A data-driven approach to forest health assessment through multivariate analysis and machine learning techniques

Abstract

Background

Himalayan forests are fragile, rich in biodiversity, and face increasing threats from anthropogenic pressures and climate change. Assessing their health is critical for sustainable forest management. This study integrated ecological indicators (tree density, size, regeneration, deforestation, slope, grazing, and erosion) with machine learning (ML) to classify forest health and identify key drivers across 37 Western Himalayan sites. Principal component analysis (PCA) reduced data dimensionality, highlighting major ecological gradients. K-means clustering was used to group forests into three distinct classes based on ecological characteristics, due to its efficiency in identifying natural patterns within multivariate data. ML models, including Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM) were trained and validated using an 80:20 train-test split and 5-fold cross-validation.

Results

PCA revealed that elevation, disturbance, and regeneration explained 74.3% variance. Forest health varied across sites, with 10 categorized as healthy, 19 as moderate, and 8 as unhealthy. Forest regeneration was highly skewed (2.67) and leptokurtic (9.8), with few sites showing high seedling abundance, while deforestation (mean = 294 stumps ha⁻¹) indicated uneven human impact. Among ML models, RF showed the best performance with a mean accuracy of 0.83, Kappa 0.87, and balanced accuracy 0.88. SVM followed with 0.75 accuracy, Kappa 0.70, and balanced accuracy 0.81. DT performed lowest with 0.66 accuracy and Kappa 0.45. Cross-validation confirmed RF’s highest mean accuracy (90.3%), followed by SVM (88.1%) and DT (65.1%). RF-based feature importance analysis showed tree DBH, height, regeneration rate, soil erosion, and tree density as key ecological drivers of forest health.

Conclusions

This study highlights ML-driven classification as a precise, scalable, and objective tool for large-scale forest health assessments. Conservation efforts should prioritize degraded forests through afforestation, slope stabilization, controlled grazing, erosion management, and continuous ecosystem monitoring. Future studies should integrate high-resolution remote sensing (e.g., Landsat, Sentinel-2) and climate datasets (e.g., temperature, precipitation, and drought indices) to enhance predictive capabilities and support long-term forest management planning. The findings underscore the value of data-driven approaches, establishing machine learning as an effective tool to enhance forest monitoring and support evidence-based forest conservation and management in the Himalayas.

Introduction

Forest ecosystems are vital for ecological balance, providing essential services such as carbon sequestration, water regulation, and soil stabilization [1,2,3,4]. In the Western Himalayas, forest ecosystems are increasingly threatened by land use changes, overexploitation of forest resources, and infrastructure development, which degrade forest structure and hinder natural regeneration [2,3,4,5]. Current climate change combined with human impacts disrupts temperature and rainfall patterns, leading to shifts in species distribution, reduced seedling survival, and accelerated ecosystem decline [3, 4].

Forest degradation in the Himalayas is driven by both natural and human-induced factors, leading to biodiversity loss and soil erosion [4,5,6]. Environmental factors such as elevation, slope, and soil stability significantly influence forest structure and regeneration [7,8,9,10]. These variables determine species composition, soil moisture retention, and ecosystem resilience. Elevation affects temperature, moisture availability, and species distribution [11, 12]. The steep Himalayan slopes are prone to landslides and erosion, especially where deforestation, overgrazing and soil erosion occur. These disturbances deplete nutrients, reduce water retention, and hinder tree growth, lowering ecosystem productivity [6,7,8,9,10]. Such conditions increase vulnerability to drought, promote pest outbreaks, and raise tree mortality rates, destabilizing forest ecosystems [13].

Among anthropogenic factors, unregulated logging, fuelwood collection, and overgrazing are well known to reduce tree density, hinder regeneration, and lower forest productivity [1, 4]. High tree stump density indicates deforestation intensity, while overgrazing compacts soil, damages seedlings, and alters species composition, further reducing biodiversity [2, 5]. Similarly, human settlements contribute to forest degradation through overextraction of timber and nontimber forest products, trampling, and introduction of invasive species, which damage natural forest regeneration and accelerate biodiversity loss [13,14,15].

From 1990 to 2020, Landsat imagery showed a decrease in forest cover by 74 km² in AJK, with the steepest decline (29.92 km²) happening between 2000 and 2010. Forest fragmentation became more pronounced, as large contiguous areas (> 500 acres) were divided into smaller patches (< 250 acres), which poses risks to biodiversity and ecosystem stability [5]. Major deforestation drivers include energy shortages leading to firewood dependence, population-driven land conversion, and unregulated forest resource extraction. Solutions to deforestation involve increasing access to alternative clean energy, improving forest management, and enforcing stricter regulations against illegal timber harvesting [7, 8].

Precise assessment of forest health is essential for effective ecosystem management and long-term biodiversity conservation [16]. Traditional field assessments can be combined with advanced machine learning techniques to improve forest ecosystem classification by utilizing environmental and structural indicators for comprehensive and accurate analysis [17]. Among widely used machine learnin models, Decision Trees are easy to interpret but prone to overfitting, while Random Forest improves accuracy and robustness by aggregating multiple trees. Support Vector Machines provide precise, margin-based classification but can struggle with overlapping classes and imbalanced data [18]. Machine learning is increasingly used in ecological assessments to predict ecosystem responses to environmental change [19, 20] and optimize environmental monitoring [21]. Machine learning facilitates classification of ecological status, biodiversity monitoring, and sustainability assessments [22]. Machine learning’s integration with remote sensing and big data is enhancing accuracy and scalability, though challenges remain in data quality and model transparency [23,24,25].

Despite ecological importance, forest health assessments in the Western Himalayas remain limited and largely non-quantitative. To fill this gap, the current study assessed forest health in the Kashmir region of the Western Himalayas by analyzing field-based indicators reflecting forest condition, regeneration capacity, and degradation processes. Although machine learning has seen growing application in ecological research, its use in forest health assessment within Himalayan ecosystems, particularly in remote and data-scarce areas, remains limited. This study is among the first to apply and compare multiple machine learning models for classifying forest health in the Western Himalayas, offering a novel, data-driven approach to understanding ecosystem condition in this ecologically sensitive region. The primary aims were to identify key field indicators of forest health, evaluate the predictive performance of different machine learning models, and determine the most accurate and reliable model for health classification.

Methodology

Study area

Azad Jammu and Kashmir (AJK) is situated in the Western Himalayas between 33°−36°N latitude and 73°−75°E longitude (Fig. 1), covering an area of 13,297 km2. It shares borders with Punjab and Khyber Pakhtunkhwa (KP) in the south and west, while the Indian-administered region of Kashmir lies to the east. The region consists of parallel highlands extending along the Kaghan Valley (KP) and the Astor District (Gilgit-Baltistan) and is characterized by a predominantly mountainous landscape with steep valleys and rocky outcrops. Soil types range from nutrient-poor upland soils and shallow loamy slopes with low organic matter to fertile alluvial soils in intermontane valleys [26, 27].

[IMAGE OMITTED: SEE PDF]

Climatic zones in AJK range from subtropical monsoon type at lower elevation to moist temperate, subalpine, and alpine zones at higher elevations. The lowlands experience hot summers, whereas the highlands maintain mild temperatures. Winters are cold, with heavy snowfall above 1200 m between November and April. Summer temperatures range between 25 °C and 34 °C, and winter temperatures between 4 °C and 10 °C. Alpine regions remain snow-covered from September to May, with temperatures dropping below 0 °C. Annual rainfall averages around 1300 mm, mainly during the monsoon season (July to September). Monsoon areas receive between 900 and 1300 mm, while non-monsoon areas get much less, about 30 to 140 mm [27, 28].

