Next Article in Journal
The Crucial Role of Data Quality Control in Hydrochemical Studies: Reevaluating Groundwater Evolution in the Jiangsu Coastal Plain, China
Previous Article in Journal
Specific Activity of Radionuclides in Cryoconite Sediments of Glaciers of the Central Caucasus (Tsey, Skazka, Bezengi), Russia
 
 
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Land Subsidence Susceptibility Modelling in Attica, Greece: A Machine Learning Approach Using InSAR and Geospatial Data

by
Vishnuvardhan Reddy Yaragunda
1,*,
Divya Sekhar Vaka
2 and
Emmanouil Oikonomou
1
1
Department of Surveying & Geoinformatics Engineering, University of West Attica, Egaleo Park Campus, St Spyridonos, 12243 Athens, Greece
2
Geofem, 1080 Nicosia, Cyprus
*
Author to whom correspondence should be addressed.
Earth 2025, 6(3), 61; https://doi.org/10.3390/earth6030061
Submission received: 11 May 2025 / Revised: 11 June 2025 / Accepted: 16 June 2025 / Published: 21 June 2025

Abstract

Land subsidence significantly threatens urban infrastructure, agricultural productivity, and environmental sustainability. This study develops a land subsidence susceptibility model by integrating Small Baseline Subset (SBAS) Interferometric Synthetic Aperture Radar (InSAR) data with key geospatial factors using machine learning approaches. The study focuses on the Attica prefecture, Greece, and utilizes SBAS InSAR data from 2015 to 2021 to extract ground deformation velocities by classifying them into four susceptibility levels: stable, low, moderate, and high. The susceptibility results indicate that stable zones constitute 58.2% of the study area, followed by low (27.2%), moderate (11.2%), and high susceptibility zones (3.4%), predominantly concentrated in areas undergoing hydrological stress and urbanization. Random Forest (RF) and XGBoost (XGB) models incorporate a comprehensive set of causal factors, including slope, aspect, land use, groundwater level, geology, and rainfall. The evaluation of the models includes accuracy metrics and confusion matrices. The XGB model achieved the highest performance, recording an accuracy of 94%, with well-balanced predictions across all susceptibility classes. Addressing class imbalance during model training improved the recall of minority classes, though with slight trade-offs in precision. Feature importance analysis identifies proximity to streams, land use, aspect, rainfall, and groundwater extraction as the most influential factors driving subsidence susceptibility. This methodology demonstrates high reliability and robustness in predicting land subsidence susceptibility, providing critical insights for land-use planning and mitigation strategies. These findings establish a scalable framework for regional and global applications, contributing to sustainable land management and risk reduction efforts.

