1. Introduction
Synthetic aperture radar (SAR) interferometry is a geodetic imaging technique that has rapidly evolved over the past 30 years. It offers significant advantages, including all-weather, all-day operational capabilities, and wide-area measurement coverage. These features enable SAR to capture high-precision, high-resolution data on ground elevation and surface deformation. As a result, SAR interferometry has become a crucial tool for monitoring ground deformation and identifying geological hazards [1,2,3,4,5,6,7,8]. With the continuous advancement of satellite technology, SAR data resolution has steadily improved, while revisit frequency has increased and coverage expanded. This progress has greatly enhanced the availability of SAR data, resulting in the accumulation of massive datasets. For example, the European Space Agency’s (ESA) Sentinel-1 mission now collects more than 10 terabytes of SAR data daily [9].
Under ideal conditions, satellite-acquired radar images can reveal greater detail about surface deformation than even the most intensive ground-based geodetic measurements [10]. The interferometric synthetic aperture radar (InSAR) technique has been widely applied in various fields, including landslides [11,12,13,14,15], ground subsidence [16,17,18,19,20], permafrost changes [21], earthquake-related deformation [22,23,24,25,26], and volcanic deformation [27,28,29]. However, the automatic processing of extensive InSAR datasets remains a substantial challenge, particularly in accurately identifying deformation signals from these datasets [9,30]. Automatically distinguishing real deformation features from large InSAR datasets has emerged as a critical challenge in the study of volcanoes and earthquakes.
Traditional methods for detecting volcanic and earthquake-related deformation from InSAR interferograms often rely on manual feature extraction and labeling processes. Handling and analyzing such large volumes of data is both time-consuming and labor-intensive. For instance, following an earthquake, traditional InSAR processing methods often take days or even weeks to analyze satellite images, leading to delayed post-disaster responses. With the rapid growth in data scale, the limitations of these methods have become increasingly apparent. Consequently, there is an urgent need to develop an efficient and accurate automated detection method. Such a method would not only significantly improve data processing speed and analytical precision but also effectively reduce labor costs, thereby providing stronger technical support for geophysical research.
Deep learning (DL) techniques, rooted in neural networks, have emerged as effective methodologies for analyzing large InSAR datasets, driving significant advancements in InSAR processing. These techniques have achieved notable developments, particularly in the classification of deformation features and noise reduction. DL has been instrumental in the identification and localization of deformation phenomena associated with volcanic activity, landslides, and earthquakes [2,9,31,32,33,34,35,36,37,38], as well as in time-series-based deformation prediction [39,40] and subsidence monitoring [41]. Moreover, DL methods have been effectively applied across various stages of InSAR data processing, including phase unwrapping, phase filtering, and atmospheric correction, significantly reducing noise and artifacts in interferograms (Figure 1). These contributions highlight the transformative potential of DL in enhancing the accuracy and efficiency of InSAR analysis.
This paper presents a comprehensive review of significant studies on detecting volcanic and earthquake-related deformation of InSAR interferograms using DL. It begins by introducing the fundamental DL networks used for InSAR data analysis, followed by a summary of advancements in model architectures and methods for dataset construction for detecting volcanic and earthquake-related deformation. Finally, the paper explores the current challenges and emerging trends for automating InSAR deformation signal identification through DL.
2. Fundamental Deep Learning Networks for InSAR Data Processing
DL facilitates pattern recognition and decision making in complex data environments [42,43,44] by constructing and training deep neural network models, such as convolutional neural networks (CNNs), recurrent neural networks (RNNs), generative adversarial networks (GANs), and the recently emerging Transformer networks. In practice, DL models are trained using labeled data, where each data point is associated with a corresponding label. Through repeated training, these models progressively extract and refine meaningful patterns and structures from the data. With each iteration, the model enhances its ability to identify and prioritize key features, improving its accuracy in classifying and predicting previously unseen data. These advancements have resulted in significant breakthroughs in image classification [45,46,47,48,49,50,51,52,53,54].
2.1. Fundamental Architectures in Deep Learning for InSAR Data Processing
DL architectures have evolved to address diverse challenges in data processing, each offering unique strengths tailored to specific tasks.
-
CNNs are foundational DL models originally designed for computer vision tasks like image classification, object detection, and segmentation. They consist of key layers, including convolutional layers, activation functions, pooling layers, and fully connected layers [55,56,57]. Convolutional layers extract features such as edges and textures using learnable kernels [58,59], while activation functions introduce nonlinearity to enhance pattern recognition. Pooling layers reduce spatial dimensions, minimizing computational cost [60,61], and fully connected layers integrate features for final predictions. The structure is shown in Figure 2a.
-
RNNs are critical for processing sequential data, such as InSAR time series, by capturing temporal dependencies through cyclic connections, as shown in Figure 2b. Originating with the Elman network [62], early RNNs faced gradient vanishing/explosion issues, which were mitigated by the introduction of Long Short-Term Memory (LSTM) networks [63] and gated recurrent units (GRUs) [64]. These advancements enabled effective long-term dependency modeling, making RNNs indispensable for natural language processing, speech recognition, and time-series prediction.
-
Introduced by Goodfellow et al. [65], GANs employ an adversarial framework between a generator and a discriminator to produce realistic synthetic data. Enhancements such as Deep Convolutional GANs (DCGANs) [66], Conditional GANs (CGANs) [67], and CycleGANs [68] have expanded their applications, from improving image quality to unpaired style transfer. Advanced models like Wasserstein GANs (WGANs) [69], BigGANs [70], and StyleGANs [71] further refined GANs, achieving stability, diversity, and high-quality outputs. The main structure of GANs is illustrated in Figure 2c.
-
Transformers, introduced by Vaswani et al. [72], revolutionized DL with an encoder–decoder architecture leveraging self-attention mechanisms. This design overcomes the limitations of RNNs and CNNs, excelling in sequence-to-sequence tasks [72,73,74]. Incorporating multi-head attention, feed-forward layers, and residual connections, Transformers have achieved remarkable success in natural language processing and computer vision, as shown in Figure 2d.
2.2. Application of Deep Learning Networks in InSAR Data Processing
In recent years, the application of deep learning networks in InSAR data processing has become increasingly widespread. As illustrated in Table 1, CNNs have significantly advanced InSAR data processing, enhancing tasks like denoising, phase unwrapping, and atmospheric correction. Early models, such as Mukherjee et al.’s denoising approach, eliminated the need for clean ground truth images, while innovations like Yang et al.’s multi-level fusion and Liu et al.’s ResNext integration improved noise suppression and segmentation [31,75,76,77,78,79]. RNNs, particularly LSTM and GRU, have excelled in sequential data analysis, with applications such as deformation monitoring and noise reduction, though challenges like high computational cost remain [39,80,81]. GAN-based models have addressed data challenges in InSAR, with applications ranging from atmospheric correction (CycleGAN, Pix2Pix) to phase unwrapping (PU-GAN) [82,83,84,85]. Transformer networks, with their ability to enhance phase feature extraction and handle missing data, have recently been applied to InSAR data, offering improved deformation signal reconstruction and predictions in complex environments like permafrost regions [86,87,88]. These advancements collectively underscore the transformative impact of deep learning techniques in improving the efficiency and accuracy of InSAR data analysis.
3. Deep Learning Architecture Models for Detecting Volcanic and Earthquake-Related Deformations from InSAR Imagery
DL techniques have shown significant potential in processing InSAR data for detecting surface deformations. For instance, CNNs excel at spatial feature extraction, effectively identifying and removing atmospheric noise, thereby enhancing the purity of deformation signals. On the other hand, RNNs are well suited for analyzing time-series data, offering unique advantages in capturing the dynamic surface deformation processes before and after earthquake events. Additionally, GANs have proven useful in data enhancement and the filling of missing data, presenting novel solutions to improve the completeness of InSAR datasets and address the challenge of insufficient training samples. This chapter explores the architecture of DL models in depth. Given that some research has leveraged transfer learning strategies to optimize model performance, we begin with an overview of transfer learning techniques. Following this, we systematically review the current state-of-the-art DL models for detecting volcanic and earthquake-related deformations from InSAR interferograms.
3.1. Transfer Learning Strategy
DL ideally requires abundant labeled training data; however, acquiring such data is often expensive, time-consuming, or impractical. Transfer learning provides a powerful solution by enabling the reuse of knowledge gained from one task to address a related, yet distinct, task [89]. This approach leverages knowledge from a source task to improve performance on a target task, reducing the dependence on large datasets. The main challenge in transfer learning lies in identifying how to generalize knowledge from the source to the target task effectively [90,91].
Figure 3 illustrates the distinction between the learning processes of traditional machine learning and transfer learning techniques. Traditional machine learning approaches attempt to learn each task independently from scratch, requiring substantial training data for optimal performance. In contrast, transfer learning leverages knowledge from previously learned tasks and applies it to a target task, especially when the latter has limited high-quality training data. This approach significantly enhances learning efficiency and performance in data-scarce scenarios.
By utilizing pre-trained models and existing datasets, transfer learning accelerates training and enhances model performance. Successful transfer learning depends on uncovering shared features between the source and target tasks, employing techniques like feature sharing and model fine-tuning. However, its effectiveness is contingent on sufficient task similarity. When tasks differ significantly, the transferred knowledge can fail to improve performance or even degrade it. For tasks such as image classification, object recognition, and target detection, transfer learning typically begins with a pre-trained model, fine-tuned for the specific task. Compared to training from scratch, this method offers several advantages: (1) pre-trained models usually outperform randomly initialized ones; (2) training converges faster; and (3) the final model exhibits better generalization and accuracy [92].
Fine-tuning bridges the theoretical and practical aspects of transfer learning [93]. It involves adapting pre-trained models to new tasks by freezing certain layers and retraining specific ones. This technique allows models to quickly meet new task requirements while retaining their ability to extract meaningful features. Fine-tuning not only reduces the labeled data required for new tasks but also significantly shortens training time, establishing itself as an essential practice in transfer learning.