Site selection and sampling

A total of 37 forest sites were sampled across different ecological zones in AJK, covering subtropical, temperate, and subalpine forests (Fig. 1). The ecological zones were categorized based on elevation and dominant flora recorded at each site (Table 1). The subtropical forests, comprising pine and mixed stands, were surveyed at 15 sites across four districts, ranging from 344 to 1425 m above sea level (a.s.l.). Temperate forests, characterized by mixed broadleaf and coniferous species, were sampled at 16 sites in five districts between 1933 and 2700 m a.s.l. In the subalpine zone, six sites across three districts were surveyed at elevations of 2790–3300 m.

Field surveys were conducted to evaluate the structural and regenerative status of forests. Elevation was measured using a GPS device to capture altitudinal variations. Tree density was estimated using systematic quadrat sampling, with five 10 × 10 m plots [23,24,25] placed at 100 m intervals per site. Tree diameter at breast height (DBH ≥ 10 cm) was recorded [29] using tape measurement, while tree height was recorded [30, 31] with a “Nikon Forestry Pro” laser range finder.

Plant species were identified in the field by the taxonomists, Dr. TH and Dr. HS. As no plant material was collected, voucher specimens were not deposited. The plant nomenclature was verified using the Flora of Pakistan (https://www.tropicos.org/Project/Pakistan) and Plants of the World Online (https://powo.science.kew.org). Seedling recruitment was quantified by counting tree seedlings (< 1.5 m in height) within five random subplots (5 × 5 m) to record fine-scale regeneration dynamics [32]. Degradation indicators assessed in the study included deforestation, erosion, grazing, and slope. Deforestation intensity was evaluated by recording stump presence within the primary 10 × 10 m sampling plots [33,34,35,36].

Soil erosion was classified into three levels: low (minimal or no visible erosion), moderate, and high, based on visual indicators such as soil displacement, rills or gullies, and ground cover condition [27]. Grazing intensity and slope inclination were assessed using the same three-tier approach. Grazing was estimated from field evidence, including browsing marks, hoof prints, dung, trampling, and understory damage, and categorized as low (few or no signs), moderate (intermittent damage, scattered dung, light trampling), or high (frequent damage, extensive trampling, abundant dung). Slope inclination was measured with a clinometer and classified as low (0–30°), moderate (31–45°), or high (> 45°), reflecting increasing erosion risk with steeper slopes.

[IMAGE OMITTED: SEE PDF]

Data preparation and analysis

Key ecological variables, including tree density, species richness, mean tree DBH and height, regeneration, deforestation, slope, grazing, and erosion, were scaled before analysis. Any non-numeric identifiers were excluded from the analysis. All numerical variables were scaled and centered where required, and missing values were removed via complete case analysis. Ordinal variables were numerically encoded before inclusion in the principal component analysis (PCA). For ML models, categorical and ordinal variables were factor-encoded as appropriate. The dataset was analyzed using correlations, PCA, and ML techniques in R (version 4.4.3) software [37]. Descriptive statistical analysis was carried out in Past (version 5.0.2) software [38].

Correlation analysis, PCA and forest health clustering

Pearson correlation coefficients among ecological variables were also computed. A complementary heatmap was generated to visualize variable interactions. PCA was conducted with standardized values to reduce dimensionality and explore variable interrelationships. Biplot visualizations were obtained, and PCA outputs were further used for exploratory clustering analysis. Similarly, to analyze ecological variations across elevation gradients, sites were classified into distinct elevation zones using a case-wise conditional transformation. The following elevation zones were defined based on vegetation: Subtropical (≤ 1425 m), Temperate (1426–2700 m), and Subalpine (2701–3300 m). To define site-level forest health conditions, all variables were scaled before clustering. K-means clustering was performed on nine indicators, including tree density, species richness, mean DBH, mean height, regeneration, deforestation, slope, grazing, and erosion. The optimal number of clusters (k = 3) was determined using the Elbow method [39] and visualized with convex hull plots.

Machine learning model development and evaluation

The study utilized three supervised machine learning algorithms, Decision Tree (DT), Random Forest (RF), and Support Vector Machine (SVM). To ensure reliable model evaluation, the dataset was split using a stratified 80:20 train-test partition based on the target variable (forest health status), preserving the proportional representation of each health class [18, 40]. Hyperparameter tuning was performed through repeated 5-fold cross-validation [40] using automated grid search (in the caret package). For the DT model, the complexity parameter (cp.) was tuned across standard values (0.01, 0.05, 0.1) to optimize the trade-off between model interpretability and accuracy. The RF model was configured with 500 trees, selected after testing for error stabilization, while the number of variables randomly selected at each split was tuned within the range of √p to p (where p is the total number of predictors). For the SVM model, a radial basis function kernel was used, with hyperparameters C and sigma optimized via grid search, along with feature scaling to normalize input variables [41,42,43,44].

Model performance was assessed using a 20% holdout test set and a confusion matrix. A resampling-based comparison was also conducted across models to identify the most effective approach. The best-performing model’s interpretability was further enhanced through variable importance analysis and cluster validation in PCA-reduced space. Predicted forest health labels from all models were appended to the original dataset, facilitating spatial overlay and comparative analyses. Bar plots were used to visualize the agreement between predicted and cluster-defined forest health classes [43, 44]. Finally, a feature importance analysis was also conducted based on the best model (RF) to elucidate the drivers [45] influencing forest health. The overall methodology adopted for this study is summarized in the figure below (Fig. 2).

[IMAGE OMITTED: SEE PDF]

Results

Descriptive analysis revealed pronounced variability in forest structural and ecological attributes across the study area (Table 2). Tree density, species richness, DBH, and tree height exhibited substantial heterogeneity, reflecting diverse stand structures. Regeneration showed particularly high variability and a positively skewed (most sites had low regeneration, with few sites showing very high values) and leptokurtic distribution (a heavy-tailed pattern, indicating sharp peaks and extreme values). Deforestation intensity was also highly variable, underscoring uneven anthropogenic pressure across sites.

[IMAGE OMITTED: SEE PDF]

Correlation patterns of forest attributes by site and environmental factors

A strong positive correlation was found between mean DBH and mean tree height, indicating consistent structural development. Elevation showed moderate positive correlations with both DBH and tree height, suggesting altitudinal influence on forest structure. Tree density was negatively correlated with DBH and height, reflecting density-related growth limitations. Other correlations among grazing, deforestation, and slope were generally weak, and no strong negative relationships were observed (Fig. 3).

[IMAGE OMITTED: SEE PDF]

Principal component analysis of ecological gradients

PCA was conducted to identify the dominant ecological gradients within the dataset and to reduce its dimensionality while preserving the most significant variance. The first principal component (PC1) accounted for 33.9% of the total variance, followed by PC2 (15.9%), PC3 (13.6%), and PC4 (10.9%). Collectively, these four components explained 74.3% of the total variance, indicating that a relatively low-dimensional subspace effectively summarized the major patterns of ecological variation within the study area.

