the advantages of ML in urban growth analysis, certain limitations should be acknowledged. The accuracy of ML models depends on the quality and availability of training data, which can be constrained by inconsistencies in historical land use datasets and satellite image resolutions. Additionally, distinguishing spectral similarities between different land cover types, such as sparse vegetation and water bodies, remains a persistent challenge. The complexity of urban systems, including informal settlements and mixed land-use areas, further complicates classification accuracy. Furthermore, ML models, particularly those utilizing deep learning architectures, require significant computational resources and parameter tuning, which may limit their applicability in regions with restricted technological infrastructure. To address these limitations, hybrid modeling approaches that integrate ML with traditional urban growth models, such as Cellular Automata–Markov could enhance predictive performance. Future studies should focus on improving data preprocessing techniques, incorporating socioeconomic factors, and leveraging high-resolution imagery to refine urban expansion predictions. Recognizing these challenges will ensure that ML applications in urban analysis remain accurate, adaptable, and relevant to policy-making in rapidly developing regions such as Saudi Arabia.

The findings of this study offer significant implications for urban planning in Saudi Arabia, particularly in addressing the trade-offs between urban expansion and ecosystem service preservation. Rapid urbanization has led to extensive land cover transformations, impacting essential ecosystem functions such as water regulation, carbon sequestration, and biodiversity conservation. These trade-offs necessitate a strategic approach to urban planning, ensuring that development policies incorporate ecosystem preservation frameworks. Sustainable urban growth strategies, including green infrastructure development, urban densification, and zoning regulations, can mitigate the adverse effects of urban sprawl. Moreover, urban planners must integrate remote sensing and GIS-based land use modeling to predict future urbanization trends and minimize ecological disruptions. By employing nature-based solutions and policies aimed at compact urban development, Saudi Arabia can balance economic growth with environmental sustainability, thereby fostering more resilient urban landscapes. The practical applications of these findings for urban planners and policymakers are crucial in guiding future urban development in Saudi Arabia. Policymakers can use the study’s insights to implement targeted land use regulations that prevent excessive urban sprawl while preserving essential ecosystem services. For example, zoning policies can be refined to promote mixed-use developments that reduce land consumption and encourage sustainable urban mobility. The identification of high-risk areas for ecosystem degradation can inform the establishment of protected zones and conservation areas, ensuring long-term ecological resilience.

While the findings of this study provide valuable insights into urban expansion, land use dynamics, and ecosystem service trade-offs, several limitations must be acknowledged. One of the primary challenges pertains to potential biases in LULC classification, which may arise due to spectral similarities between certain land cover types. Despite the use of advanced machine learning and deep learning models, classification errors remain a possibility, particularly in distinguishing sparse vegetation from barren land and differentiating water bodies with seasonal fluctuations. The accuracy of classification is influenced by sensor resolution, atmospheric conditions, and variations in reflectance properties, which may introduce uncertainties in the interpretation of land cover transitions. Future studies could mitigate these issues by incorporating higher-resolution satellite imagery, multi-temporal data fusion techniques, and additional spectral indices to enhance classification accuracy. Another limitation relates to the generalizability of the study’s findings across different ecological and geographical contexts. The Markov transition model, while effective in capturing land cover changes and transition probabilities, assumes that past trends will continue into the future, potentially overlooking abrupt environmental shifts or anthropogenic interventions. Additionally, cross-regional comparisons of ecological restoration efforts present inherent challenges, as the factors influencing ecosystem recovery—such as climate variability, land management policies, and socio-economic drivers—differ significantly between regions. The extent to which restoration efforts in Abha can be extrapolated to other urban landscapes with varying ecological conditions remains uncertain. To improve cross-regional applicability, future research could incorporate multi-region case studies and employ hybrid modeling approaches that integrate socio-economic and climatic variables to refine land cover transition predictions. The resilience index method used to quantify ecosystem restoration presents another area for potential refinement. Although recovery metrics were derived based on the transition of land cover classes from built-up areas to natural land types, the ecological functionality of restored areas was not explicitly considered. Land cover conversion alone does not necessarily imply full ecosystem recovery, as factors such as biodiversity, soil quality, and hydrological processes influence the effectiveness of restoration efforts.