3.2. Deep Learning Models for Volcanic Deformation Detection
This section aims to provide a discussion of current DL models applied in the field of volcano deformation detection, especially those advanced architectures that can effectively process InSAR image data.
Anantrasirichai et al. (2018) were among the first to demonstrate the potential of machine learning algorithms for detecting volcanic ground deformation from large-scale InSAR data, paving the way for automated deformation monitoring through DL models [2]. In subsequent studies (Anantrasirichai et al., 2019a, 2019b) [30,37], the authors extended this approach by employing a CNN architecture, specifically leveraging a pre-trained AlexNet model. The model was initially trained with real or synthetic data, after which it underwent targeted retraining based on expert feedback on positive sample classifications, improving the overall accuracy of deformation detection. Building on these early advancements, Valade et al. [94] proposed a CNN-based volcano detection model designed to extract deformation signals from InSAR imagery. This model integrated a fully convolutional autoencoder and a ResNet-based heuristic module [47]. Unlike Anantrasirichai et al.’s approach, which fine-tuned an existing AlexNet architecture, Valade’s model was built from scratch, offering more flexibility in output generation and demonstrating superior performance in volcano deformation detection. Further contributions to the field were made by Fadhillah et al. [95], who applied DL techniques to detect prehistoric volcanic deformation. Similar to Anantrasirichai et al., they employed transfer learning in their model architecture, utilizing pre-trained networks such as ResNet-18, ResNet-50, and AlexNet. A comparative analysis of these networks revealed that each architecture achieved over 85% accuracy, highlighting the effectiveness of transfer learning in enhancing model performance.
In addition to the network models based on supervised learning, semi-supervised and self-supervised learning models are also used to extract volcano deformation. For instance, Bountos et al. [96] proposed a self-supervised comparative learning model for binary classification of volcanic turbulence detection, based on the SimCLR framework [97]. The model architecture comprises two primary components: a self-supervised encoder and a fully connected layer specifically designed for supervised classification tasks. The self-supervised training process, based on SimCLR, optimizes the similarity between augmented views of the same example in the latent space, allowing the model to learn useful representations for detection. Bountos et al. also revealed that transfer learning from models trained in optical data, such as ImageNet, provides less meaningful features for the volcanic detection tasks [96]. To address these limitations, Beker et al. [98] proposed a novel DL model specifically designed to detect subtle long-term volcanic deformations. Unlike previous studies, this approach leverages a more advanced CNN architecture—Inception ResNet v2. Beker et al. also introduced an innovative training strategy: rather than relying on fine-tuning or transfer learning from pre-trained models on optical data, the model was trained exclusively using InSAR data from scratch [98]. This model demonstrated comparable performance to the leading short-term deformation detection models of its time, in terms of both AUC [99] and ROC [100] metrics. Notably, it was able to detect deformations as small as 5 mm/year—more than an order of magnitude smaller than previously detectable deformations.
Beyond DL-based detection of volcanic deformation, CNNs have also been adapted to simultaneously perform classification and localization tasks [45,101,102]. Gaddes et al. [34] proposed a CNN-based “two-head model” that not only classifies different types of deformations but also localizes them within interferograms. The first part of the model uses the five convolutional blocks of the VGG16 architecture [45] to extract features, which are then passed through a large fully connected layer. This layer connects to two distinct “heads”, the upper head for classification and the lower head for localization, enabling the model to simultaneously perform both tasks.
Although CNNs have made significant advancements in the detection and interpretation of deformation from InSAR data, their application remains limited by the scarcity of training data and the model’s insufficient consideration of global contextual information. To address these challenges, Abdallah et al. [103] proposed a volcano deformation detection method based on the Vision Transformer (ViT) model, which leverages the self-attention mechanism to effectively capture global dependencies. The model architecture consists primarily of a patch embedding layer, a Transformer encoder, and a pooling layer. In contrast to CNNs, which primarily focus on local features, this method demonstrates substantial improvements in both deformation classification accuracy and localization accuracy, thereby validating its superiority and effectiveness.
3.3. Deep Learning Models for Earthquake Deformation Detection
Compared to the extensive research on volcanic deformation detection using InSAR, the application of DL to earthquake-related InSAR deformation detection remains relatively limited. This section reviews several DL models applied to earthquake deformation analysis, highlighting their unique advantages and contributions to the field.
Similar to volcanic deformation detection, many earthquake deformation models are based on CNNs. Brengman et al. [104] developed SarNet, the first CNN architecture specifically designed for detecting, localizing, and classifying coseismic deformations in InSAR interferograms. SarNet builds upon the residual network (ResNet) introduced by He et al. [47], which is known for mitigating high training errors in deep networks [105,106,107]. The model is initially trained on synthetic datasets simulating both wrapped and unwrapped earthquake deformations, alongside synthetic noise representing atmospheric and terrain effects. By using randomly initialized weights, SarNet learns to recognize deformation patterns in synthetic data. Subsequently, the model employs transfer learning to adapt to real interferogram data. During this process, weights trained on synthetic data are transferred to the real-data model, with layers progressively fine-tuned while retaining the robustness of its early convolutional layers. SarNet takes single-channel interferometric images as input and outputs the likelihood of deformation presence. Additionally, Class Activation Maps (CAMs) [108,109] are used to locate deformation areas in the interferograms. Experimental results show that SarNet achieves an accuracy of 99.74% on validation datasets and 85.22% on real InSAR data.
da Silva and Motlagh [10,110] further explored the performance of different CNN architectures, such as InceptionV3, VGG19, and ResNet50V2, for earthquake deformation detection. Their findings suggest that transfer learning can achieve high classification performance even with limited labeled data. Motlagh also observed that deeper networks tend to perform better in earthquake-related tasks, although the conclusions are constrained by the specific datasets and tasks analyzed.
DL models have also been extended to InSAR time-series analysis. Rouet-Leduc et al. [111] pioneered the application of DL for extracting fault-related deformation from noisy InSAR time-series data. They proposed a convolutional autoencoder architecture consisting of 11 convolutional layers, designed to differentiate between deformation signals and atmospheric noise. This model uses InSAR time series and ground elevation maps as inputs, producing cumulative deformation maps over the observation interval by mitigating atmospheric distortions. Notably, this approach does not require prior knowledge of fault locations or slip behavior. Li et al. [38] adopted Rouet-Leduc’s model to reconstruct post-earthquake deformation using InSAR time-series data, while Zhu et al. [35] developed a similar DL framework to extract coseismic displacements, particularly for small- to medium-sized earthquakes (Mw < 6.5). Apart from extracting ground deformation from InSAR images, some researchers have also retrieved ground displacement fields caused by earthquakes from remote sensing optical images. Montagnon et al. proposed a deep convolutional neural network method, GeoFlowNet, based on the U-net architecture, which retrieves full-scale earthquake-induced ground displacement fields from optical satellite images with sub-pixel accuracy [112].
3.4. Summary of Deep Learning Models for Detecting Volcanic and Earthquake Deformation
In this section, we summarize the advantages and shortcomings of DL-based deformation analysis for earthquake and volcano studies. Basically, these DL applications can be categorized into four core approaches, based on their model architectures and training strategies.
-
Transfer learning using pre-trained networks offers the advantages of reduced training time and diminished reliance on large labeled datasets. However, a significant limitation lies in the rigidity of the pre-defined model architecture, which may fail to adequately capture the unique characteristics of InSAR data. This misalignment can compromise the model’s adaptability to InSAR-specific features, resulting in suboptimal performance and restricted flexibility. Furthermore, the inherent differences in data acquisition methods and the nature of information captured by optical and InSAR images introduce distributional discrepancies. These disparities can pose substantial challenges when transferring models pre-trained on optical datasets to InSAR applications. In some cases, this mismatch may lead to negative transfer, where the performance of the pre-trained model is inferior to that of a model trained from scratch using InSAR data.
-
Retraining existing network models with InSAR datasets aims to optimize the model for specific InSAR tasks, improving generalization and accuracy. However, a major limitation of this approach is its heavy reliance on large-scale InSAR training samples.
-
The newly designed model architecture integrates multiple complex modules, greatly enhancing flexibility. However, this increased complexity also brings challenges, including higher computational demands, longer training times, and a greater risk of overfitting. Despite these drawbacks, the model offers significant advantages, as it is specifically optimized for detecting volcano and earthquake deformation.
-
Self-supervised learning models based on the self-attention mechanism can enhance model performance by focusing on key information, efficiently processing long sequential data, capturing global dependencies, and leveraging unlabeled data or self-generated signals when labeled data are scarce. However, these models are more complex and require significant training time and resources. For example, a basic Vision Transformer model [113] requires 18 billion floating-point operations (FLOPs) to process images, where as the lightweight CNN model GhostNet [114,115] achieves similar performance with only about 600 million FLOPs. FLOP refers to the total number of floating-point operations, including multiplication, division, addition, and subtraction. Therefore, there is an urgent need for more efficient Transformer models to enable the deployment of Vision Transformers on resource-constrained devices [116].
4. Deep Learning Dataset Creation Method for Volcanic and Seismic Deformation Detection
Datasets, providing essential samples and accurate labels for model training, are fundamental to DL tasks. A well-constructed dataset, rich in samples and accurately labeled, provides the necessary “learning resources” for the model, enabling it to better capture data patterns and characteristics. This, in turn, enhances the model’s accuracy and robustness in real-world applications. Moreover, the quality of the datasets directly affects the model’s generalization ability. High-quality datasets that are diverse, representative, and comprehensive allow models to not only perform well on existing data but also generalize effectively to new data.
The volume of InSAR data generated daily is enormous; however, InSAR images capturing volcanic and earthquake deformation remain relatively scarce compared to those without such information, leading to a significant dataset imbalance. In other words, deformation interferograms are far fewer than interferograms dominated by noise. To address this imbalance, researchers have implemented strategies such as “Data Augmentation” to increase the number of deformation interferograms and synthesize deformation images based on physical models. This section examines the effectiveness of these two approaches in mitigating data imbalance and provides a comprehensive overview of current methods for creating datasets aimed at detecting volcanic and earthquake deformations.