The PCA biplot revealed the distribution of sampling sites along PC1 and PC2, with loading vectors representing the contribution of each ecological variable. Elevation, tree density, slope inclination, grazing intensity, soil erosion, deforestation, and regeneration exhibited strong loadings on PC1, suggesting that this component primarily captured gradients associated with site elevation, forest structure, and disturbance intensity. In contrast, tree species richness, mean DBH, and mean tree height were loaded onto PC2, reflecting their stronger association with tree size and canopy structure rather than environmental disturbance (Fig. 4).

[IMAGE OMITTED: SEE PDF]

Forest type-specific structure and regeneration across elevation and distance gradients

The relationship between tree density and seedling recruitment varied across forest types (Fig. 5). In subalpine forests, mean tree density showed a strong negative correlation with seedling recruitment (r = −0.804, p < 0.01) while this correlation was weaker (r = −0.312) in subtropical forests. Temperate forests exhibited a slightly positive correlation (r = 0.106) between tree density and regeneration. The influence of past disturbances on seedling regeneration also varied. In temperate forests, a positive correlation between stump presence and seedling recruitment (r = 0.570, p < 0.05) was observed. In subtropical forests, a negative correlation was found (r = −0.357), while subalpine forests showed a weak association (r = 0.153).

[IMAGE OMITTED: SEE PDF]

Tree density demonstrated a negative correlation with increasing distance from forest human settlements (r = −0.370, p < 0.05). Seedling recruitment exhibited a weak positive trend (r = 0.020), while stump presence declined with distance (r = −0.051). The relationship between tree density and stump presence varied across forest types. In subtropical forests, a strong positive correlation (r = 0.636, p < 0.05) indicated high biomass loss due to logging. In subalpine forests, a negative correlation (r = −0.363) showed significant forest cover reduction. In temperate forests, the correlation was weak (r = −0.108), indicating a more balanced disturbance-regeneration dynamics (Fig. 5).

Cluster analysis of forest health classifications

The K-means cluster analysis in PCA space identified three distinct groups: Healthy, Moderate, and Unhealthy sites, each representing different levels of forest structural integrity, regeneration potential, and environmental resilience. The healthy cluster (cluster 1) included 10 sites, characterized by higher tree density, diverse species composition, and minimal disturbance. These forests showed low deforestation rates, limited grazing pressure, stable soil conditions, high seedling recruitment, and strong ecological resilience. Sites such as Sia Azad Pattan, Shahdara, and Sudhan Galli belonged to this category (Fig. 6).

[IMAGE OMITTED: SEE PDF]

The moderate cluster (cluster 2) consisted of 19 sites with intermediate forest health. These forests maintained relatively stable tree density and regeneration levels but showed signs of environmental stress or past disturbances. Locations such as Sakhi Abad Kakrol, Taobut, and Dao Khan exhibited mixed effects of deforestation, grazing, and soil erosion. The unhealthy cluster (cluster 3) included 8 sites, such as Jarri Kas, Mirpur Sector C1, and Mehmood Gali, showing clear signs of ecological degradation (Fig. 6). These forests were marked by lower regeneration rates, higher deforestation impacts, intensive grazing pressure, and pronounced soil erosion. The clustering results indicated that these sites face significant ecological challenges, declining forest health, and reduced resilience against disturbances.

Model performance evaluation and comparison

The classification performance of three machine learning models (DT, RF, and SVM) showed that the RF model performed best, achieving the highest overall mean accuracy (0.83), a strong Kappa statistic (0.87), and balanced accuracy (0.88), indicating excellent agreement between predicted and actual forest health classes. The SVM model achieved an accuracy of 0.75. The confusion matrix showed one misclassification of a Healthy site as Unhealthy, while the rest were correctly predicted. A Kappa of 0.70 and balanced accuracy of 0.81 reflected good classification performance with minor errors. The DT model had the lowest performance, with an accuracy of 0.66. It correctly classified two Healthy and two Moderate classes but failed to identify any Unhealthy sites. The Kappa statistic (0.45) indicated moderate-to-low agreement. Mean sensitivity and specificity were 0.69 and 0.62, respectively, while 0 sensitivity for the Unhealthy class highlighted the model’s failure to detect this category.

Cross-validation also confirmed that the RF model had the highest mean accuracy (90.3%), with values ranging from 80 to 100%, demonstrating consistent and superior performance. The SVM model followed closely with a mean accuracy of 88.11% (range: 66.7 − 100%), showing strong generalization capability but slightly more variability. The DT model exhibited the lowest mean accuracy (65.1%), with wider variability (range: 33.3 − 85.7%), indicating inconsistent classification performance (Fig. 7).

[IMAGE OMITTED: SEE PDF]

Feature importance analysis

Analysis of feature importance based on Random Forest (RF) showed that mean DBH emerged as the most influential factor, with an importance score of 99, followed by tree height (87.83), regeneration (40.27), soil erosion (31.48), and tree density (30.85). Moderate levels of importance were observed for the number of tree species (26.17), slope gradient (22.63), and grazing pressure (21.65). Deforestation (15.72) and elevation (11.61) exhibited comparatively lower importance (Fig. 8), although they are still relevant within the multivariate context.

[IMAGE OMITTED: SEE PDF]

Discussion

Forest structure and ecological dynamics in the Western Himalayas

This study examined forest structure and ecological dynamics by investigating the role of elevation and environmental factors in shaping species composition and forest health. Tree density and species richness decline at higher elevations due to harsher climates, shorter growing seasons, and nutrient limitations, while lower elevations support greater diversity under more favourable conditions [13, 46]. Tree sizes and structure indicate dynamic forest composition where larger trees characterize undisturbed sites, while an abundance of smaller trees suggests regeneration following past disturbances [47].

Although overall regeneration patterns appear moderate, site-specific variations are influenced by light availability, soil fertility, and disturbance history, which collectively shape forest ecosystem dynamics. Anthropogenic activities, particularly deforestation and logging near settlements, significantly alter forest structure, reduce ecological resilience, and shift species composition [3, 48]. In Himalayan forests, pressures such as steep slopes, grazing, and soil erosion contribute to habitat instability and suppress regeneration [14].

The disturbance-regeneration relationship is strongly shaped by species traits, canopy structure, and microclimatic conditions. In subalpine forests, shade-intolerant conifers with conservative growth strategies (e.g., Abies pindrow) experience regeneration constraints under high tree density, where limited light and prolonged cold exposure reduce establishment success. Temperate forests, composed of mixed species and moderate canopy cover, exhibit a more balanced regeneration response due to intermediate shade tolerance and trait variability. In subtropical forests, fast-growing, shade-tolerant species regenerate quickly, but intensive logging near human settlements reduces seed sources and degrades soil, weakening regenerative capacity. The study reveals the complexity of ecosystem interactions and highlights the need for integrated conservation strategies that balance ecological sustainability with socio-economic demands. Long-term monitoring and adaptive management are crucial for preserving these forests amid ongoing climatic and anthropogenic pressures [37, 49].

A decline in tree density, species richness, and regeneration with increasing elevation reflects the structural dynamics of mountain forest ecosystems. Elevation influences temperature, moisture, and soil properties, shaping forest composition and creating ecological mosaics [10, 13] which ultimately impacts tree density. Steep declines with elevation indicate reduced forest resilience, biodiversity, and regeneration capacity, highlighting the need for elevation-specific management to sustain ecosystem functioning and recovery.