5. Conclusions

This study assesses urban expansion, LULC dynamics, and ecosystem service trade-offs in Abha, Saudi Arabia, using advanced ML, DL models, and the Markov transition model. Findings reveal significant barren land reduction and cropland expansion due to rapid urbanization, with peak ESV losses (USD 125,662.7) between 2017 and 2020, followed by partial recovery (USD 96,769.5) through land management. ML/DL models improved LULC classification accuracy, while Markov modeling provided probabilistic land transition insights. The increased Shannon entropy indicates increasing urban sprawl, underscoring the need for strategic land-use planning and sustainable management.

To balance development with sustainability, zoning regulations should restrict urban encroachment into ecologically sensitive areas. Green infrastructure (urban green spaces, reforestation) must be integrated to mitigate urban heat islands, while sustainable cropland management should prevent excessive agricultural expansion. Data-driven spatial planning, leveraging high-resolution satellite imagery and predictive modeling, should guide urbanization policies. Adaptive land management strategies must protect water bodies/vegetation via sustainable land-use incentives and mandatory EIAs. Community-based restoration efforts should be strengthened for conservation and afforestation. Despite strong ML and DL classification, challenges persist in differentiating sparse vegetation and seasonal water bodies, leading to potential biases. Future research should refine classification via multi-temporal fusion and spectral indices. Urban growth modeling can be enhanced with cellular automata and agent-based simulations for non-linear expansion analysis. Further studies on ecosystem recovery dynamics and long-term restoration efficiency are essential for sustainable urbanization.

Institutional Review Board Statement

Not applicable.

Informed Consent Statement

Not applicable.

Data Availability Statement

The datasets used and/or analyzed during the current study are available from the corresponding author on reasonable request.

Acknowledgments

The author is thankful to the undergraduate students of the College of Engineering, King Khalid University, for assisting in data collection.

Conflicts of Interest

The author declares no conflicts of interest.

Footnotes

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Figures and Tables

Figure 1. Study area.

Figure 2. Training and validation loss curves for a 1D CNN model.

Figure 3. Confusion matrices of ML and DL models for LULC classification evaluating RF, ANN, GBM, LightGBM, and 1D CNN models, highlighting classification accuracy and misclassification trends across land cover classes.

Figure 4. ROC curves and AUC values for Random Forest, ANN, GBM, LightGBM, and 1D CNN models for six land cover classes.

Figure 5. Spatiotemporal distribution of LULC classes for the years (a) 2014, (b) 2017, (c) 2020, and (d) 2023.

Figure 6. Land cover area for different classes for the years 2014, 2017, 2020, and 2023.

Figure 7. Probability-based Markov transition matrices depicting the dynamic land cover changes between 2014–2017, 2017–2020, 2020–2023, and overall, for 2014–2023, quantifying transformation trends among LULC categories.

Figure 8. Temporal analysis (2014 to 2023) showing trends in urban growth metrics over time, including urban growth rate (top left), Shannon’s entropy (top right), urban fragmentation (bottom left), and urban edge growth (bottom right), highlighting spatial and structural changes in urban expansion.

Table 1

Calculated ESV over different years based on the methodology of Costanza et al. (1997) [53].

Land Cover Class	2014 (USD/year)	2017 (USD/year)	2020 (USD/year)	2023 (USD/year)
Built-up	0	0	0	0
Dense Vegetation	9,654,678	11,826,021	11,707,465	14,621,217
Sparse Vegetation	6,911,224	11,098,752	9,872,528	11,079,632
Water Bodies	5,735,800	7,719,700	6,266,140	6,767,300
Barren Land	0	0	0	0
Cropland	1,565,472	3,230,672	2,517,396	3,453,960

Table 2

Ecosystem service loss due to urban sprawl (USD/year).