4.1. Data Augmentation
It is widely recognized that larger datasets generally result in more robust DL models [117,118]. However, insufficient data often hinder effective training, and the process of manually collecting and labeling large datasets is challenging. To address issues related to small datasets and uneven data distribution—which frequently cause overfitting and poor model generalization—researchers have developed various image data augmentation techniques. As shown in Figure 4, these techniques generate new samples that resemble, but are not identical to, the original data by applying a range of transformation operations. Data augmentation methods can be broadly categorized into four main types:
Geometric transformations: These include operations such as rotation, scaling, flipping, translation, mirroring, and cropping.
Pixel-level transformations: These involve adjustments to contrast, brightness, or color, or the application of histogram equalization to enhance image quality.
Filtering techniques: These encompass methods like sliding windows, median filtering, and Gaussian filtering to process image data.
Generative adversarial networks (GANs): As an emerging approach, GANs have shown significant promise for data augmentation by generating synthetic samples that closely resemble real data [119].
Combining GAN-based augmentation with deformation detection models could open exciting avenues for future research.
Anantrasirichai et al. (2018) [2] tested a pre-trained network model on archived data from the Envisat mission, fine-tuning it for application to the Sentinel dataset, which contained over 30,000 images of 900 volcanoes. The Envisat data included only 300 positive samples (indicating volcanic deformation), while negative samples (containing only noise) outnumbered positives by more than 100-fold. To address this severe imbalance, data augmentation techniques were used to expand the positive sample set. The 300 positive samples from the Envisat dataset were increased to approximately 10,000 using transformations such as panning, flipping (horizontal and vertical), rotation, and warping. This study marked the first application of data augmentation methods for volcanic deformation detection, effectively demonstrating their utility in InSAR imagery. Similarly, Bountos et al. [96] employed self-supervised contrastive learning for volcanic deformation detection, emphasizing the importance of elastic transformations in data augmentation. This approach was particularly effective, likely due to the stripe-like patterns characteristic of InSAR data.
4.2. Dataset Synthesized Based on Physical Models
Although data augmentation expands existing datasets, its effectiveness is constrained by the limited availability of real interferograms. For large-scale DL models, data augmentation alone often falls short of achieving the necessary dataset size and diversity. Furthermore, it may still result in overfitting, as known ground deformation features are insufficient to represent the global diversity of deformation patterns [120]. An alternative solution involves synthesizing datasets based on physical models.
4.2.1. Research on Volcano Dataset Synthesis
In volcanic deformation detection, Valade et al. [94] proposed a method that fused procedural noise with empirical rules to synthesize interferograms, accounting for factors like ground deformation, atmospheric phase delays, geometrical distortions from ortho-rectification, and phase decorrelation due to vegetation. However, these synthesized data lacked a clear physical foundation. In contrast, Anantrasirichai et al. [30] attributed phase changes in interferograms to a combination of ground deformation, stratified atmospheric delays, and turbulent atmospheric delays. By synthesizing data based on a physical model, they effectively incorporated specialized knowledge, providing a more grounded approach to interferogram synthesis. In another study, Anantrasirichai et al. [37] re-wrapped unwrapped InSAR time-series images. By reducing the wrap interval to increase the density of interference fringes, they lowered the detection threshold, enhancing sensitivity. However, excessive reduction in the wrap interval could lead to the CNNs incorrectly identifying features caused by atmospheric noise as ground deformation. Thus, determining the optimal over-wrapping threshold to balance detection sensitivity and accuracy remains a challenge worth further exploration.
Fadhillah et al. [95] not only synthesized deformation, but also simulated noise, including atmospheric disturbances, topographic noise, and orbital errors, based on a physical model. They then combined synthesized deformation with synthesized noise to generate interferograms. Beker et al. [98] utilized variogram modeling to extract statistical parameters for simulating residual atmospheric noise, further improving the accuracy of simulated noise in real data. The latest research, such as that by Gaddes et al. [34], also trains models using synthetic interferogram data. Gaddes et al. developed SyInterferoPy, an open-source Python package that generates volcano deformation based on physical models, considering noise from terrain-related atmospheric delays, turbulent atmospheric delays, and phase decorrelation effects. Abdallah et al. [103] further optimized the SyInterferoPy generator by incorporating orbital errors into the noise model and adding coseismic deformations from various faults into the deformation model.
4.2.2. Research on Earthquake Dataset Synthesis
In earthquake deformation detection, the technology for synthesizing surface deformation data has become relatively mature. Most studies rely on the Okada elastic half-space displacement model [121] to model coseismic deformations. By randomly generating source parameters such as fault strike, dip angle, slip amount, depth, length, and width, researchers can simulate diverse earthquake deformation scenarios. The focus has gradually shifted to methods for synthesizing noise to improve the consistency between simulated and real observed interferograms.
Brengman et al. [104] synthesized atmospheric noise with spatial wavelengths ranging from 10 to 100 km in their study of seismic deformation detection [122]. They also simulated terrain-related noise by scaling the Digital Elevation Model (DEM) and introduced decorrelated noise from vegetation growth and terrain slope variations. This laid the foundation for using synthetic noise to train models in seismic deformation detection. Rouet-Leduc et al. [111] used spatially correlated Gaussian noise to simulate atmospheric turbulence delays at different wavelengths and employed a quadratic function of phase elevation to model topography-related atmospheric delays [122,123]. They also simulated common decorrelation and phase unwrapping errors in real data by adding random pixels to the synthetic data. Li et al. [38] focused on post-earthquake deformation, using synthetic deformation to model the time-series characteristics of post-seismic deformation. After synthesizing ground LOS (Line-of-Sight) deformation from sliding faults, they applied time decay functions to transform single LOS deformation into a time series. They treated data with no significant post-seismic signal before the earthquake as residual noise and used real noise interferograms from before the earthquake for model training. Finally, Zhu et al. [35] proposed a novel method to simulate the multi-scale characteristics of spatially correlated noise for extracting coseismic deformation time series. They combined second-order polynomials, sinusoidal functions, and exponential decay functions to model the spatially correlated noise in InSAR data, adding random noise to enhance the realism of the simulated interferograms.
4.3. Summary of Dataset Creation Methods for Deep Learning to Detect Volcanic and Earthquake Deformation
In DL-based InSAR deformation detection tasks, the availability and quality of data are crucial factors influencing detection performance. The choice of dataset significantly impacts model performance. In studies of volcanic and seismic deformation detection using DL, dataset construction methods can be broadly categorized into two approaches: one involves applying data augmentation techniques to real-world data, while the other generates synthetic data based on physical models. Table 2 compares the performance of various datasets in deformation detection tasks.
Data augmentation techniques enhance the diversity and richness of InSAR images by applying various transformations, thereby improving the generalization ability of DL models. This approach maximizes the use of existing real data by simulating different scenarios, helping the model better adapt to complex environments. However, applying data augmentation in volcanic and seismic deformation detection presents several challenges. One core task of DL models is to distinguish between deformation and noise, and models are highly sensitive to noise. Therefore, selecting appropriate augmentation methods is crucial. Improper techniques may introduce artificial noise, which could disrupt the learning process and reduce detection accuracy. Moreover, the effectiveness of data augmentation depends heavily on the quantity of original data. In this field, real positive samples (i.e., images containing volcanic and seismic deformations) are scarce. Even with augmentation, the amount of new data generated remains limited, restricting its application.
Alternatively, generating synthetic data based on physical models involves constructing models of earthquakes, volcanoes, and noise to produce large volumes of simulated data. The advantage of this approach lies in its flexibility to control the data generation process. Researchers can adjust parameters to generate datasets that meet specific needs. Synthetic data also provide key information that is often difficult to obtain from real data. The model’s generalization ability largely depends on the accuracy of the physical model. Although extensive research has led to effective physical models for simulating ground deformation from earthquakes and volcanic activity, generating synthetic noise remains challenging. Atmospheric noise, due to its complexity and variability, is difficult to model accurately, making precise noise generation a significant challenge in current research.
5. Challenges and Future Research Directions
5.1. InSAR Training Sample Scarcity Problem
Supervised learning plays a critical role in monitoring volcanic and seismic deformation. Networks trained using this approach typically require a large amount of labeled data for parameter training and evaluation. Compared to computer vision tasks, InSAR data samples are relatively scarce due to two main factors: (1) the inherent difficulty in acquiring InSAR data, and (2) the lack of high-quality labeled datasets [129]. To address the challenge of limited training data, researchers have explored several solutions, including data augmentation techniques and synthetic simulation of InSAR data. Future research may focus on the following directions:
Application of GAN-based image data augmentation in InSAR data: GANs have shown great promise in generating high-quality synthetic images. Through adversarial training, GANs produce synthetic data that closely resemble real samples, thus enriching datasets. For instance, Chen et al. [78] utilized GANs to simulate atmospheric delay data, providing a novel approach to alleviating sample scarcity in InSAR analysis.
Development of small-sample DL models: Small-sample learning refers to the ability to build models that can solve real-world problems using a limited number of labeled samples, which is highly valuable in DL applications [130]. While small-sample learning alleviates the problem of limited data, it still faces considerable challenges in addressing data scarcity.
Data-sharing initiatives: Despite efforts to increase the size and diversity of training datasets, current InSAR datasets remain limited. Collaborative sharing of datasets and code among researchers and institutions can accelerate advancements in this field.
5.2. Deep Learning Model Generalization Problem
The generalization capability of DL models is critical for their performance in real-world scenarios. Typically, models trained on InSAR data perform well in specific regions where they have effectively learned the features of the local InSAR interferograms. However, when applied to new regions with partially similar but distinct characteristics, their performance often degrades due to differences in data distribution. To address this issue, the following strategies are proposed:
Transfer learning: Fine-tuning or retraining pre-trained model weights using new samples from the target region can improve the model’s adaptability to new scenarios, thereby enhancing its generalization performance.