Despite site-specific variations, tree height and DBH remain relatively stable due to species functional traits. While species richness and deforestation show moderate heterogeneity, selective logging impacts vary, with some areas experiencing significant disturbance while others remain relatively undisturbed [50]. Areas with low regeneration require targeted reforestation, while high-deforestation zones necessitate stricter logging controls [51]. Effective slope management, along with controlling grazing and erosion, is vital for sustainable land use. While mature trees safeguard ecological stability, effective conservation requires an interdisciplinary understanding of the drivers of deforestation and forest degradation [52]. Integrating ecological assessments supports this by linking forest dynamics with climate and land-use monitoring to guide sustainable management [53].

Influence of elevation, disturbance, and growth patterns on forest structure

The strong positive correlation between DBH and tree height reflects typical allometric relationships, where taller trees accumulate biomass, enhancing forest productivity, carbon sequestration, and habitat stability [54]. Elevation moderately affects tree DBH, height, and crown architecture by imposing climatic stress and limiting nutrient availability, which restricts growth in harsher high-elevation environments [10, 13]. Reduced competition allows certain trees to grow larger by providing easier access to vital resources like sunlight, water, and nutrients [47]. Correlations between elevation and human activities such as grazing and deforestation suggest complex interactions influenced by accessibility and land-use practices. Lower elevations experience greater human disturbance, whereas selective resource extraction persists at higher elevations [55].

A weak negative correlation between tree density and size (DBH and height) indicates that denser forests often contain younger or smaller trees, likely reflecting post-disturbance regeneration or species-specific adaptations [56]. The negligible correlation between deforestation, grazing, and tree size implies that their immediate effects on forest structure may be influenced by long-term ecological processes that buffer short-term disturbances. Analysis specifies a quantitative perspective on these dynamics, indicating the need for site-specific strategies to maintain ecological balance while addressing human pressures [57, 58].

Forest structure and disturbance dynamics

In each forest type, tree density differently influences seedling recruitment. A strong negative correlation in subalpine forests may indicate that limited forest extent, coupled with climatic, ecological, and anthropogenic pressures, constrain recruitment. In contrast, a slight positive correlation in temperate forests may suggest that more favourable conditions, including better forest extent, canopy cover, and moderated microclimate, facilitate seedling recruitment [46, 59]. Disturbance effects vary by forest types. In temperate forests, moderate and selective logging can promote regeneration by creating light-rich canopy gaps and reducing competition, enhancing structural diversity [60]. In subtropical forests, soil degradation and seed bank loss from logging hinder recruitment. In subalpine forests, weak disturbance-regeneration links suggest environmental constraints like low temperatures and erosion-prone, poor soils are more limiting than disturbance intensity [61, 62].

Literature review revealed that in the Garhwal Himalaya, tree density for Cedrus deodara can reach about 700 trees ha−1, with saplings around 250 ha−1 and seedlings about 160 ha−1, indicating relatively good regeneration. Quercus leucotrichophora at lower altitudes shows about 90 trees ha−1, but saplings (250 ha−1) and seedlings (160 ha−1) densities suggest active regeneration, though still less than mature trees [63]. Similarly, Pinus roxburghii and mixed oak-pine forests show highly reduced sapling and seedling counts, likely due to fire and grazing [64]. Other factors, including poor seed viability, shade intolerance, slow growth, nutrient-poor soils, and low moisture, also hinder seedling establishment. Similarly, harsh climates and short growing seasons make regeneration sparse and disturbance-sensitive [64,65,66].

Spatial factors further shape forest dynamics. Tree density declines with distance from settlements due to increased human pressures like grazing and firewood collection, while seedling recruitment shows a weak positive trend, likely from reduced competition [3, 47]. Stump presence decreases with distance, indicating both higher past and recent logging intensity near accessible areas, highlighting the need for sustainable harvesting regulations. Logging impacts differ across forest types [67]. In subtropical forests, the strong correlation between tree density and stump presence implies substantial structural modifications driven by anthropogenic activities, necessitating reforestation and sustainable harvesting strategies [7, 28, 48, 52, 68].

In subalpine forests, the negative correlation indicates that logging-induced canopy loss is not offset by sufficient regeneration, emphasizing the need for stricter conservation measures. In temperate forests, the weak negative correlation implies a relatively stable regeneration-logging dynamic, though continuous monitoring is essential to assess long-term sustainability [50, 51, 68, 69]. This study found that many correlation patterns are statistically limited but still hold ecological significance by highlighting either mild or long-term processes shaping forest structure. The findings highlight the need for canopy management and regeneration support in subalpine forests, sustainable harvesting and soil stabilization in subtropical forests, and gap-based management in temperate forests. Stricter logging controls, invasive species management, and community-driven conservation are essential for long-term monitoring and ecological stability [5, 16, 46, 58].

Forest health classification and machine learning integration

This study utilized ordination and classification techniques to examine ecological variables and spatial distribution in Western Himalayan forests. PCA ordination revealed two key dimensions shaping forest structure: one driven by environmental and anthropogenic disturbances (elevation, tree density, slope, grazing, erosion, deforestation, and regeneration), indicating increased disturbances at higher elevations; the other capturing biological attributes (species richness, mean DBH, and height), where greater species richness and larger trees signify mature, resilient forests essential for biodiversity. Sites distributed along PC1 experience intensified disturbances, necessitating targeted conservation efforts. Conversely, sites distributed along PC2 represent mature forests that serve as biodiversity benchmarks. Conservation strategies should prioritize reforestation, soil stabilization, and controlled grazing in degraded areas, while protecting biodiversity-rich forests and preserving large remnant trees is crucial for sustaining long-term ecological resilience and structural integrity [2, 47, 57, 67].

Cluster analysis categorized forest sites into healthy, moderate, and unhealthy classes, reflecting forest resilience and disturbance levels. Healthy forests, characterized by high tree density, species richness, and minimal human interference, indicate strong ecosystem stability. Moderately healthy forests show early degradation signs from grazing and selective deforestation but retain recovery potential. Unhealthy forests, marked by severe deforestation, erosion, and low regeneration, require urgent intervention through reforestation, soil stabilization, and regulated land use [68, 69].

ML enhances forest health classification by improving accuracy, automating monitoring, and predicting degradation trends. DT and RF models effectively model ecological interactions and identify key degradation factors, while SVMs perform well with limited data to capture complex environmental relationships [70, 71]. Unsupervised learning, such as K-means clustering, aids large-scale forest segmentation for conservation planning [72, 73]. ML and remote sensing enable real-time monitoring of forest degradation, regeneration, and illegal logging, while predictive models assess climate change impacts to support adaptive conservation. AI-powered mobile tools improve field data collection and analysis. To maximize their potential, stronger integration between remote sensing data and ground-level ecological observations is needed, ensuring that large-scale monitoring aligns with field realities for effective, data-driven forest management [74, 75].

Machine learning for forest health: performance, challenges, and applications

Selecting an ML model is crucial for forest health classification. This study found RF had the highest accuracy, followed by SVM, while DT performed the lowest. RF achieved the best agreement and balanced accuracy by effectively obtaining complex ecological patterns. Its ability to handle large datasets and non-linear relationships makes it ideal for large-scale monitoring. However, its computational demands, lack of interpretability, and reliance on extensive labelled data may still limit its applicability in conservation policy and data-scarce environments [22].