Year Interval	Dense Vegetation Loss	Sparse Vegetation Loss	Water Bodies Loss	Cropland Loss	Total Loss
2014–2017	−15,885,120	−41,810,560	−5,246,050	−16,648,720	−79,590,450
2017–2020	−1,196,340	−12,261,680	−14,491,190	−7,125,220	−35,074,430
2020–2023	−29,264,040	−12,773,600	−1,014,865	−9,366,840	−52,419,345

Table 3

Ecosystem service recovery following urban sprawl (USD/year).

Year Interval	Dense Vegetation Recovery	Sparse Vegetation Recovery	Water Bodies Recovery	Cropland Recovery	Total Recovery
2014–2017	+21,736,560	+41,882,720	+4,937,270	+16,648,720	+85,205,270
2017–2020	+1,076,610	+11,966,080	+15,008,620	+6,769,260	+34,820,570
2020–2023	+29,854,170	+12,899,920	+1,071,685	+9,958,520	+53,784,295

References

1. Ma, Z.; Duan, X.; Wang, L.; Wang, Y.; Kang, J.; Yun, R. A Scenario Simulation Study on the Impact of Urban Expansion on Terrestrial Carbon Storage in the Yangtze River Delta, China. Land; 2023; 12, 297. [DOI: https://dx.doi.org/10.3390/land12020297]

2. Zhong, J.; Liu, J.; Jiao, L.; Lian, X.; Xu, Z.; Zhou, Z. Assessing the Comprehensive Impacts of Different Urbanization Process on Vegetation Net Primary Productivity in Wuhan, China, from 1990 to 2020. Sustain. Cities Soc.; 2021; 75, 103295. [DOI: https://dx.doi.org/10.1016/j.scs.2021.103295]

3. Megersa, W.; Deribew, K.T.; Abreha, G.; Liqa, T.; Moisa, M.B.; Hailu, S.; Worku, K. Stochastic Modeling of Urban Growth Using the CA-Markov Chain and Multi-Scenario Prospects in the Tropical Humid Region of Ethiopia: Mettu. Geocarto Int.; 2023; 38, 2240285. [DOI: https://dx.doi.org/10.1080/10106049.2023.2240285]

4. Zhang, X.; Feng, T.; Zhao, S.; Yang, G.; Zhang, Q.; Qin, G.; Liu, L.; Long, X.; Sun, W.; Gao, C. et al. Elucidating the Impacts of Rapid Urban Expansion on Air Quality in the Yangtze River Delta, China. Sci. Total Environ.; 2021; 799, 149426. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2021.149426]

5. Li, S.; He, Y.; Xu, H.; Zhu, C.; Dong, B.; Lin, Y.; Si, B.; Deng, J.; Wang, K. Impacts of Urban Expansion Forms on Ecosystem Services in Urban Agglomerations: A Case Study of Shanghai-Hangzhou Bay Urban Agglomeration. Remote Sens.; 2021; 13, 1908. [DOI: https://dx.doi.org/10.3390/rs13101908]

6. Dhanaraj, K.; Angadi, D.P. Urban Expansion Quantification from Remote Sensing Data for Sustainable Land-Use Planning in Mangaluru, India. Remote Sens. Appl.; 2021; 23, 100602. [DOI: https://dx.doi.org/10.1016/j.rsase.2021.100602]

7. Huang, C.; Xu, N. Quantifying Urban Expansion from 1985 to 2018 in Large Cities Worldwide. Geocarto Int.; 2022; 37, pp. 18356-18371. [DOI: https://dx.doi.org/10.1080/10106049.2022.2142957]

8. Egidi, G.; Halbac-Cotoara-Zamfir, R.; Cividino, S.; Quaranta, G.; Salvati, L.; Colantoni, A. Rural in Town: Traditional Agriculture, Population Trends, and Long-Term Urban Expansion in Metropolitan Rome. Land; 2020; 9, 53. [DOI: https://dx.doi.org/10.3390/land9020053]

9. Rai, R.; Yili, Z.; Paudel, B.; Khanal, N.R.; Acharya, B.K. Satellite Image-Based Monitoring of Urban Land Use Change and Assessing the Driving Factors in Pokhara and Bharatpur Metropolitan Cities, Gandaki Basin, Nepal. J. Resour. Ecol.; 2020; 11, 87. [DOI: https://dx.doi.org/10.5814/j.issn.1674-764x.2020.01.009]