Refinement of noise modeling: InSAR images typically contain various noise sources, such as atmospheric noise, decorrelation, unwrapping errors, and orbital errors, which can significantly affect model performance, especially in new scenes. Therefore, improving the model’s ability to accurately model noise is key to enhancing generalization. By analyzing the noise characteristics in different scenes and developing more precise noise models, the model’s performance in new scenarios can be effectively improved.
5.3. Deep Learning Model Interpretability Problem
The interpretability of DL models remains a significant challenge, particularly in scientific fields where understanding the underlying processes is crucial. In contrast to traditional mechanistic models that explicitly represent physical or theoretical processes, DL models leverage complex, multi-layer neural networks that simulate human-like “induction” and “inference” mechanisms to address specific tasks. This data-driven approach often bypasses the exploration of fundamental physical processes inherent in the data, focusing primarily on direct pattern recognition. Although such models can yield impressive performance, their “black-box” nature restricts interpretability, hindering the ability to analyze or trust their internal mechanisms. This limitation presents substantial barriers to the broader adoption of DL in scientific domains that require rigorous validation and transparency [44].
To mitigate this challenge, recent advancements in Explainable AI (XAI) have become increasingly critical. XAI methodologies aim to enhance the transparency of models by providing interpretable insights into decision-making processes, thus enabling researchers and practitioners to gain a deeper understanding of how models arrive at their conclusions [131,132,133,134]. In the context of InSAR data analysis, XAI could facilitate the interpretation of model outputs, aiding researchers in validating the model’s findings and promoting its acceptance within the scientific community. A particularly promising approach involves the integration of physics-based models with contemporary DL techniques. By embedding well-established physical laws into the architecture or training process of DL models, these hybrid approaches can enhance both the interpretability and the robustness of the models, ensuring that their predictions remain consistent with known scientific principles. Consequently, future research should prioritize the development of novel XAI techniques, particularly those that combine mechanistic models with DL approaches, to improve the interpretability and applicability of DL models in geophysical research [135,136,137].
6. Conclusions
This review provides a comprehensive analysis of the advancements and challenges in applying DL techniques to detect volcanic and seismic InSAR deformations. The DL methods, particularly CNNs, have significantly advanced the processing and interpretation of InSAR data; other architectures such as GANs, RNNs, and Transformer models are for addressing specific issues. The success of CNNs in tasks such as denoising, phase unwrapping, and deformation classification has solidified their position as the cornerstone of DL applications in InSAR data processing. GANs have demonstrated strong potential in generating synthetic datasets to mitigate data imbalance, RNNs excel in modeling temporal patterns in deformation time series, and Transformer models provide superior capabilities for capturing global dependencies in InSAR data. However, critical challenges persist, including the scarcity of labeled datasets, the imbalance between deformation and noise samples, the limited generalization capability of DL models across diverse geological settings, and the interpretability challenges posed by the “black-box” nature of DL algorithms, particularly in the geophysical domain.
To address these limitations, future research must prioritize the development of robust benchmark datasets, improved noise modeling techniques, and the integration of physics-based principles to ensure consistency between DL model outputs and known geophysical processes. The exploration of emerging paradigms, such as self-supervised learning, domain adaptation, and transfer learning, offers promising solutions to improve model adaptability and efficiency. Furthermore, fostering open data-sharing platforms, incorporating expert knowledge into dataset construction, and embedding geophysical constraints into DL models will enhance their applicability in real-world scenarios. With these advancements, DL methods are expected to achieve greater accuracy and reliability in volcanic and seismic hazard monitoring, enabling more effective real-time decision making and improved predictions of geological events.
Conceptualization, X.L. and Y.Z.; methodology, X.L. and Y.Z.; resources, X.L., Y.Z. and X.S.; investigation, X.L. and Y.Z.; data curation, X.L., Y.Z., X.S., Z.W., W.G. and G.Z.; writing—original draft preparation, X.L., Y.Z., X.S., Z.W., W.G. and G.Z.; writing—review and editing, Y.Z. and Z.W.; visualization, X.L., Y.Z. and X.S.; supervision, Y.Z., X.S. and Z.W.; project administration, Y.Z. and W.G.; funding acquisition, Y.Z. and Z.W. All authors have read and agreed to the published version of the manuscript.
No new datasets were generated or analyzed during the current study. Data sharing is not applicable to this article.
The authors are grateful to the reviewers for their constructive comments and valuable assistance in improving the manuscript.
The authors declare 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.
Enlarge this image.Figure 1. InSAR data processing based on deep learning. (a) The primary deep learning architectures utilized in InSAR data processing, including CNNs, RNNs, GANs, and Transformers. (b) DL is applied to various stages of InSAR data processing, including deformation detection, atmospheric correction, phase filtering, and phase unwrapping.
Enlarge this image.Figure 2. The main architectures of CNNs, RNNs, GANs, and Transformer networks. (a) CNNs primarily consist of an input layer, convolutional layers, pooling layers, fully connected layers, and an output layer. (b) RNNs consist of input layers, recurrent hidden layers, and an output layer for sequence tasks. (c) GANs consist of a generator and a discriminator, which are trained together in a competitive manner. (d) Transformers consist of an encoder and a decoder, both using self-attention and feed-forward layers.
Enlarge this image.Figure 3. Different learning processes between traditional machine learning and transfer learning. (a) Traditional machine learning approaches learn each task independently, starting from scratch. (b) Transfer learning utilizes knowledge gained from previous tasks and applies it to a target task.
Enlarge this image.Figure 4. Data augmentation methods. (a) Geometric transformation-based data augmentation involves techniques like zoom, rotation, mirroring, and flipping to expand the training datasets. (b) Pixel-level transformation-based data augmentation modifies individual pixel values, such as brightness, contrast, and color, to enhance the datasets. (c) Filtering-based data augmentation involves applying filters like blurring, sharpening, and noise to diversify the training datasets. The original InSAR interferogram data were downloaded from the COMET-LiCS Sentinel-1 InSAR portal (https://comet.nerc.ac.uk/comet-lics-portal/ (accessed on 1 December 2024)).
Application of deep learning networks in InSAR data processing.
| Fundamental Network Architectures | Tasks of InSAR Data | Key Contributions | References |
|---|---|---|---|
| CNNS | Phase filtering | A method based on an InSAR image denoising filter and coherence estimation is proposed | [ |
| Phase filtering | Constructing a CNN model with multi-scale features to suppress noise | [ | |
| Landslide identification and monitoring | Introducing deformable convolutional layers, ResNext blocks, and attention mechanisms to quickly identify landslides | [ | |
| Phase unwrapping | Introducing interference coherence as a model input feature to improve phase unwrapping accuracy | [ | |
| Atmospheric correction | Proposing an Attentional Deep Residual U-Net (ARU-Net) to mitigate atmospheric noise | [ | |
| Subsidence monitoring and phase unwrapping | A dual-model approach is proposed for mine subsidence monitoring and phase unwrapping | [ | |
| RNNs | Land subsidence prediction | Proving that LSTM is superior to CNN in monitoring land subsidence | [ |
| Deformation and atmospheric delay prediction | Predicting the deformation and atmospheric delay phase of the GB-InSAR time series at the next moment | [ | |
| Atmospheric correction | A bidirectional RNN with a gated recurrent unit (GRU) is proposed to remove atmospheric noise from InSAR time series | [ | |
| GANs | Phase unwrapping | Proposing PU-GAN for 2D phase wrapping | [ |
| Atmospheric correction | Using Pix2Pix GAN to effectively reduce the tropospheric phase delay | [ | |
| Deformation synthesis | InSAR deformation signal synthesis | [ | |
| Atmospheric correction | Successfully generating realistic fringe patterns that closely resemble the patterns of InSAR interferograms | [ | |
| Transformers | Phase filtering | Introducing deformable convolution to further extract local phase features | [ |
| Data interpolation | Increasing the spatial density of surface motion samples | [ | |
| InSAR data reconstruction | Reconstructing permafrost regions using InSAR data | [ |
Comparison of various DL-based deformation detection methods and the datasets used.
| References | Data | ||||||
|---|---|---|---|---|---|---|---|
| Type | Wrap | Task | Train Set (Synth/Real) | TP 1/P 2 | FP 3/N 4 | AUC 5 | |
| [ | TS 6 | Unwrapped | Denoising | ? 7/0 | - | - | - |
| [ | IFG 8 | Wrapped | Classification | 0/15,125 | 38/42 | 242/30,207 | 0.995 1 |
| [ | IFG | Wrapped | Classification | 910,000/0 | 38/42 | 295/30,207 | 0.9312 |
| [ | TS | Unwrapped | Denoising | ?/0 | - | - | - |
| [ | IFG | Wrapped | Classification | 330,848/836 | 124/142 | 201/633 | 0.8644 |
| [ | IFG | Both | Classification | 0/15,478 | /252 | 3826 | 0.864 |
| [ | IFG | Unwrapped | Classification, | 20,000/0 | 8/20 | 6/32 | - |
| 20,000/20,000 | 16/20 | 5/32 | - | ||||
| [ | TS | Unwrapped | Denoising | 225,000,000/0 | - | - | - |
| [ | IFG | Wrapped | Classification | 0/20,000 | 42/42 | 1327/30,207 | 0.882 |
| [ | IFG | Wrapped | Classification | 0/7536 | 31/32 | 4/32 | - |
| [ | IFG | Wrapped | Classification | 0/>1,200,000 | 699/836 | 176/1648 | - |
| [ | TS | Wrapped | Classification | 910,000/0 | ?/? | ?/? | 0.9894 |
1 True positive (TP): These are the cases where the model correctly predicts the positive class. 2 Positive (P): The sample in the datasets that belongs to the positive class. 3 False positive (FP): These are the cases where the model incorrectly predicts the positive class. 4 Negative (N): The sample in the datasets that belongs to the negative class. 5 Area Under the Receiver Operating Characteristic Curve (AUC): A widely used metric to evaluate the performance of classification models. 6 TS: InSAR time-series interferograms. 7 ?: No specific data are mentioned or provided in the article. 8 IFG: Single InSAR interferogram.