SVM was the second-best model, but exhibited minor misclassifications between Healthy and Unhealthy forests, highlighting sensitivity to overlapping conditions. SVMs effectively handle high-dimensional, non-linear data through kernel transformations but are highly sensitive to kernel choice and hyperparameter tuning (e.g., C and gamma). Inadequate tuning can lead to overfitting or underfitting, particularly in small, noisy, or imbalanced datasets. SVMs also struggle with class overlap and offer limited interpretability, which constrains ecological inference. These limitations necessitate rigorous cross-validation, robust parameter optimization, and, where applicable, integration with more interpretable models [70, 76]. DT had the lowest accuracy and showed limitations in classifying Unhealthy forests. Cross-validation revealed high DT variability, while its simplicity aids rapid assessments and policy communication. However, overfitting and poor handling of noisy data reduce its predictive reliability [77].

Most classification metrics range from 0 to 1, with exceptions. Accuracy, sensitivity, specificity, and balanced accuracy follow this range, where 0 indicates total misclassification and 1 represents accurate classification. However, the Kappa statistic ranges from − 1 to 1, where − 1 indicates worse-than-random agreement, 0 represents chance agreement, and 1 indicates complete agreement. These metrics are widely used to assess ML classification models [78, 79]. The effectiveness of these models depends on data quality, including accurate labelling, sufficient sample size, and balanced class representation. Imbalanced datasets, where degraded forests are underrepresented, can bias models toward healthier classifications [22, 76, 77]. Moreover, ML models primarily rely on historical data, limiting their ability to anticipate novel ecological factors such as climate change-induced shifts in vegetation dynamics [80].

RF-based feature importance showed that forest health is strongly linked to stand structure and regeneration. Forest structural attributes like DBH and height were top predictors, reflecting maturity, biomass, and canopy complexity [81]. These attributes better capture forest condition than topographic variables. A LiDAR-based assessment confirms tree size metrics as most useful for biomass prediction, while reviews highlight DBH, height, and canopy traits as important forest health indicators [82]. Regeneration, reflected by seedling abundance, indicates recovery potential. In Himalayan forests, poor regeneration is a known issue [83]. In the study area, seedling abundance also emerged as a primary indicator of forest health. While topographic variables (e.g., elevation, slope) define habitat conditions, they are less effective than structural and regenerative metrics in reflecting current forest health.

Comparative analysis underscores that RF is the most effective method for extensive forest monitoring, while SVM prove beneficial for differentiating forest conditions within moderate-sized datasets [84,85,86,87,88]. DT, by presenting lower accuracy, also hold potential as a rapid assessment tool. The integration of these models could enhance classification accuracy by utilizing DT for preliminary classification, RF for refining predictions, and SVM for targeted ecological assessments [89,90,91]. Nonetheless, the limitations inherent in ML, including data dependence (concerning quality, quantity, and accuracy), computational costs, and challenges related to interpretability, must be confronted to facilitate their practical application in forest conservation [22, 76, 77, 84]. Addressing these challenges necessitates expert involvement, the development of hybrid modelling approaches, the integration of artificial intelligence (AI) techniques, along with the implementation of accurate and appropriate data collection methods, thus improving transparency and usability within conservation projects [75, 76, 80].

This study may not comprehensively summarize the overall health of forests in the region due to constraints in spatial and temporal coverage; however, it effectively illustrates the applicability of ML models utilizing basic structural data from local forest types. Frameworks for real-time monitoring can support adaptive conservation and the sustainable management of forests [18, 25, 70, 84]. Ultimately, data-driven approaches can inform adaptive management strategies, optimize resource allocation, and strengthen conservation planning [92, 93]. Future research should incorporate suitable AI tools that integrate deep learning, remote sensing, and geospatial analytics to enhance the precision of models. This approach can strengthen interpretability, thus facilitating policymaking.

Conclusion

This study evaluated forest ecosystem health in the Kashmir Himalayas by combining detailed field-based structural data with both unsupervised (PCA, clustering) and supervised (RF, DT, SVM) machine learning approaches. Unlike earlier work focused mainly on remote sensing or limited variables, it used a broader set of nine ecological and disturbance indicators (tree density, species richness, DBH, height, regeneration, deforestation, slope, grazing, and erosion) to capture key patterns shaping forest condition. The analysis revealed significant structural variation across sites, particularly in tree density, species richness, and regeneration. Three ML models (RF, DT, and SVM) were then used to classify forest health, with RF demonstrating the highest accuracy, followed by SVM and DT. Tree DBH, height, and regeneration emerged as key predictors of forest health, showing strong associations with disturbance and elevation. Healthy forests showed higher growth rates, regeneration and low disturbance, while Unhealthy ones exhibited increased rates of deforestation, low seedling recruitment, and greater environmental stress. These classifications provided a valuable framework for forest monitoring and management. The findings underscore the importance of integrating ML with ecological assessments to identify vulnerable forest ecosystems and facilitate targeted conservation. As highlighted by the RF variable importance analysis, focusing on the protection of larger trees and the restoration of degraded sites, characterized by low regeneration, reduced DBH, and high disturbance, can support biodiversity and enhance ecosystem resilience. Similarly, disturbance drivers such as grazing, deforestation, and erosion require targeted management. Integrated and ensembled predictive ML models into forest planning can support adaptive management and spatial prioritization. Future research should incorporate additional environmental variables and remote sensing data to refine models and enable broader-scale forest health monitoring across the Himalayan region.

Data availability

Data is contained within the article.

References

Biswas T, Subramanian S, Niyogi R, Mylliemngap W, Mitra S, Das DK, et al. Conservation challenges and adaptation strategies for Indian Himalayan biodiversity in a changing climate. In: Editor A, Editor B, editors. Climate change impact on Himalayan biodiversity. Springer Nature Switzerland; 2025. pp. 179–224.

Irfan U, Shaheen H, Manzoor M, Nasar S, Khan RW, Gillani SW, et al. Population structure, floral diversity, habitat geography, and conservation status of Himalayan horse chestnut (Aesculus indica) in Western Himalayan moist temperate forest ecosystems of Kashmir region. Front Glob Change. 2025;7:1388607. https://doi.org/10.3389/ffgc.2024.1388607.

Haile AA, Seid A, Mekonnen AB, Wassie WA, Yemata G, Yihune E, et al. Human disturbances and their impact on Woody species diversity in sacred church forests in West Gojjam zone, Northwestern Ethiopia. Trees People. 2025;19:100776. https://doi.org/10.1016/j.tfp.2025.100776.

Rawat N, Pandey N, Joshi A, Pandey D. Influence of climate change on Himalayan ecosystem: a review. Clim Change Impact Himal Biodivers. 2025;18:415–24.

Ahmad A, Ahmad SR, Gilani H, Nowosad J. Assessment of forest fragmentation and ecological dynamics in Western Himalayan region over three decades (1990–2020). Environ Monit Assess. 2025;197(2):1–19.

Raj DA, Kumar S, Mariappan S, Sooryamol K, Kalambukattu JG. Potential of pine forest in controlling soil erosion in Himalayan region - Investigation using fallout radionuclide (137Cs) measurements. Evol Earth. 2023;2:100031. https://doi.org/10.1016/j.eve.2024.100031.