10. Gao, C.; Feng, Y.; Wang, R.; Lei, Z.; Chen, S.; Tang, X.; Xi, M. 50-Year Urban Expansion Patterns in Shanghai: Analysis Using Impervious Surface Data and Simulation Models. Land; 2023; 12, 2065. [DOI: https://dx.doi.org/10.3390/land12112065]

11. Tripathi, S. Do Macroeconomic Factors Promote Urbanization? Evidence from BRICS Countries. Asia-Pac. J. Reg. Sci.; 2020; 5, pp. 397-426. [DOI: https://dx.doi.org/10.1007/s41685-020-00179-z]

12. Gilbert, K.M.; Shi, Y. Urban Growth Monitoring and Prediction Using Remote Sensing Urban Monitoring Indices Approach and Integrating CA-Markov Model: A Case Study of Lagos City, Nigeria. Sustainability; 2023; 16, 30. [DOI: https://dx.doi.org/10.3390/su16010030]

13. Islam, M.A.; Rickty, T.A.; Das, P.K.; Haque, M.B. Modeling and Forecasting Urban Sprawl in Sylhet Sadar Using Remote Sensing Data. Proc. Eng. Technol. Innov.; 2023; 23, pp. 23-35. [DOI: https://dx.doi.org/10.46604/peti.2023.9617]

14. Huang, C.; Zhuang, Q.; Meng, X.; Guo, H.; Han, J. An Improved Nightlight Threshold Method for Revealing the Spatiotemporal Dynamics and Driving Forces of Urban Expansion in China. J. Environ. Manag.; 2021; 289, 112574. [DOI: https://dx.doi.org/10.1016/j.jenvman.2021.112574]

15. Xu, T.; Gao, J.; Coco, G.; Wang, S. Urban Expansion in Auckland, New Zealand: A GIS Simulation via an Intelligent Self-Adapting Multiscale Agent-Based Model. Int. J. Geogr. Inf. Sci.; 2020; 34, pp. 2136-2159. [DOI: https://dx.doi.org/10.1080/13658816.2020.1748192]

16. Tang, J.; Di, L. Past and future trajectories of farmland loss due to rapid urbanization using Landsat imagery and the Markov-CA model: A case study of Delhi, India. Remote Sens.; 2019; 11, 180. [DOI: https://dx.doi.org/10.3390/rs11020180]

17. Warrag, E.I.; Mallick, J.; Singh, R.K.; Khan, R.A. Status of Dieback of Juniperus Procera (African Pencil Cedar) in Natural Stands and Plantation in Alsouda Highlands, Saudi Arabia. Appl. Ecol. Environ. Res.; 2019; 17, pp. 2325-2338. [DOI: https://dx.doi.org/10.15666/aeer/1702_23252338]

18. Abulfatih, H.A. Wild Plants from Abha & the Surrounding Areas; Saudi Publishing & Distribution: Riyadh, Saudi Arabia, 1984; Available online: https://www.amazon.com/Wild-Plants-Abha-Surrounding-Areas/dp/B00113M1GQ (accessed on 14 March 2021).

19. Bindajam, A.A.; Mallick, J.; AlQadhi, S.; Singh, C.K.; Hang, H.T. Impacts of Vegetation and Topography on Land Surface Temperature Variability over the Semi-Arid Mountain Cities of Saudi Arabia. Atmosphere; 2020; 11, 762. [DOI: https://dx.doi.org/10.3390/atmos11070762]

20. da Silva, R.F.; Moran, E.; Viña, A.; Millington, J.D.A.; Dou, Y.; Vieira, S.A.; Lopez, M.C.; Liu, J. Toward a Forest Transition across the Brazilian Atlantic Forest Biome. Front. For. Glob. Change; 2023; 6, 1071495. [DOI: https://dx.doi.org/10.3389/ffgc.2023.1071495]