References
1. Sun, Q.; Zhang, L.; Ding, X.; Hu, J.; Liang, H. Investigation of slow-moving landslides from ALOS/PALSAR images with TCPInSAR: A case study of Oso, USA. Remote Sens.; 2014; 7, pp. 72-88. [DOI: https://dx.doi.org/10.3390/rs70100072]
2. Anantrasirichai, N.; Biggs, J.; Albino, F.; Hill, P.; Bull, D. Application of machine learning to classification of volcanic deformation in routinely generated InSAR data. J. Geophys. Res. Solid Earth; 2018; 123, pp. 6592-6606. [DOI: https://dx.doi.org/10.1029/2018JB015911]
3. Wu, S.; Yang, Z.; Ding, X.; Zhang, B.; Zhang, L.; Lu, Z. Two decades of settlement of Hong Kong International Airport measured with multi-temporal InSAR. Remote Sens. Environ.; 2020; 248, 111976. [DOI: https://dx.doi.org/10.1016/j.rse.2020.111976]
4. Kang, Y.; Lu, Z.; Zhao, C.; Xu, Y.; Kim, J.-W.; Gallegos, A.J. InSAR monitoring of creeping landslides in mountainous regions: A case study in Eldorado National Forest, California. Remote Sens. Environ.; 2021; 258, 112400. [DOI: https://dx.doi.org/10.1016/j.rse.2021.112400]
5. Liu, Z.; Qiu, H.; Zhu, Y.; Liu, Y.; Yang, D.; Ma, S.; Zhang, J.; Wang, Y.; Wang, L.; Tang, B. Efficient identification and monitoring of landslides by time-series InSAR combining single-and multi-look phases. Remote Sens.; 2022; 14, 1026. [DOI: https://dx.doi.org/10.3390/rs14041026]
6. Yang, F.; Jiang, Z.; Ren, J.; Lv, J. Monitoring, prediction, and evaluation of mountain geological hazards based on InSAR technology. Sci. Program.; 2022; 2022, 2227049. [DOI: https://dx.doi.org/10.1155/2022/2227049]
7. Zheng, Z.; Xie, C.; He, Y.; Zhu, M.; Huang, W.; Shao, T. Monitoring potential geological hazards with different InSAR algorithms: The case of western Sichuan. Remote Sens.; 2022; 14, 2049. [DOI: https://dx.doi.org/10.3390/rs14092049]
8. Zhong, J.; Li, Q.; Zhang, J.; Luo, P.; Zhu, W. Risk Assessment of Geological Landslide Hazards Using D-InSAR and Remote Sensing. Remote Sens.; 2024; 16, 345. [DOI: https://dx.doi.org/10.3390/rs16020345]
9. Silva, B.; Sousa, J.J.; Lazecky, M.; Cunha, A. Deformation fringes detection in SAR interferograms using deep learning. Procedia Comput. Sci.; 2022; 196, pp. 151-158. [DOI: https://dx.doi.org/10.1016/j.procs.2021.11.084]
10. Motlagh, A.B. Deep Learning Application to Detect Earth’s Surface Deformation from Earthquakes with SAR Interferometric Techniques. Master’s Thesis; Faculty of Sciences, University of Porto: Porto, Portugal, 2023.
11. Wang, Y.; Dong, J.; Zhang, L.; Deng, S.; Zhang, G.; Liao, M.; Gong, J. Automatic detection and update of landslide inventory before and after impoundments at the Lianghekou reservoir using Sentinel-1 InSAR. Int. J. Appl. Earth Obs. Geoinf.; 2023; 118, 103224. [DOI: https://dx.doi.org/10.1016/j.jag.2023.103224]
12. Yang, C.; Wen, Y.; Pan, X.; Yuan, D. Monitoring and analysis of landslide deformation based on SBAS-InSAR. Bull. Surv. Mapp.; 2023; pp. 12-17. [DOI: https://dx.doi.org/10.13474/j.cnki.11-2246.2023.0320]
13. Chen, X.; Zhang, F.; Wang, X.; Guo, F.; Jiang, Y. Research on the landslides disaster based on time series InSAR technique. Eng. Surv. Mapp.; 2020; 29, pp. 45-49.
14. Zhang, L.; Liao, M.; Dong, J.; Xu, Q.; Gong, J. Early Detection of Landslide Hazards in Mountainous Areas of West China Using Time Series SAR Interferometry—A Case Study of Danba, Sichuan. Geomat. Inf. Sci. Wuhan Univ.; 2018; 43, pp. 2039-2049.
15. Liao, M.; Dong, J.; Li, M.; Ao, M.; Zhang, L.; Shi, X. Radar remote sensing for potential landslidesdetection and deformation monitoring. Natl. Remote Sens. Bull.; 2021; 25, pp. 332-341. [DOI: https://dx.doi.org/10.11834/jrs.20210162]
16. Zhang, P.; Guo, Z.; Guo, S.; Xia, J. Land subsidence monitoring method in regions of variable radar reflection characteristics by integrating PS-InSAR and SBAS-InSAR techniques. Remote Sens.; 2022; 14, 3265. [DOI: https://dx.doi.org/10.3390/rs14143265]
17. Khan, J.; Ren, X.; Hussain, M.A.; Jan, M.Q. Monitoring land subsidence using PS-InSAR technique in Rawalpindi and Islamabad, Pakistan. Remote Sens.; 2022; 14, 3722. [DOI: https://dx.doi.org/10.3390/rs14153722]
18. Cigna, F.; Tapete, D. Present-day land subsidence rates, surface faulting hazard and risk in Mexico City with 2014–2020 Sentinel-1 IW InSAR. Remote Sens. Environ.; 2021; 253, 112161. [DOI: https://dx.doi.org/10.1016/j.rse.2020.112161]
19. Gao, E.; Fan, D.; Fu, B.; Yong, Q.; Lan, Y. Land Subsidence Monitoring of Nanjing Area Based on PS-InSAR and SBAS Technology. J. Geod. Geodyn.; 2019; 39, pp. 158-163.
20. Pu, C.; Xu, Q.; Jiang, Y.; Zhao, K.; He, P.; Zhang, H.; Li, H. Analysis of Land Subsidence Distribution and Influencing Factors in Yan’an New District Based on Time Series InSAR. Geomat. Inf. Sci. Wuhan Univ.; 2020; 45, pp. 1728-1738.
21. Zhao, R.; Li, Z.-W.; Feng, G.-C.; Wang, Q.-J.; Hu, J. Monitoring surface deformation over permafrost with an improved SBAS-InSAR algorithm: With emphasis on climatic factors modeling. Remote Sens. Environ.; 2016; 184, pp. 276-287. [DOI: https://dx.doi.org/10.1016/j.rse.2016.07.019]
22. Yu, C.; Li, Z.; Chen, J.; Hu, J.-C. Small magnitude co-seismic deformation of the 2017 Mw6.4 Nyingchi earthquake revealed by InSAR measurements with atmospheric correction. Remote Sens.; 2018; 10, 684. [DOI: https://dx.doi.org/10.3390/rs10050684]
23. Suresh, D.; Yarrakula, K. InSAR based deformation mapping of earthquake using Sentinel 1A imagery. Geocarto Int.; 2020; 35, pp. 559-568. [DOI: https://dx.doi.org/10.1080/10106049.2018.1544289]
24. Liu, C.; Bai, L.; Hong, S.; Dong, Y.; Jiang, Y.; Li, H.; Zhan, H.; Chen, Z. Coseismic deformation of the 2021 Mw7.4 Maduo earthquake from joint inversion of InSAR, GPS, and teleseismic data. Earthq. Sci.; 2021; 34, pp. 436-446. [DOI: https://dx.doi.org/10.29382/eqs-2021-0050]
25. Zhao, L.; Chen, Z.; Xie, L.; Zhu, Z.; Xu, W. Coseismic deformation and slip model of the 2024 MW7.0 Wushi earthquake obtained from InSAR observation. Rev. Geophys. Planet. Phys.; 2024; 55, pp. 453-460.
26. Qu, C.; Zhao, D.; Shan, X.; Yue, C.; Chen, H.; Wu, D.; Liu, L.; Zhang, G. Coseismic and postseismic deformation of the 2001 MW7.8 Kunlun Mountain earthquake and its loading effect on the 2021 MW7.3 Madoi earthquake. Chin. J. Geophys.; 2023; 66, pp. 2741-2756.