Mir S. Energy needs drive deforestation in Pakistan-administered Kashmir. Eco-Business. 2023 Aug 30. https://www.eco-business.com/news/energy-needs-drive-deforestation-in-pakistan-administered-kashmir/. Accessed 03 Apr 2025.

Shafiq M, Mahmood S. Spatial assessment of forest cover change in Azad kashmir, Pakistan. Adv GIS. 2022;2(2):62–9.

Liang W, Feng Y, Li F, Zhao Y, Zhu J, Yao J, Chang C, Wei X. Drivers of vegetation and soil determine natural regeneration of a single plantation at different slope positions. Front Glob Change. 2023;6:1167284. https://doi.org/10.3389/ffgc.2023.1167284.

Negi VS, Pandey A, Singh A, Bahukhandi A, Pharswan DS, Gaira KS, et al. Elevation gradients alter vegetation attributes in mountain ecosystems of Eastern himalaya, India. Front Glob Change. 2024;7:1381488. https://doi.org/10.3389/ffgc.2024.1381488.

Joshi VC, Sundriyal RC, Arya D. Impacts of climate warming on biodiversity, ecosystem functionality, and services in the Himalayan landscape. Climate change impact on Himalayan biodiversity. Cham: Springer Nature Switzerland; 2025. pp. 337–51.

Goheer MA, Fatima L, Farah H, Hassan SS, Abbas N. Assessment of change in forest land, carbon stock and carbon emissions of KPK, Pakistan for past three decades using Geospatial techniques. J Water Clim Change. 2023;14(2):442–53.

Pandey KP, Wellstein C, Bräuning A, Bhuju DR. Climatic influence on growth performance of Abies spectabilis in the Himalayas. Forests. 2025;16(3):473.

Kumar A, Kumar S. Impact of climate change on Himalayan ecosystem and traditional crops. Clim Change Impact Himal Biodivers. 2025;18:393–413.

Engert JE, Souza CM Jr, Kleinschroth F, Bignoli DJ, Costa SCP, Botelho J Jr, et al. Explosive growth of secondary roads is linked to widespread tropical deforestation. Curr Biol. 2025. https://doi.org/10.1016/j.cub.2025.02.017.

Lv Z, Li S, Xu X, Lei J, Peng Z. Assessment of vegetation restoration potential in central Asia. J Environ Manage. 2025;374:124089. https://doi.org/10.1016/j.jenvman.2025.124089.

Dhyani S. Impact of drivers of biodiversity loss on mountain ecosystems: assessing the need for ecosystem health assessments in Indian Himalayan region. Singh K, Ribeiro MC, Calicioglu O, Biodiversity and bioeconomy, Editors. Elsevier; 2024. pp. 25–46.

Qadeer A, Shakir M, Wang L, Talha SM. Evaluating machine learning approaches for aboveground biomass prediction in fragmented high-elevated forests using multi-sensor satellite data. Remote Sens Appl Soc Environ. 2024;36:101291. https://doi.org/10.1016/j.rsase.2024.101291.

Dharma VL, Nurtanio NK, Nugroho FS, Anggreainy MS, Kurniawan A. A review on machine learning methods for water quality prediction. In: Proc 4th Int Conf Artif Intell Data Sci (AiDAS); 2023 Aug 8–9; Ipoh, Malaysia. IEEE; 2023. pp. 131–8.

Olden JD, Lawler JJ, Poff NL. Machine learning methods without tears: a primer for ecologists. Q Rev Biol 2008;83(2):171–93.

Pyayt AL, Mokhov II, Lang B, Krzhizhanovskaya VV, Meijer RJ. Machine learning methods for environmental monitoring and flood protection. World Acad Sci Eng Technol. 2011;78:118–23.

Alotaibi E, Nassif N. Artificial intelligence in environmental monitoring: in-depth analysis. Discov Artif Intell. 2024;4:84. https://doi.org/10.1007/s44163-024-00198-1.

Reichstein M, Camps-Valls G, Stevens B, Jung M, Denzler J, Carvalhais N, Prabhat F. Deep learning and process Understanding for data-driven Earth system science. Nature. 2019;566(7743):195–204.

Sarker IH. Machine learning: algorithms, real-world applications and research directions. SN Comput Sci. 2021;2(3):160.

Fassnacht FE, White JC, Wulder MA, Næsset E. Remote sensing in forestry: current challenges, considerations and directions. Forestry. 2024;97(1):11–37.

GoAJK. Government of the State of Azad Jammu & Kashmir. Statistical Year Book 2019. AJ&K Bureau of Statistics, Planning & Development Department. 2019. https://pndajk.gov.pk/uploadfiles/downloads/AJ&K%20Statistical%20Year%20Book%202019.pdf. Accessed 03 Apr 2025.

Khan RWA, Quratulain M, Shaheen H, Hussain K, Mehmood AB. Local scale variations in soil organic carbon sequestration in lesser Himalayan coniferous and mixed forests: implications for sustainability. J Ecol Environ. 2025;49:02. https://doi.org/10.5141/jee.24.088.

Khan MZ, Baig T, Ali S, Shah A. Mountain studies: understanding and managing mountains for people and nature. In: Khan MZ, editor. Karakoram International University, Gilgit Pakistan. 2022. https://kiu.edu.pk/cms/wp-content/uploads/2023/02/Mountain-Studies-Book-Online-Version.pdf. Accessed 03 Apr 2025.

Mattia SB, Sesay S. Ground forest inventory and assessment of carbon stocks in Sierra leone, West Africa. Natural resources management and biological sciences. IntechOpen; 2020. https://doi.org/10.5772/intechopen.88950.

Sullivan MJP, Lewis SL, Hubau W, Qie L, Baker TR, Banin LF, et al. Field methods for sampling tree height for tropical forest biomass Estimation. Methods Ecol Evol. 2018;9(5):1179–89.

Wu H, Xu H, Tian X, Zhang W, Lu C. Multistage sampling and optimization for forest volume inventory based on Spatial autocorrelation analysis. Forests. 2023;14(2):250. https://doi.org/10.3390/f14020250.

Hetzer J, Huth A, Wiegand T, Dobner HJ, Fischer R. An analysis of forest biomass sampling strategies across scales. Biogeosciences. 2020;17(6):1673–83.

Majeed M, Khan AM, Habib T, Anwar MM, Sahito HA, Khan N, et al. Vegetation analysis and environmental indicators of an arid tropical forest ecosystem of Pakistan. Ecol Indic. 2022;142:109291. https://doi.org/10.1016/j.ecolind.2022.109291.

Tarin MWK, Nizami SM, Jundong R, Lingyan C, You H, Farooq TH, Gilani MM, Ifthikar J, Tayyab M, Zheng Y. Range vegetation analysis of Kherimurat scrub forest, Pakistan. Int J Dev Sustain. 2017;6(10):1319–33.

Gairola S, Rawal RS, Todaria NP. Forest vegetation patterns along an altitudinal gradient in sub-alpine zone of West himalaya, India. Afr J Plant Sci. 2008;2(6):42–8.

Shrestha BB, Moe SR. Plant diversity and composition associated with Himalayan musk deer latrine sites. Zool Ecol. 2015;25(4):295–304.