21. Ouma, Y.O.; Keitsile, A.; Nkwae, B.; Odirile, P.; Moalafhi, D.; Qi, J. Urban Land-Use Classification Using Machine Learning Classifiers: Comparative Evaluation and Post-Classification Multi-Feature Fusion Approach. Eur. J. Remote Sens.; 2023; 56, 2173659. [DOI: https://dx.doi.org/10.1080/22797254.2023.2173659]

22. Belete, D.M.; Huchaiah, M.D. Grid Search in Hyperparameter Optimization of Machine Learning Models for Prediction of HIV/AIDS Test Results. Int. J. Comput. Appl.; 2021; 44, pp. 875-886. [DOI: https://dx.doi.org/10.1080/1206212X.2021.1974663]

23. Siregar, A.A.; Soim, S.; Fadhli, M. Optimizing Malware Detection Using Back Propagation Neural Network and Hyperparameter Tuning. Indones. J. Artif. Intell. Data Min.; 2023; 6, 220. [DOI: https://dx.doi.org/10.24014/ijaidm.v6i2.24731]

24. Fan, M.; Xiao, K.; Sun, L.; Zhang, S.; Xu, Y. Automated Hyperparameter Optimization of Gradient Boosting Decision Tree Approach for Gold Mineral Prospectivity Mapping in the Xiong’ershan Area. Minerals; 2022; 12, 1621. [DOI: https://dx.doi.org/10.3390/min12121621]

25. Hartanto, A.D.; Nur Kholik, Y.; Pristyanto, Y. Stock Price Time Series Data Forecasting Using the Light Gradient Boosting Machine (LightGBM) Model. JOIV Int. J. Inform. Vis.; 2023; 7, pp. 2270-2279. [DOI: https://dx.doi.org/10.30630/joiv.7.4.1740]

26. Anggoro, D.A.; Mukti, S.S. Performance Comparison of Grid Search and Random Search Methods for Hyperparameter Tuning in Extreme Gradient Boosting Algorithm to Predict Chronic Kidney Failure. Int. J. Intell. Eng. Syst.; 2021; 14, pp. 198-207. [DOI: https://dx.doi.org/10.22266/IJIES2021.1231.19]

27. Swaminatha Rao, L.P.; Jaganathan, S. Adaptive Bayesian Contextual Hyperband: A Novel Hyperparameter Optimization Approach. IAES Int. J. Artif. Intell. (IJ-AI); 2024; 13, 775. [DOI: https://dx.doi.org/10.11591/ijai.v13.i1.pp775-785]

28. Wan, Q.; Xiao, Z.; Yu, Y.; Liu, Z.; Wang, K.; Li, D. A Hyperparameter-Free Attention Module Based on Feature Map Mathematical Calculation for Remote-Sensing Image Scene Classification. IEEE Trans. Geosci. Remote Sens.; 2024; 62, 5600318. [DOI: https://dx.doi.org/10.1109/TGRS.2023.3335627]

29. Thylashri, S.; Rajalakshmi, R.N. Wetland Mapping by Fusion of Deep Learning and Ensemble Model for Enhancing Prediction Outcomes. Fusion Pract. Appl.; 2024; 14, pp. 178-189. [DOI: https://dx.doi.org/10.54216/fpa.140115]

30. Arrechea-Castillo, D.A.; Solano-Correa, Y.T.; Muñoz-Ordóñez, J.F.; Pencue-Fierro, E.L.; Figueroa-Casas, A. Multiclass Land Use and Land Cover Classification of Andean Sub-Basins in Colombia with Sentinel-2 and Deep Learning. Remote Sens.; 2023; 15, 2521. [DOI: https://dx.doi.org/10.3390/rs15102521]

31. Mulenga, C.; Kaonga, P.; Hamoonga, R.; Mazaba, M.L.; Chabala, F.; Musonda, P. Predicting Mortality in Hospitalized COVID-19 Patients in Zambia: An Application of Machine Learning. Glob. Health; 2023; 2023, 8921220. [DOI: https://dx.doi.org/10.1155/2023/8921220]