27. Gündüz, H.İ.; Yılmaztürk, F.; Orhan, O. An investigation of volcanic ground deformation using InSAR observations at Tendürek Volcano (Turkey). Appl. Sci.; 2023; 13, 6787. [DOI: https://dx.doi.org/10.3390/app13116787]
28. Spaans, K.; Hooper, A. InSAR processing for volcano monitoring and other near-real time applications. J. Geophys. Res. Solid Earth; 2016; 121, pp. 2947-2960. [DOI: https://dx.doi.org/10.1002/2015JB012752]
29. Bemelmans, M.J.; Biggs, J.; Poland, M.; Wookey, J.; Ebmeier, S.; Diefenbach, A.K.; Syahbana, D. High-resolution InSAR reveals localized pre-eruptive deformation inside the crater of Agung Volcano, Indonesia. J. Geophys. Res. Solid Earth; 2023; 128, e2022JB025669. [DOI: https://dx.doi.org/10.1029/2022JB025669]
30. Anantrasirichai, N.; Biggs, J.; Albino, F.; Bull, D. A deep learning approach to detecting volcano deformation from satellite imagery using synthetic datasets. Remote Sens. Environ.; 2019; 230, 111179. [DOI: https://dx.doi.org/10.1016/j.rse.2019.04.032]
31. Liu, Y.; Yao, X.; Gu, Z.; Zhou, Z.; Liu, X.; Chen, X.; Wei, S. Study of the automatic recognition of landslides by using InSAR images and the improved mask R-CNN model in the Eastern Tibet Plateau. Remote Sens.; 2022; 14, 3362. [DOI: https://dx.doi.org/10.3390/rs14143362]
32. Khalili, M.A.; Coda, S.; Calcaterra, D.; Di Martire, D. Synergistic use of SAR satellites with deep learning model interpolation for investigating of active landslides in Cuenca, Ecuador. Geomat. Nat. Hazards Risk; 2024; 15, 2383270. [DOI: https://dx.doi.org/10.1080/19475705.2024.2383270]
33. Cai, J.; Zhang, L.; Dong, J.; Guo, J.; Wang, Y.; Liao, M. Automatic identification of active landslides over wide areas from time-series InSAR measurements using Faster RCNN. Int. J. Appl. Earth Obs. Geoinf.; 2023; 124, 103516. [DOI: https://dx.doi.org/10.1016/j.jag.2023.103516]
34. Gaddes, M.; Hooper, A.; Albino, F. Simultaneous classification and location of volcanic deformation in SAR interferograms using a convolutional neural network. Earth Space Sci.; 2024; 11, e2024EA003679. [DOI: https://dx.doi.org/10.1029/2024EA003679]
35. Zhu, C.; Li, X.; Wang, C.; Zhang, B.; Li, B. Deep Learning-Based Coseismic Deformation Estimation from InSAR Interferograms. IEEE Trans. Geosci. Remote Sens.; 2024; 62, 5203610. [DOI: https://dx.doi.org/10.1109/TGRS.2024.3357190]
36. Reinisch, E.C.; Abolt, C.J.; Swanson, E.M.; Rouet-Leduc, B.; Snyder, E.E.; Sivaraj, K.; Solander, K.C. Advancing the Limits of InSAR to Detect Crustal Displacement from Low-Magnitude Earthquakes through Deep Learning. Remote Sens.; 2024; 16, 2019. [DOI: https://dx.doi.org/10.3390/rs16112019]
37. Anantrasirichai, N.; Biggs, J.; Albino, F.; Bull, D. The application of convolutional neural networks to detect slow, sustained deformation in InSAR time series. Geophys. Res. Lett.; 2019; 46, pp. 11850-11858. [DOI: https://dx.doi.org/10.1029/2019GL084993]
38. Li, C.; Xi, X.; Zhang, G.; Song, X.; Shan, X. A Deep-Learning Neural Network for Postseismic Deformation Reconstruction from InSAR Time Series. IEEE Trans. Geosci. Remote Sens.; 2024; 62, 4505214. [DOI: https://dx.doi.org/10.1109/TGRS.2024.3389772]
39. Han, J.; Yang, H.; Liu, Y.; Lu, Z.; Zeng, K.; Jiao, R. A deep learning application for deformation prediction from ground-based insar. Remote Sens.; 2022; 14, 5067. [DOI: https://dx.doi.org/10.3390/rs14205067]
40. Ma, P.; Zhang, F.; Lin, H. Prediction of InSAR time-series deformation using deep convolutional neural networks. Remote Sens. Lett.; 2020; 11, pp. 137-145. [DOI: https://dx.doi.org/10.1080/2150704X.2019.1692390]
41. Radman, A.; Akhoondzadeh, M.; Hosseiny, B. Integrating InSAR and deep-learning for modeling and predicting subsidence over the adjacent area of Lake Urmia, Iran. GIScience Remote Sens.; 2021; 58, pp. 1413-1433. [DOI: https://dx.doi.org/10.1080/15481603.2021.1991689]
42. Goodfellow, I. Deep Learning; MIT-Press: Cambridge, MA, USA, 2016.
43. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature; 2015; 521, pp. 436-444. [DOI: https://dx.doi.org/10.1038/nature14539] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/26017442]
44. Sarker, I.H. Deep learning: A comprehensive overview on techniques, taxonomy, applications and research directions. SN Comput. Sci.; 2021; 2, 420. [DOI: https://dx.doi.org/10.1007/s42979-021-00815-1]
45. Simonyan, K.; Zisserman, A. Very deep convolutional networks for large-scale image recognition. arXiv; 2014; arXiv: 1409.1556
46. Szegedy, C.; Liu, W.; Jia, Y.; Sermanet, P.; Reed, S.; Anguelov, D.; Erhan, D.; Vanhoucke, V.; Rabinovich, A. Going deeper with convolutions. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Boston, MA, USA, 7–12 June 2015; pp. 1-9.
47. He, K.; Zhang, X.; Ren, S.; Sun, J. Deep residual learning for image recognition. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Las Vegas, NV, USA, 27–30 June 2016; pp. 770-778.
48. Amodei, D.; Ananthanarayanan, S.; Anubhai, R.; Bai, J.; Battenberg, E.; Case, C.; Casper, J.; Catanzaro, B.; Cheng, Q.; Chen, G. Deep speech 2: End-to-end speech recognition in english and mandarin. Proceedings of the International Conference On Machine Learning; New York, NY, USA, 19–24 June 2016; pp. 173-182.
49. Hinton, G.; Deng, L.; Yu, D.; Dahl, G.E.; Mohamed, A.-R.; Jaitly, N.; Senior, A.; Vanhoucke, V.; Nguyen, P.; Sainath, T.N. Deep neural networks for acoustic modeling in speech recognition: The shared views of four research groups. IEEE Signal Process. Mag.; 2012; 29, pp. 82-97. [DOI: https://dx.doi.org/10.1109/MSP.2012.2205597]
50. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. Proceedings of the 26th Annual Conference on Neural Information Processing Systems (NIPS 2012); Lake Tahoe, NV, USA, 3–6 December 2012; Volume 25.
51. Farabet, C.; Couprie, C.; Najman, L.; LeCun, Y. Learning hierarchical features for scene labeling. IEEE Trans. Pattern Anal. Mach. Intell.; 2012; 35, pp. 1915-1929. [DOI: https://dx.doi.org/10.1109/TPAMI.2012.231]
52. Tompson, J.J.; Jain, A.; LeCun, Y.; Bregler, C. Joint training of a convolutional network and a graphical model for human pose estimation. Proceedings of the 2014 Annual Conference on Neural Information Processing Systems (NIPS 2014); Montreal, QC, Canada, 8–13 December 2014; Volume 27.
53. Mikolov, T.; Deoras, A.; Povey, D.; Burget, L.; Černocký, J. Strategies for training large scale neural network language models. Proceedings of the 2011 IEEE Workshop on Automatic Speech Recognition & Understanding; Hawaii, HI, USA, 11–15 December 2011; pp. 196-201.
54. Sainath, T.N.; Kingsbury, B.; Mohamed, A.-R.; Dahl, G.E.; Saon, G.; Soltau, H.; Beran, T.; Aravkin, A.Y.; Ramabhadran, B. Improvements to deep convolutional neural networks for LVCSR. Proceedings of the 2013 IEEE Workshop on Automatic Speech Recognition and Understanding; Olomouc, Czech Republic, 8–12 December 2013; pp. 315-320.
55. Cong, S.; Zhou, Y. A review of convolutional neural network architectures and their optimizations. Artif. Intell. Rev.; 2023; 56, pp. 1905-1969. [DOI: https://dx.doi.org/10.1007/s10462-022-10213-5]
56. O’Shea, K. An introduction to convolutional neural networks. arXiv; 2015; arXiv: 1511.08458
57. Krichen, M. Convolutional neural networks: A survey. Computers; 2023; 12, 151. [DOI: https://dx.doi.org/10.3390/computers12080151]
58. Pandian, J.A.; Kanchanadevi, K.; Kumar, V.D.; Jasińska, E.; Goňo, R.; Leonowicz, Z.; Jasiński, M. A five convolutional layer deep convolutional neural network for plant leaf disease detection. Electronics; 2022; 11, 1266. [DOI: https://dx.doi.org/10.3390/electronics11081266]
59. Lu, T.-C. CNN Convolutional layer optimisation based on quantum evolutionary algorithm. Connect. Sci.; 2021; 33, pp. 482-494. [DOI: https://dx.doi.org/10.1080/09540091.2020.1841111]
60. Galanis, N.-I.; Vafiadis, P.; Mirzaev, K.-G.; Papakostas, G.A. Convolutional neural networks: A roundup and benchmark of their pooling layer variants. Algorithms; 2022; 15, 391. [DOI: https://dx.doi.org/10.3390/a15110391]
61. Jie, H.J.; Wanda, P. RunPool: A dynamic pooling layer for convolution neural network. Int. J. Comput. Intell. Syst.; 2020; 13, pp. 66-76. [DOI: https://dx.doi.org/10.2991/ijcis.d.200120.002]
62. Elman, J.L. Finding structure in time. Cogn. Sci.; 1990; 14, pp. 179-211. [DOI: https://dx.doi.org/10.1207/s15516709cog1402_1]
63. Hochreiter, S. Long Short-term Memory. Neural Computation; MIT-Press: Cambridge, MA, USA, 1997.
64. Chung, J.; Gulcehre, C.; Cho, K.; Bengio, Y. Empirical evaluation of gated recurrent neural networks on sequence modeling. arXiv; 2014; arXiv: 1412.3555
65. Goodfellow, I.; Pouget-Abadie, J.; Mirza, M.; Xu, B.; Warde-Farley, D.; Ozair, S.; Courville, A.; Bengio, Y. Generative adversarial nets. Proceedings of the 2014 Annual Conference on Neural Information Processing Systems (NIPS 2014); Montreal, QC, Canada, 8–13 December 2014; Volume 27.
66. Radford, A. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv; 2015; arXiv: 1511.06434
67. Mirza, M. Conditional generative adversarial nets. arXiv; 2014; arXiv: 1411.1784
68. Zhu, J.-Y.; Park, T.; Isola, P.; Efros, A.A. Unpaired image-to-image translation using cycle-consistent adversarial networks. Proceedings of the IEEE International Conference on Computer Vision; Honolulu, HI, USA, 21–26 July 2017; pp. 2223-2232.
69. Arjovsky, M.; Chintala, S.; Bottou, L. Wasserstein generative adversarial networks. Proceedings of the International Conference on Machine Learning; Sydney, Australia, 6–11 August 2017; pp. 214-223.