R Core Team. R: A language and environment for statistical computing (Version 4.4.3). R Foundation for Statistical Computing. 2025. https://www.R-project.org

Hammer Ø, Harper DAT, Ryan PD. PAST: paleontological statistics software package for education and data analysis. Palaeontol Electron. 2001;4(1):9.

Santoso T, Darmawan A, Sari N, Syadza MAF, Himawan ECB, Rahman WA. Clusterization of agroforestry farmers using K-means cluster algorithm and elbow method. J Sylva Lestari. 2023;11(1):107–22.

Mushagalusa CA, Fandohan AB, Glèlè Kakaï R. Random forest and Spatial cross-validation performance in predicting species abundance distributions. Environ Syst Res. 2024;13(1):23. https://doi.org/10.1186/s40068-024-00297-7.

Rocha SJSSD, Torres CMME, Villanova PH, Tavares Júnior IDS, Rufino MPMX, Romero FMB, et al. Machine learning methods: modeling net growth in the Atlantic forest of Brazil. Ecol Inf. 2024;81:102564. https://doi.org/10.1016/j.ecoinf.2024.102564.

Allgaier J, Pryss R. Cross-validation visualized: a narrative guide to advanced methods. Mach Learn Knowl Extr. 2024;6(2):1378–88.

Elsobky AM, Keshk AE, Malhat MG. A comparative study for different resampling techniques for imbalanced datasets. Int J Comput Inf. 2023;10(3):147–56.

Xiao J, Wang Y, Chen J, Xie L, Huang J. Impact of resampling methods and classification models on the imbalanced credit scoring problems. Inf Sci. 2021;569:508–26.

Gregorutti B, Michel B, Saint-Pierre P. Correlation and variable importance in random forests. Stat Comput. 2017;27:659–78.

Gyedu AK, Damptey FG, Barnes VR, Bentsi-Enchill F, Sarfo DA, Mohammed I, Opuni-Frimpong E. Multiple anthropogenic and Climatic factors drive tree species attributes in ecologically distinct sacred groves in Ghana. Int J Res. 2025;2025:9944297. https://doi.org/10.1155/ijfr/9944297.

Zhang H, Li Y, Xu K, Yu L, He P, Zeng S, et al. Effects of anthropogenic disturbance on the structure, competition, and succession of Abies ziyuanensis communities. Forests. 2024;15(6):1001. https://doi.org/10.3390/f15061001.

Khadka S, He HS, Bardhan S. Investigating the Spatial pattern of white oak (Quercus Alba L.) mortality using ripley’s K function across the ten States of the Eastern US. Forests. 2024;15(10):1809.

Kaplan A, Khan MN, Hayat K, Iqbal M, Ali B, Wahab S, et al. Coupling environmental factors and climate change: impacts on plants and vegetation growth patterns in ecologically sensitive regions. In: Fahad S, Saud S, Nawaz T, Gu L, Ahmad M, Zhou R, editors. Environment, climate, plant and vegetation growth. Cham: Springer; 2024. https://doi.org/10.1007/978-3-031-69417-2_11.

Kissa DO, Nzunda EF, Tweheyo M, Lussetti D, Ssekuubwa E, Sheil D. Composition and diversity of understory and canopy species vary along a logging gradient in an African semi-deciduous tropical rainforest. Forests. 2024;16(1):6. https://doi.org/10.3390/f16010006.

Iype AH, Suryawanshi K, Khanyari M. Restoration in the Western himalaya: a systematic review of current efforts and implications for the future. Restor Ecol. 2025;e14378. https://doi.org/10.1111/rec.14378.

Pourfarrashzadeh F, Madadi A, Gharachorlu M, Sareskanrood SA. Spatial-statistical modeling of deforestation from an ecogeomorphic approach in typical hyrcanian forests, Northern Iran. Environ Monit Assess. 2025;197(3):324.

Gervasi V, Aragno P, Salvatori V, Caniglia R, Angelis DD, Fabbri E, et al. Estimating distribution and abundance of wide-ranging species with integrated Spatial models: opportunities revealed by the first Wolf assessment in south-central Italy. Ecol Evol. 2024;14(5):e11285. https://doi.org/10.1002/ece3.11285.

Dhiman V, Kumar A. Species exhibiting positive association demonstrate high above-ground biomass accumulation in the subtropical Himalayan forest ecosystem, India. Appl Ecol Environ Res. 2025;23(1):1433–52.

López-García J, Cruz-Bello GM, de Lourdes Manzo-Delgado L. A long-term analysis, modeling and drivers of forest recovery in central Mexico. Environ Monit Assess. 2024;197(1):87.

Gorrod E, White L, Bedward M, Hope B, Chu N, McAllister D. Restoration thinning accelerates small tree growth but May slow large tree growth in a multi-age flood-dependent forest. J Appl Ecol. 2024;61(7):1560–72.

Iheaturu CJ, Hepner S, Batchelor JL, Agonvonon GA, Akinyemi FO, Wingate VR, Ifejika Speranza C. Integrating UAV lidar and multispectral data to assess forest status and map disturbance severity in a West African forest patch. Ecol Inf. 2024;84:102876. https://doi.org/10.1016/j.ecoinf.2024.102876.

Köpeczi-Bócz T, Farkas-Kis M, Bencekri M, Csiba-Herceg Á. Resilient cultural landscapes: adaptive management and social innovation in heritage conservation. Int J Urban Sci. 2025;29(1):247–72.

Hu X, Duan G, Jin Y, Cheng Y, Liang F, Lian Z, Li F, Wang Y, Chen H. Study on the natural regeneration characteristics and influencing factors of typical Quercus forests in Northern China. Forests. 2025;16(2):250. https://doi.org/10.3390/f16020250.

Zhou L, Thakur MP, Jia Z, Hong Y, Yang W, An S, Zhou X. Light effects on seedling growth in simulated forest canopy gaps vary across species from different successional stages. Front Glob Change. 2023;5:1088291. https://doi.org/10.3389/ffgc.2022.1088291.

Ganie PA, Posti R, Bharti VS, Sehgal VK, Sarma D, Pandey PK. Erosion landscape characterization in the Himalayan basin: insights from Geospatial data and multi-criteria evaluation. Environ Monit Assess. 2025;197(1):1–46.

Robles D, Boulanger Y, Pascual J, Danneyrolles V, Bergeron Y, Drobyshev I. Timber harvesting was the most important factor driving changes in vegetation composition, as compared to climate and fire regime shifts, in the mixedwood temperate forests of Temiscamingue since AD 1830. Landsc Ecol. 2025;40(2):26. https://doi.org/10.1007/s10980-025-02043-x.

Saha S, Rajwar GS, Kumar M. Forest structure, diversity and regeneration potential along altitudinal gradient in Dhanaulti of Garhwal himalaya. Syst. 2016;25(2):e058. https://doi.org/10.5424/fs/2016252-07432.

Kacker S, Krishna S, Das A, Shahabuddin G. Patterns of tree regeneration and their implications for succession in Himalayan pine-oak forests, India. Ecol Manage. 2024;562:121941. https://doi.org/10.1016/j.foreco.2024.121941.

Khaine I, Woo SY, Kwak M, Lee SH, Je SM, You H, et al. Factors affecting natural regeneration of tropical forests across a precipitation gradient in Myanmar. Forests. 2018;9(3):143. https://doi.org/10.3390/f9030143.