32. Carrington, A.M.; Manuel, D.G.; Fieguth, P.W.; Ramsay, T.; Osmani, V.; Wernly, B.; Bennett, C.; Hawken, S.; Magwood, O.; Sheikh, Y. et al. Deep ROC Analysis and AUC as Balanced Average Accuracy, for Improved Classifier Selection, Audit and Explanation. IEEE Trans. Pattern Anal. Mach. Intell.; 2023; 45, pp. 329-341. [DOI: https://dx.doi.org/10.1109/TPAMI.2022.3145392] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35077357]

33. Taz, N.H.; Islam, A.; Mahmud, I. A Comparative Analysis of Ensemble Based Machine Learning Techniques for Diabetes Identification. Proceedings of the 2021 2nd International Conference on Robotics, Electrical and Signal Processing Techniques (ICREST); Dhaka, Bangladesh, 5–7 January 2021; pp. 1-6.

34. Nainggolan, S.P.; Sinaga, A. Comparative Analysis of Accuracy of Random Forest and Gradient Boosting Classifier Algorithm for Diabetes Classification. Sebatik; 2023; 27, pp. 97-102. [DOI: https://dx.doi.org/10.46984/sebatik.v27i1.2157]

35. Xu, J.; Cheng, J.; Yan, H.; Park, J.H.; Qi, W. Dynamic Event-Triggered Control for Semi-Markov Singularly Perturbed Systems With Generally Transition Rates. IEEE Trans. Syst. Man Cybern. Syst.; 2024; 54, pp. 225-235. [DOI: https://dx.doi.org/10.1109/TSMC.2023.3304968]

36. Kao, Y.; Han, Y.; Zhu, Y.; Shu, Z. Stability Analysis of Delayed Discrete Singular Piecewise Homogeneous Markovian Jump Systems With Unknown Transition Probabilities via Sliding-Mode Approach. IEEE Trans. Automat. Control; 2024; 69, pp. 315-322. [DOI: https://dx.doi.org/10.1109/TAC.2023.3262444]

37. Yousefi Kebriya, A.; Nadi, M. Examining Air Pollution Continuity in Tehran Province Using Markov Chain Model. Iran. J. Energy Environ.; 2024; 15, pp. 201-210. [DOI: https://dx.doi.org/10.5829/IJEE.2024.15.02.10]

38. Rai, K. Urban Sprawl Analysis Using GIS Application of Bhagur City. Int. J. Res. Appl. Sci. Eng. Technol.; 2021; 9, pp. 861-868. [DOI: https://dx.doi.org/10.22214/ijraset.2021.38077]

39. Das, S.; Angadi, D.P. Assessment of Urban Sprawl Using Landscape Metrics and Shannon’s Entropy Model Approach in Town Level of Barrackpore Sub-Divisional Region, India. Model. Earth Syst. Environ.; 2020; 7, pp. 1071-1095. [DOI: https://dx.doi.org/10.1007/s40808-020-00990-9]

40. Sarif, M.O.; Gupta, R.D. Comparative Evaluation between Shannon’s Entropy and Spatial Metrics in Exploring the Spatiotemporal Dynamics of Urban Morphology: A Case Study of Prayagraj City, India (1988–2018). Spat. Inf. Res.; 2021; 29, pp. 961-979. [DOI: https://dx.doi.org/10.1007/s41324-021-00406-5]

41. Getu, K.; Bhat, H.G. Analysis of Spatio-Temporal Dynamics of Urban Sprawl and Growth Pattern Using Geospatial Technologies and Landscape Metrics in Bahir Dar, Northwest Ethiopia. Land Use Policy; 2021; 109, 105676. [DOI: https://dx.doi.org/10.1016/j.landusepol.2021.105676]

42. Singh, L.; Singh, H.; Singh, S. Managing Natural Resources and Environmental Challenges in the Face of Urban Sprawl in Indian Himalayan City of Jammu. J. Indian Soc. Remote Sens.; 2020; 48, pp. 1139-1147. [DOI: https://dx.doi.org/10.1007/s12524-020-01133-4]