1. Introduction

Land subsidence is a geohazard that occurs when the land surface lowers over time due to natural or anthropogenic causes [1,2,3]. Subsidence due to natural processes occurs gradually as sediment compacts under overlying layers, carbonate rocks dissolve, and tectonic activity or isostatic adjustments contribute to ground deformation [4,5,6]. A range of human-made activities can lead to or worsen subsidence processes, including new settlements and underground excavations (mining and tunnelling) that can cause sudden or progressive ground collapse [7,8], as well as the compaction of susceptible aquifer systems by severe fluid withdrawal (groundwater or gas) [9,10,11]. In particular, urban areas exhibit high vulnerability to the effects of subsidence due to a high concentration of important projects and structures (e.g., bridges, highways, underground life-line pipes, subway), which increases exposure to hazardous conditions. In most cases, land subsidence is a slow-moving and gradual process that can be detected or revealed by instrumental methods; sometimes, the subsidence velocity can reach up to decimetres per year (e.g., eastern Rome, Italy) [12]. These subsiding movements can damage roads and pipelines, fracture service lines, cause differential settlement of buildings, and lead to seawater intrusion and flooding in coastal and fluvial environments [13]. There are significant limitations to conventional monitoring methods like levelling surveys, stratigraphic measurements, and GNSS (Global Navigation Satellite Systems) techniques [14,15]. These methods are characterized by time-consuming and labour-intensive processes, and they depend on fixed monitoring points, having limited capability to detect subtle deformations over large areas. Therefore, they are insufficient to address the evolving needs of land subsidence monitoring in dynamic urban environments, where subsidence patterns continuously shift in response to diverse human activities [16]. The advancements in remote sensing and geoinformation technology present a promising and complementary solution for land subsidence monitoring. Interferometric Synthetic Aperture Radar (InSAR) offers significant advantages, including high accuracy, all-weather operability, and extensive applicability for effectively monitoring land subsidence [17].
Time-series InSAR techniques can mitigate the impacts of temporal and spatial decorrelation, as well as atmospheric phase delay [18,19]. Over the past two decades, two primary InSAR methodologies—namely, Persistent Scatterer (PS) and Small Baseline Subset (SBAS)—have been developed to measure surface displacement caused by both human activities and natural processes. These methods enhance displacement accuracy by systematically removing atmospheric, topographic, and signal noise effects [20,21]. The Persistent Scatterer Interferometric Synthetic Aperture Radar (PS-InSAR) technique is particularly effective in highly coherent regions, such as urban areas with extensive man-made structures [22,23,24]. However, generating a sufficient number of PS points in complex terrains, such as mountainous or densely forested regions, can pose challenges and compromise the accuracy of monitoring outcomes [9,19,25]. An alternative and robust technique is the SBAS InSAR approach, which constructs differential interferometric pairs by applying carefully defined spatial and temporal baseline thresholds [26,27]. Deformation results are then derived using methods such as least squares estimation or singular value decomposition, ensuring reliable monitoring outcomes [28].
The assessment of land subsidence susceptibility (LSS) involves two approaches, namely knowledge-driven and data-driven approaches [29,30]. The former, for example the Analytic Hierarchy Process, is generally considered more effective than the latter [31]. Data-driven approaches include frequency ratio [32], weight of evidence [33], and evidential belief function [5], and they offer a mathematical framework to objectively summarize the relationships between influencing factors and land subsidence. However, their applicability remains limited due to their reliance on linear assumptions and sensitivity to subjective factor weighting, which can lead to biases in susceptibility mapping [34]. Machine learning (ML) methods have been successfully applied to geological hazard land subsidence susceptibility prediction [35,36,37], with the most common being the support vector machine (SVM), logistic regression, and decision trees [38,39]. Particularly, random forest (RF) is recognized for achieving ideal LSM performance [40,41,42,43]. ML has demonstrated significant feasibility for predicting LSS across various regions worldwide, utilizing different algorithms for specific geohazard conditions. For example, in China SVM and artificial neural networks were used to assess land subsidence in rapidly urbanizing regions, such as Shanghai and Guangzhou, where groundwater extraction and infrastructure development have led to ground deformation [44,45]. In Iran, RF and boosted regression trees were applied to map subsidence susceptibility linked to aquifer depletion [33]. Among ML models, RF and XGB classifiers proved to be increasingly accurate in highlighting geohazard susceptibility using traditional meteorological, geological, and remote sensing data, compared to other ML classifiers [46].
Advanced machine learning methods [47,48] utilize land subsidence incident impact factors as input features to uncover underlying mathematical relationships that facilitate accurate predictions. ML techniques are adopted in this study due to their advantage in handling complex, non-linear relationships among multiple contributing factors, without requiring prior assumptions about data distribution. This makes them particularly suitable for integrating with SBAS-InSAR-derived displacement data and various geospatial factors to enhance the reliability of susceptibility mapping. However, many approaches overlook the spatiotemporal variability of land subsidence. Leveraging SBAS-InSAR technology can be suitable for diverse geographical settings, such as Attica, Greece’s capital region, encompassing urban regions, rangelands, vegetation, and varied terrains [49]. By implementing the capabilities of SBAS-InSAR data over extended time series, it is possible to define more accurate susceptibility zones that capture land subsidence’s spatiotemporal continuity.
This study highlights the efficacy of combining SBAS-InSAR and ML methodologies to produce reliable susceptibility maps across diverse landscapes, including urban areas and regions with complex terrain morphology. Despite advancements in geohazard assessment, Attica lacks a comprehensive LSS model that integrates long-term InSAR deformation data with geospatial factors. Previous studies in Greece primarily focused on localized ground deformation monitoring [50,51], but a region-wide susceptibility assessment remains limited. Given the increasing urbanization, groundwater extraction, and geological complexity of Attica, a systematic susceptibility model is essential for identifying high-risk areas and mitigating potential damage. This study aims to address the above gap by developing an ML-based LSS Model that incorporates SBAS-InSAR-derived deformation rates between the period of 2015–2021, alongside key geospatial factors such as slope, aspect, curvature, plan and profile curvature, flow direction, distance to stream, road, faults, geology, groundwater level, Normalized Difference Vegetation Index (NDVI), and rainfall. By using both RF and Extreme Gradient Boosting (XGB), the proposed framework aims at increasing the accuracy and interpretability of subsidence susceptibility predictions. The results demonstrate the potential of the proposed framework for improving geohazard management strategies and promoting sustainable development.