Schwab N, Bürzle B, Bobrowski M, Böhner J, Chaudhary RP, Scholten T, et al. Predictors of the success of natural regeneration in a Himalayan treeline ecotone. Forests. 2022;13(3):454. https://doi.org/10.3390/f13030454.

Masanja SN, Shirima DD, Zahabu EM. Relationship between Deadwood structural diversity and carbon stock along environmental and disturbance gradients in tropical dry forests. Folia Oecol. 2025;52(1):1–13.

Owens G, Heinsohn R, Gibbons P, Stojanovic D. Quantifying forest degradation, deforestation and land use change in vital swift Parrot breeding habitat. Sci Rep. 2025;15(1):1–16.

Yuan Y, Li H, Wang H, Liu T, Xue Z, Song J, Xu L. Riparian forest health assessment in the Valley area of the Irtysh river basin. Forests. 2025;16(2):373. https://doi.org/10.3390/f16020373.

Daviran M, Ghezelbash R, Hajihosseinlou M, Maghsoudi A. Uncertainty quantification in genetic algorithm-optimized artificial intelligence-based mineral prospectivity models: automated hyperparameter tuning for support vector machines and random forest. Model Earth Syst Environ. 2025;11. https://doi.org/10.1007/s40808-024-02176-z. Article 10.

Habeeb HN, Mustafa YT. Analyzing the impact of geoenvironmental factors on the Spatiotemporal dynamics of forest cover via random forest. Earth. 2025;6(1):3. https://doi.org/10.3390/earth6010003.

Marino S, Brugiapaglia E, Miraglia N, Persichilli C, De Angelis M, Pilla F, Di Brita A. Modelling of the above-ground biomass and ecological composition of semi-natural grasslands on the strength of remote sensing data and machine learning algorithms. Ecol Inf. 2024;82:102740. https://doi.org/10.1016/j.ecoinf.2024.102740.

Hosingholizade A, Erfanifard Y, Alavipanah SK, Millan VE, Mielcarek M, Pirasteh S, Stereńczak K. Assessment of pine tree crown delineation algorithms on UAV data: from K-means clustering to CNN segmentation. Forests. 2025;16(2):228. https://doi.org/10.3390/f16020228.

Hasan R, Farabi SF, Kamruzzaman M, Bhuyan MK, Nilima SI, Shahana A. AI-driven strategies for reducing deforestation. Am J Eng Technol. 2024;6(06):6–20.

Sharma MK, Prasad SVGVA, Anusha G, Das A, Sambathkumar M. Tech-driven solutions for environmental conservation by AI collaboration processes. In: Arun J, editor. Environmental applications of carbon-based materials. IGI Global; 2024. pp. 1–29.

Yadav BR. Machine learning algorithms: optimizing efficiency in AI applications. Int J Eng Manage Res. 2024;14(5):49–57.

Bongomin O, Nzila C, Mwasiagi JI, Maube O. Exploring insights in biomass and waste gasification via ensemble machine learning models and interpretability techniques. Int J Energy Res. 2024;1:6087208. https://doi.org/10.1155/2024/6087208.

Sathyanarayanan S, Tantri BR. Confusion matrix-based performance evaluation metrics. Afr J Biomed Res. 2024;27(4s):4023–31.

Shabani F, Kumar L, Ahmadi M. Assessing accuracy methods of species distribution models: AUC, specificity, sensitivity, and the true skill statistic. Global J Human-Soc Sci: B Geo-Sciences Environ Sci Disaster Manage. 2018;18(1):1–10.

Hamdan A, Ibekwe KI, Etukudoh EA, Umoh AA, Ilojianya VI. AI and machine learning in climate change research: A review of predictive models and environmental impact. World J Adv Res Rev. 2024;21(1):1999–2008.

Lutz JA, Larson AJ, Freund JA, Swanson ME, Bible KJ. The importance of large-diameter trees to forest structural heterogeneity. PLoS ONE. 2013;8(12):e82784. https://doi.org/10.1371/journal.pone.0082784.

Estrada JS, Fuentes A, Reszka P, Auat Cheein F. Machine learning assisted remote forestry health assessment: a comprehensive state of the Art review. Front Plant Sci. 2023;14:1139232. https://doi.org/10.3389/fpls.2023.1139232.

Dhamala MK, Aryal PC, Suwal MK, Bhatta S, Bhuju DR. Population structure and regeneration of himalayan endemic Larix species in three high-altitude valleys in Nepal himalaya. J Ecol Environ. 2020;44:22. https://doi.org/10.1186/s41610-020-00166-7.

Liu Z, Peng C, Work T, Candau JN, DesRochers A, Kneeshaw D. Application of machine-learning methods in forest ecology: recent progress and future challenges. Environ Rev. 2018;26(4):339–50.

Xu X, Fu X, Zhao H, Liu M, Xu A, Ma Y. Three-dimensional reconstruction and geometric morphology analysis of lunar small craters within the patrol range of the Yutu-2 Rover. Remote Sens. 2023;15(17):4251. https://doi.org/10.3390/rs15174251.

Li F, Shan Y, Li M, Guo Y, Liu C. Insights for river restoration: The impacts of vegetation canopy length and canopy discontinuity on riverbed evolution. Water Resour Res. 2024;60(7):p.e2023WR036473. https://doi.org/10.1029/2023WR036473.

Feng X, Wang Z, Wu X, Huang S, Li J, Lai C, Zeng Z, Lin G. Tracking 3D drought events across global river basins: Climatology, spatial footprint, and temporal changes. Geophys Res Lett. 2025;52(3):p.e2024GL111442. https://doi.org/10.1029/2024GL111442.

He Q, Zhan J, Liu X, Dong C, Tian D, Fu Q. Multispectral polarimetric bidirectional reflectance research of plant canopy. Opt Lasers Eng. 2025;184:p.108688. https://doi.org/10.1016/j.optlaseng.2024.108688.

Tithi AC, Sohel MSI, Jakariya M, Rana P. Analyzing ecosystem functions in Bangladesh’s forests: a historical MODIS study. Geol, Ecol, Landsc. 2024;1–13. https://doi.org/10.1080/24749508.2024.2334516.

Tamiru B, Soromessa T, Warkineh B, Legese G. Modelling selected soil chemical properties using TerrSet and Random Forest: a case of Hangadi watershed, Oromia, Ethiopia. Geol, Ecol, Landsc. 2024;1-17. https://doi.org/10.1080/24749508.2024.2429842.

Yang G, Qiu H, Wang N, Yang D, Liu Y. Tracking 35-year dynamics of retrogressive thaw slumps across permafrost regions of the Tibetan Plateau. Remote Sens Environ. 2025;325:114786. https://doi.org/10.1016/j.rse.2025.114786.

Zhang J, Yan F, Lyne V, Wang X, Su F, Cao Q, He B. Monitoring of ecological security patterns based on long-term land use changes in Langsa Bay, Indonesia. Int J Digit Earth. 2025;18(1):2495740. https://doi.org/10.1080/17538947.2025.2495740.

Zheng Y, Li J, Zhu T, Li J. Experimental and MPM modelling of widened levee failure under the combined effect of heavy rainfall and high riverine water levels. Comput Geotech. 2025;184:107259. https://doi.org/10.1016/j.compgeo.2025.107259.