43. Al-Hameedi, W.M.M.; Chen, J.; Faichia, C.; Al-Shaibah, B.; Nath, B.; Kafy, A.-A.; Hu, G.; Al-Aizari, A. Remote Sensing-Based Urban Sprawl Modeling Using Multilayer Perceptron Neural Network Markov Chain in Baghdad, Iraq. Remote Sens.; 2021; 13, 4034. [DOI: https://dx.doi.org/10.3390/rs13204034]

44. Gong, Y.; Cai, M.; Yao, L.; Cheng, L.; Hao, C.; Zhao, Z. Assessing Changes in the Ecosystem Services Value in Response to Land-Use/Land-Cover Dynamics in Shanghai from 2000 to 2020. Int. J. Environ. Res. Public Health; 2022; 19, 12080. [DOI: https://dx.doi.org/10.3390/ijerph191912080] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/36231381]

45. Oloo, F.; Murithi, G.; Jepkosgei, C. Quantifying Tree Cover Loss in Urban Forests within Nairobi City Metropolitan Area from Earth Observation Data. Proceedings of the 1st International Electronic Conference on Forests—Forests for a Better Future: Sustainability, Innovation, Interdisciplinarity; Online, 15–30 November 2020; 78.

46. Lennert, J.; Farkas, J.Z.; Kovács, A.D.; Molnár, A.; Módos, R.; Baka, D.; Kovács, Z. Measuring and Predicting Long-Term Land Cover Changes in the Functional Urban Area of Budapest. Sustainability; 2020; 12, 3331. [DOI: https://dx.doi.org/10.3390/su12083331]

47. Suharyadi, R.; Umarhadi, D.A.; Awanda, D.; Widyatmanti, W. Exploring Built-Up Indices and Machine Learning Regressions for Multi-Temporal Building Density Monitoring Based on Landsat Series. Sensors; 2022; 22, 4716. [DOI: https://dx.doi.org/10.3390/s22134716] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/35808212]

48. Jafar, R.; Awad, A.; Hatem, I.; Jafar, K.; Awad, E.; Shahrour, I. Multiple Linear Regression and Machine Learning for Predicting the Drinking Water Quality Index in Al-Seine Lake. Smart Cities; 2023; 6, pp. 2807-2827. [DOI: https://dx.doi.org/10.3390/smartcities6050126]

49. Das, T.; Shahfahad,; Naikoo, M.W.; Talukdar, S.; Parvez, A.; Rahman, A.; Pal, S.; Asgher, M.S.; Islam, A.R.M.T.; Mosavi, A. Analysing Process and Probability of Built-Up Expansion Using Machine Learning and Fuzzy Logic in English Bazar, West Bengal. Remote Sens.; 2022; 14, 2349. [DOI: https://dx.doi.org/10.3390/rs14102349]

50. Dou, W.; Sun, S.; Xu, P. The Design of Jujube Irrigation System Using Linear Regression Analysis, BP Neural Network and Random Forest. INMATEH Agric. Eng.; 2023; 70, pp. 165-172. [DOI: https://dx.doi.org/10.35633/inmateh-70-16]

51. Egbueri, J.C.; Agbasi, J.C. Performances of MLR, RBF-NN, and MLP-NN in the Evaluation and Prediction of Water Resources Quality for Irrigation Purposes under Two Modeling Scenarios. Geocarto Int.; 2022; 37, pp. 14399-14431. [DOI: https://dx.doi.org/10.1080/10106049.2022.2087758]

52. Hu, Y.; Wei, F.; Fu, B.; Zhang, W.; Sun, C. Ecosystems in China Have Become More Sensitive to Changes in Water Demand since 2001. Commun. Earth Environ.; 2023; 4, 444. [DOI: https://dx.doi.org/10.1038/s43247-023-01105-9]

53. Costanza, R.; d’Arge, R.; De Groot, R.; Farber, S.; Grasso, M.; Hannon, B.; Limburg, K.; Naeem, S.; O’neill, R.V.; Paruelo, J. et al. The value of the world’s ecosystem services and natural capital. Nature; 1997; 387, pp. 253-260. [DOI: https://dx.doi.org/10.1038/387253a0]