70. Brock, A. Large Scale GAN Training for High Fidelity Natural Image Synthesis. arXiv; 2018; arXiv: 1809.11096
71. Karras, T.; Laine, S.; Aila, T. A style-based generator architecture for generative adversarial networks. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; Long Beach, CA, USA, 15–20 June 2019; pp. 4401-4410.
72. Vaswani, A. Attention is all you need. Proceedings of the 31st Conference on Neural Information Processing Systems (NIPS 2017); Long Beach, CA, USA, 4–9 December 2017.
73. Carion, N.; Massa, F.; Synnaeve, G.; Usunier, N.; Kirillov, A.; Zagoruyko, S. End-to-end object detection with transformers. Proceedings of the European Conference on Computer Vision; Glasgow, UK, 23–28 August 2020; pp. 213-229.
74. Zhang, Q.; Lu, H.; Sak, H.; Tripathi, A.; McDermott, E.; Koo, S.; Kumar, S. Transformer transducer: A streamable speech recognition model with transformer encoders and RNN-T loss. Proceedings of the ICASSP 2020—2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP); Barcelona, Spain, 4–8 May 2020; pp. 7829-7833.
75. Mukherjee, S.; Zimmer, A.; Kottayil, N.K.; Sun, X.; Ghuman, P.; Cheng, I. CNN-based InSAR denoising and coherence metric. Proceedings of the 2018 IEEE Sensors; New Delhi, India, 28–31 October 2018; pp. 1-4.
76. Yang, W.; He, Y.; Yao, S.; Zhang, L.; Cao, S.; Wen, Z. An InSAR interferogram filtering method based on multi-level feature fusion CNN. Sensors; 2022; 22, 5956. [DOI: https://dx.doi.org/10.3390/s22165956]
77. Sica, F.; Calvanese, F.; Scarpa, G.; Rizzoli, P. A CNN-based coherence-driven approach for InSAR phase unwrapping. IEEE Geosci. Remote Sens. Lett.; 2020; 19, 4003705. [DOI: https://dx.doi.org/10.1109/LGRS.2020.3029565]
78. Chen, Y.; Bruzzone, L.; Jiang, L.; Sun, Q. ARU-net: Reduction of atmospheric phase screen in SAR interferometry using attention-based deep residual U-net. IEEE Trans. Geosci. Remote Sens.; 2020; 59, pp. 5780-5793. [DOI: https://dx.doi.org/10.1109/TGRS.2020.3021765]
79. Wu, Z.; Wang, T.; Wang, Y.; Wang, R.; Ge, D. Deep learning for the detection and phase unwrapping of mining-induced deformation in large-scale interferograms. IEEE Trans. Geosci. Remote Sens.; 2021; 60, 5216318. [DOI: https://dx.doi.org/10.1109/TGRS.2021.3121907]
80. Chen, Y.; He, Y.; Zhang, L.; Chen, Y.; Pu, H.; Chen, B.; Gao, L. Prediction of InSAR deformation time-series using a long short-term memory neural network. Int. J. Remote Sens.; 2021; 42, pp. 6919-6942. [DOI: https://dx.doi.org/10.1080/01431161.2021.1947540]
81. Zhao, Z.; Wu, Z.; Zheng, Y.; Ma, P. Recurrent neural networks for atmospheric noise removal from InSAR time series with missing values. ISPRS J. Photogramm. Remote Sens.; 2021; 180, pp. 227-237. [DOI: https://dx.doi.org/10.1016/j.isprsjprs.2021.08.009]
82. Zhou, L.; Yu, H.; Pascazio, V.; Xing, M. PU-GAN: A one-step 2-D InSAR phase unwrapping based on conditional generative adversarial network. IEEE Trans. Geosci. Remote Sens.; 2022; 60, 5221510. [DOI: https://dx.doi.org/10.1109/TGRS.2022.3145342]
83. Ghosh, B.; Haghshenas Haghighi, M.; Motagh, M.; Maghsudi, S. Using generative adversarial networks for extraction of insar signals from large-scale Sentinel-1 interferograms by improving tropospheric noise correction. ISPRS Ann. Photogramm. Remote Sens. Spat. Inf. Sci.; 2021; 3, pp. 57-64. [DOI: https://dx.doi.org/10.5194/isprs-annals-V-3-2021-57-2021]
84. Zhou, Z.; Sun, X.; Yang, F.; Wang, Z.; Goldsbury, R.; Cheng, I. GANInSAR: Deep Generative Modeling for Large-Scale InSAR Signal Simulation. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens.; 2024; 17, pp. 5303-5316. [DOI: https://dx.doi.org/10.1109/JSTARS.2024.3361444]
85. Rongier, G.; Rude, C.M.; Herring, T.; Pankratius, V. An attempt at improving atmospheric corrections in InSAR using cycle-consistent adversarial networks. arXiv; 2020; [DOI: https://dx.doi.org/10.31223/X5M594]
86. Zheng, S.; Zhang, X.; Pu, L.; Hu, Y.; Shi, J.; Wei, S. An Insar Phase Filtering Method based on Transformer Network. Proceedings of the IGARSS 2022—2022 IEEE International Geoscience and Remote Sensing Symposium; Kuala Lumpur, Malaysia, 17–22 July 2022; pp. 1284-1287.
87. Orlandi, D.; Galatolo, F.A.; Cimino, M.G.; La Rosa, A.; Pagli, C.; Perilli, N. Enhancing land subsidence awareness via InSAR data and Deep Transformers. Proceedings of the 2022 IEEE Conference on Cognitive and Computational Aspects of Situation Management (CogSIMA); Salerno, Italy, 6–10 June 2022; pp. 98-103.
88. Wang, J.; Li, C.; Li, L.; Huang, Z.; Wang, C.; Zhang, H.; Zhang, Z. InSAR time-series deformation forecasting surrounding Salt Lake using deep transformer models. Sci. Total Environ.; 2023; 858, 159744. [DOI: https://dx.doi.org/10.1016/j.scitotenv.2022.159744]
89. Zhuang, F.; Qi, Z.; Duan, K.; Xi, D.; Zhu, Y.; Zhu, H.; Xiong, H.; He, Q. A comprehensive survey on transfer learning. Proc. IEEE; 2020; 109, pp. 43-76. [DOI: https://dx.doi.org/10.1109/JPROC.2020.3004555]
90. Pan, S.J.; Yang, Q. A survey on transfer learning. IEEE Trans. Knowl. Data Eng.; 2009; 22, pp. 1345-1359. [DOI: https://dx.doi.org/10.1109/TKDE.2009.191]
91. Weiss, K.; Khoshgoftaar, T.M.; Wang, D. A survey of transfer learning. J. Big Data; 2016; 3, 9. [DOI: https://dx.doi.org/10.1186/s40537-016-0043-6]
92. Torrey, L.; Shavlik, J. Transfer learning. Handbook of Research on Machine Learning Applications and Trends: Algorithms, Methods, and Techniques; IGI Global: Hershey, PA, USA, 2010; pp. 242-264.
93. Vrbančič, G.; Podgorelec, V. Transfer learning with adaptive fine-tuning. IEEE Access; 2020; 8, pp. 196197-196211. [DOI: https://dx.doi.org/10.1109/ACCESS.2020.3034343]
94. Valade, S.; Ley, A.; Massimetti, F.; D’Hondt, O.; Laiolo, M.; Coppola, D.; Loibl, D.; Hellwich, O.; Walter, T.R. Towards global volcano monitoring using multisensor sentinel missions and artificial intelligence: The MOUNTS monitoring system. Remote Sens.; 2019; 11, 1528. [DOI: https://dx.doi.org/10.3390/rs11131528]
95. Fadhillah, M.F.; Hakim, W.L.; Park, S.; Kim, D.; Park, Y.-C.; Kim, C.-H.; Lee, C.-W. Surface deformation simulation for InSAR detection using a machine learning approach on the hantangang river volcanic field: A case study on the orisan mountain. Front. Environ. Sci.; 2022; 10, 968120. [DOI: https://dx.doi.org/10.3389/fenvs.2022.968120]
96. Bountos, N.I.; Papoutsis, I.; Michail, D.; Anantrasirichai, N. Self-supervised contrastive learning for volcanic unrest detection. IEEE Geosci. Remote Sens. Lett.; 2021; 19, 3003905. [DOI: https://dx.doi.org/10.1109/LGRS.2021.3104506]
97. Chen, T.; Kornblith, S.; Norouzi, M.; Hinton, G. A simple framework for contrastive learning of visual representations. Proceedings of the International Conference on Machine Learning; Shenzhen, China, 15–17 February 2020; pp. 1597-1607.
98. Beker, T.; Ansari, H.; Montazeri, S.; Song, Q.; Zhu, X.X. Deep learning for subtle volcanic deformation detection with InSAR data in central volcanic zone. IEEE Trans. Geosci. Remote Sens.; 2023; 61, 5218520. [DOI: https://dx.doi.org/10.1109/TGRS.2023.3318469]
99. Lobo, J.M.; Jiménez-Valverde, A.; Real, R. AUC: A misleading measure of the performance of predictive distribution models. Glob. Ecol. Biogeogr.; 2008; 17, pp. 145-151. [DOI: https://dx.doi.org/10.1111/j.1466-8238.2007.00358.x]
100. Metz, C.E. Basic principles of ROC analysis. Seminars in Nuclear Medicine; WB Saunders: Philadelphia, PA, USA, 1978; pp. 283-298.
101. Redmon, J. You only look once: Unified, real-time object detection. Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition; Las Vegas, NV, USA, 27–30 June 2016.
102. He, K.; Gkioxari, G.; Dollár, P.; Girshick, R. Mask R-CNN. Proceedings of the IEEE International Conference on Computer Vision; Honolulu, HI, USA, 21–26 July 2017; pp. 2961-2969.