54. Liu, Z.; Liu, Z.; Zhou, Y.; Huang, Q. Distinguishing the Impacts of Rapid Urbanization on Ecosystem Service Trade-Offs and Synergies: A Case Study of Shenzhen, China. Remote Sens.; 2022; 14, 4604. [DOI: https://dx.doi.org/10.3390/rs14184604]

55. Marino, D.; Barone, A.; Marucci, A.; Pili, S.; Palmieri, M. Impact of Land Use Changes on Ecosystem Services Supply: A Meta Analysis of the Italian Context. Land; 2023; 12, 2173. [DOI: https://dx.doi.org/10.3390/land12122173]

56. Yang, S.; Su, H. Multi-Scenario Simulation of Ecosystem Service Values in the Guanzhong Plain Urban Agglomeration, China. Sustainability; 2022; 14, 8812. [DOI: https://dx.doi.org/10.3390/su14148812]

57. Lahon, D.; Sahariah, D.; Debnath, J.; Nath, N.; Meraj, G.; Kumar, P.; Hashimoto, S.; Farooq, M. Assessment of Ecosystem Service Value in Response to LULC Changes Using Geospatial Techniques: A Case Study in the Merbil Wetland of the Brahmaputra Valley, Assam, India. ISPRS Int. J. Geo-Inf.; 2023; 12, 165. [DOI: https://dx.doi.org/10.3390/ijgi12040165]

58. Nandiyanto, A.B.D.; Hofifah, S.N. How to Conduct Paired-t-Test SPSS: Comprehension in Adsorption with Bibliometric. Int. J. Eval. Res. Educ. (IJERE); 2024; 13, 151. [DOI: https://dx.doi.org/10.11591/ijere.v13i1.27541]

59. Wu, Y.; Huang, Z.; Han, D.; Qiu, X.; Pan, Y. Evolution of Urban Ecosystem Service Value and a Scenario Analysis Based on Land Utilization Changes: A Case Study of Hangzhou, China. Sustainability; 2023; 15, 8274. [DOI: https://dx.doi.org/10.3390/su15108274]

60. Hirpa, B.A.; Adane, G.B.; Asrat, A.; Nedaw, D.; Song, C.; Roh, M.; Lee, W.-K. Urban Sprawl at the Expense of Cultivated Land: Decadal Land Use and Land Cover Changes and Future Projections in the Upper Awash Basin of Central Ethiopia. Front. Ecol. Evol.; 2023; 11, 1160987. [DOI: https://dx.doi.org/10.3389/fevo.2023.1160987]

61. Asadi, M.; Oshnooei-Nooshabadi, A.; Saleh, S.-S.; Habibnezhad, F.; Sarafraz-Asbagh, S.; Van Genderen, J.L. Urban Sprawl Simulation Mapping of Urmia (Iran) by Comparison of Cellular Automata–Markov Chain and Artificial Neural Network (ANN) Modeling Approach. Sustainability; 2022; 14, 15625. [DOI: https://dx.doi.org/10.3390/su142315625]

62. Wu, Y.; Gu, C.; Zhang, Y. Towards Sustainable Management of Urban Ecological Space: A Zoning Approach Hybridized by Ecosystem Service Value and Ecological Risk Assessment. Land; 2022; 11, 1220. [DOI: https://dx.doi.org/10.3390/land11081220]

63. Chen, H.; Dong, N.; Liang, X.; Huang, H. Spatiotemporal Evaluation of Regional Land Use Dynamics and Its Potential Ecosystem Impact under Carbon Neutral Pathways in the Guangdong–Hong Kong–Macao Greater Bay Area. Remote Sens.; 2023; 15, 5749. [DOI: https://dx.doi.org/10.3390/rs15245749]

64. Zheng, H.; Guo, L.; He, B.; Shen, Y.; Yuan, W.; Yang, W.; Li, T.; Zheng, H. The Great Climate Mitigation Potential of Cropland Ecosystem Management in China. Earths Future; 2023; 11, e2023EF003586. [DOI: https://dx.doi.org/10.1029/2023EF003586]