103. Abdallah, M.; Younis, S.; Wu, S.; Ding, X. Automated deformation detection and interpretation using InSAR data and a multi-task ViT model. Int. J. Appl. Earth Obs. Geoinf.; 2024; 128, 103758. [DOI: https://dx.doi.org/10.1016/j.jag.2024.103758]
104. Brengman, C.M.; Barnhart, W.D. Identification of surface deformation in InSAR using machine learning. Geochem. Geophys. Geosyst.; 2021; 22, e2020GC009204. [DOI: https://dx.doi.org/10.1029/2020GC009204]
105. Lu, Z.; Jiang, X.; Kot, A. Deep coupled resnet for low-resolution face recognition. IEEE Signal Process. Lett.; 2018; 25, pp. 526-530. [DOI: https://dx.doi.org/10.1109/LSP.2018.2810121]
106. Targ, S.; Almeida, D.; Lyman, K. Resnet in resnet: Generalizing residual architectures. arXiv; 2016; arXiv: 1603.08029
107. Yamazaki, M.; Kasagi, A.; Tabuchi, A.; Honda, T.; Miwa, M.; Fukumoto, N.; Tabaru, T.; Ike, A.; Nakashima, K. Yet another accelerated SGD: Resnet-50 training on ImageNet in 74.7 seconds. arXiv; 2019; arXiv: 1903.12650
108. Bazzani, L.; Bergamo, A.; Anguelov, D.; Torresani, L. Self-taught object localization with deep networks. Proceedings of the 2016 IEEE Winter Conference on Applications of Computer Vision (WACV); Lake Placid, NY, USA, 7–10 March 2016; pp. 1-9.
109. Zeiler, M. Visualizing and Understanding Convolutional Networks. Proceedings of the 13th European Conference on Computer Vision—ECCV 2014; Zurich, Switzerland, 6–12 September 2014.
110. da Silva, B.M.F. Automatic Detection of Earthquake’s Deformations in SAR Interferograms. Master’s Thesis; Faculty of Sciences, University of Porto: Porto, Portugal, 2021.
111. Rouet-Leduc, B.; Jolivet, R.; Dalaison, M.; Johnson, P.A.; Hulbert, C. Autonomous extraction of millimeter-scale deformation in InSAR time series using deep learning. Nat. Commun.; 2021; 12, 6480. [DOI: https://dx.doi.org/10.1038/s41467-021-26254-3] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/34759266]
112. Montagnon, T.; Hollingsworth, J.; Pathier, E.; Marchandon, M.; Dalla Mura, M.; Giffard-Roisin, S.; Dalla, M. GeoFlowNet: Fast and Accurate Sub-pixel Displacement Estimation From Optical Satellite Images based on Deep Learning. J. Latex Cl. Files; 2024; 13, hal-04782000.
113. Dosovitskiy, A. An image is worth 16 × 16 words: Transformers for image recognition at scale. arXiv; 2020; arXiv: 2010.11929
114. Han, K.; Wang, Y.; Tian, Q.; Guo, J.; Xu, C.; Xu, C. Ghostnet: More features from cheap operations. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition; New Orleans, LA, USA, 18–24 June 2020; pp. 1580-1589.
115. Han, K.; Wang, Y.; Zhang, Q.; Zhang, W.; Xu, C.; Zhang, T. Model rubik’s cube: Twisting resolution, depth and width for tinynets. Adv. Neural Inf. Process. Syst.; 2020; 33, pp. 19353-19364.
116. Han, K.; Wang, Y.; Chen, H.; Chen, X.; Guo, J.; Liu, Z.; Tang, Y.; Xiao, A.; Xu, C.; Xu, Y. A survey on vision transformer. IEEE Trans. Pattern Anal. Mach. Intell.; 2022; 45, pp. 87-110. [DOI: https://dx.doi.org/10.1109/TPAMI.2022.3152247]
117. Halevy, A.; Norvig, P.; Pereira, F. The unreasonable effectiveness of data. IEEE Intell. Syst.; 2009; 24, pp. 8-12. [DOI: https://dx.doi.org/10.1109/MIS.2009.36]
118. Sun, C.; Shrivastava, A.; Singh, S.; Gupta, A. Revisiting unreasonable effectiveness of data in deep learning era. Proceedings of the IEEE International Conference on Computer Vision; Honolulu, HI, USA, 21–26 July 2017; pp. 843-852.
119. Shorten, C.; Khoshgoftaar, T.M. A survey on image data augmentation for deep learning. J. Big Data; 2019; 6, 60. [DOI: https://dx.doi.org/10.1186/s40537-019-0197-0]
120. He, H.; Garcia, E.A. Learning from imbalanced data. IEEE Trans. Knowl. Data Eng.; 2009; 21, pp. 1263-1284.
121. Okada, Y. Surface deformation due to shear and tensile faults in a half-space. Bull. Seismol. Soc. Am.; 1985; 75, pp. 1135-1154. [DOI: https://dx.doi.org/10.1785/BSSA0750041135]
122. Lohman, R.B.; Simons, M. Some thoughts on the use of InSAR data to constrain models of surface deformation: Noise structure and data downsampling. Geochem. Geophys. Geosystems; 2005; 6, Q01007. [DOI: https://dx.doi.org/10.1029/2004GC000841]
123. Sudhaus, H.; Sigurjón, J. Improved source modelling through combined use of InSAR and GPS under consideration of correlated data errors: Application to the June 2000 Kleifarvatn earthquake, Iceland. Geophys. J. Int.; 2009; 176, pp. 389-404. [DOI: https://dx.doi.org/10.1111/j.1365-246X.2008.03989.x]
124. Sun, J.; Wauthier, C.; Stephens, K.; Gervais, M.; Cervone, G.; La Femina, P.; Higgins, M. Automatic detection of volcanic surface deformation using deep learning. J. Geophys. Res. Solid Earth; 2020; 125, e2020JB019840. [DOI: https://dx.doi.org/10.1029/2020JB019840]
125. Anantrasirichai, N.; Albino, F.; Hill, P.; Bull, D.; Biggs, J. Detecting volcano deformation in InSAR using deep learning. arXiv; 2018; arXiv: 1803.00380
126. Bountos, N.I.; Michail, D.; Papoutsis, I. Learning from synthetic InSAR with vision transformers: The case of volcanic unrest detection. IEEE Trans. Geosci. Remote Sens.; 2022; 60, 4509712. [DOI: https://dx.doi.org/10.1109/TGRS.2022.3180891]
127. Albino, F.; Biggs, J.; Yu, C.; Li, Z. Automated methods for detecting volcanic deformation using Sentinel-1 InSAR time series illustrated by the 2017–2018 unrest at Agung, Indonesia. J. Geophys. Res. Solid Earth; 2020; 125, e2019JB017908. [DOI: https://dx.doi.org/10.1029/2019JB017908]
128. Gaddes, M.; Hooper, A.; Bagnardi, M. Using machine learning to automatically detect volcanic unrest in a time series of interferograms. J. Geophys. Res. Solid Earth; 2019; 124, pp. 12304-12322. [DOI: https://dx.doi.org/10.1029/2019JB017519]
129. Jiang, L.; Shao, Y.; Zhou, Z.; Ma, P.; Wang, T. A review of intelligent InSAR data processing: Recent advancements, challenges and prospects. Acta Geod. Cartogr. Sin.; 2024; 53, pp. 1037-1056.
130. Vinyals, O.; Blundell, C.; Lillicrap, T.; Wierstra, D. Matching networks for one shot learning. Proceedings of the 30th Annual Conference on Neural Information Processing Systems (NIPS 2016); Barcelona, Spain, 5–10 December 2016; Volume 29.
131. Gilpin, L.H.; Bau, D.; Yuan, B.Z.; Bajwa, A.; Specter, M.; Kagal, L. Explaining explanations: An overview of interpretability of machine learning. Proceedings of the 2018 IEEE 5th International Conference on Data Science and Advanced Analytics (DSAA); Turin, Italy, 1–3 October 2018; pp. 80-89.
132. Carvalho, D.V.; Pereira, E.M.; Cardoso, J.S. Machine learning interpretability: A survey on methods and metrics. Electronics; 2019; 8, 832. [DOI: https://dx.doi.org/10.3390/electronics8080832]
133. Linardatos, P.; Papastefanopoulos, V.; Kotsiantis, S. Explainable AI: A review of machine learning interpretability methods. Entropy; 2020; 23, 18. [DOI: https://dx.doi.org/10.3390/e23010018] [PubMed: https://www.ncbi.nlm.nih.gov/pubmed/33375658]
134. Rudin, C.; Chen, C.; Chen, Z.; Huang, H.; Semenova, L.; Zhong, C. Interpretable machine learning: Fundamental principles and 10 grand challenges. Stat. Surv.; 2022; 16, pp. 1-85. [DOI: https://dx.doi.org/10.1214/21-SS133]
135. Thuerey, N.; Holl, P.; Mueller, M.; Schnell, P.; Trost, F.; Um, K. Physics-based deep learning. arXiv; 2021; arXiv: 2109.05237
136. Shen, H.; Jiang, M.; Li, J.; Zhou, C.; Yuan, Q.; Zhang, L. Coupling model-and data-driven methods for remote sensing image restoration and fusion: Improving physical interpretability. IEEE Geosci. Remote Sens. Mag.; 2022; 10, pp. 231-249. [DOI: https://dx.doi.org/10.1109/MGRS.2021.3135954]
137. Shen, H.; Zhang, L. Mechanism-learning coupling paradigms for parameter inversion and simulation in earth surface systems. Sci. China Earth Sci.; 2023; 66, pp. 568-582. [DOI: https://dx.doi.org/10.1007/s11430-022-9999-9]
© 2025 by the authors. Licensee MDPI, Basel, Switzerland. This article is an open access article distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Details
; Gong, Wenyu 2
; Zhang, Guohong 4
1 State Key Laboratory of Earthquake Dynamics and Forecasting, China Earthquake Administration, Beijing 100029, China;
2 State Key Laboratory of Earthquake Dynamics and Forecasting, China Earthquake Administration, Beijing 100029, China;
3 College of Oceanography and Space Informatics, China University of Petroleum (East China), Qingdao 266580, China;
4 State Key Laboratory of Earthquake Dynamics and Forecasting, China Earthquake Administration, Beijing 100029, China;