2. Materials and Methods

2.1. Study Area

This study focuses on the Attica prefecture and part of the Corinth region, as shown in Figure 1. Attica is located in southern Greece, covers approximately 3808 km2, and features a diverse landscape that includes coastal plains, rugged mountains, and densely urbanized areas, particularly in the Athens metropolitan region, Greece’s capital. The area holds significant importance, as it is home to nearly half of Greece’s population and serves as the country’s cultural, economic, and industrial hub. The Corinth region, on the other hand, is known for its complex geology and tectonic activity, making it equally relevant for studying geohazards like land subsidence. The topography of Attica is highly variable, with elevations ranging from less than 100 m near the coastline to over 1400 m in the mountainous areas. Situated within the back-arc domain of the Hellenic Arc, the region is characterized by a mixture of metamorphic rocks, Alpine basement units, and Quaternary deposits [52]. These geological features are further complicated by the presence of active faults, such as the Avlon-Malakasa and Parnitha faults, which significantly influence the region’s tectonic behaviour. Similarly, the Corinth region shares many geological and tectonic characteristics, with prominent rift zones and seismic activity contributing to the complex dynamics of land subsidence [53].
The geology of the study area primarily consists of limestone, dolomite, flysch, and schist formations, along with post-Alpine deposits, such as marls and alluvium (Figure 1). These geological units are particularly significant in understanding land subsidence due to their varying degrees of consolidation and erosion potential [52]. Additionally, the region’s hydrological system, including streams and groundwater reservoirs, interacts with the underlying lithology and further influences subsidence processes in vulnerable areas.
These formations contribute significantly to the varying susceptibility to land subsidence across the region. The eastern and central parts of the region, including around Penteli, Glyfada, and the Saronikos Gulf, are dominated by limestones and marbles. In contrast, clastics and conglomerates are found predominantly in the Corinth and Megara areas (Western Attica), contributing to relatively unstable conditions in terms of mechanical integrity, particularly where these loose sedimentary formations are prone to compaction and water retention. Similarly, flysch deposits in localized parts of western Attica provide a mixed composition of sandstones, shales, and clay, often leading to differential stability due to their heterogeneous nature. Ophiolites, schists, and other metamorphic rocks occur sporadically throughout the region, especially along mountainous terrains. These formations exhibit variable permeability and mechanical behaviour, depending on their fracturing and weathering. Active faults, such as the Penteli, Megara 3, Saronic Gulf, Kalamaki, Akrokorinthos, and Voula Hill faults, are prominently featured in the geological map shown in Figure 1.
Given its susceptibility to land subsidence driven by both natural factors (tectonic activity and groundwater extraction) and human interventions (urban expansion, mining activities, and infrastructure development), ref. [51] both Attica and Corinth present an ideal setting for this study.

2.2. Data Preparation

Land subsidence data for the selected areas were obtained from Sentinel-1 Single Look Complex (SLC) C-band images. A total of 198 images within the period from December 2015 to December 2021 were acquired on the ascending track utilized for SBAS-InSAR analysis. The dataset for this study consists of SLC products acquired in Interferometric Wide Swath (IW) mode, providing coverage of approximately 250 km of surface area with a spatial resolution of 5 × 20 m. The SLC data contain both amplitude and phase information, which are essential for detailed terrain features and the monitoring of various environmental phenomena [56,57].
These images were sourced on 4 May 2023 from the Alaska Satellite Facility (ASF) (https://search.asf.alaska.edu accessed on 10 May 2025). The Sentinel-1 C-band images used in this study provide medium-resolution radar data with a wavelength of approximately 5.6 cm. The C-band is largely unaffected by atmospheric conditions, allowing penetration of clouds and rain, thus ensuring consistent data acquisition regardless of weather conditions. In addition, the Shuttle Radar Topography Mission (SRTM) 1 Arc-Second Digital Elevation Model (DEM), with a 30 m spatial resolution, was obtained from the USGS Earth Explorer (EarthExplorer) on 4 May 2023 for SBAS-InSAR processing. The processing was carried out using the ENVI SARscape software v5.6.2.
The factors used in this model include slope, aspect, curvature, plan and profile curvature, flow direction, distance to stream, distance to roads, NDVI, geology, groundwater levels, land use, precipitation, and distance to faults (Figure 2). These factors were selected based on the literature and their influence on land subsidence susceptibility, capturing physical, environmental, and anthropogenic conditions [45,58,59].
The first seven GIS layers—slope, aspect, curvature, plan curvature, profile curvature, flow direction, and distance to streams—were generated from a high spatial resolution (10 m) DEM using ArcGIS v10.5. Groundwater level data were obtained from historical datasets provided by the Institute of Geological and Mineral Exploration (IGME), Greece, covering the analysis period (2015–2021). Precipitation data were sourced from the Hellenic National Meteorological Service (HNMS) and spatially interpolated using spline interpolation in ArcGIS. The highest recorded rainfall value was 475 mm, and the lowest was 19.2 mm. The total precipitation data from each station were averaged from annual precipitation data to assess the overall precipitation pattern within the study area from 2016 to 2021. The geological data layer was derived from digitized geological maps originally developed by IGME, representing various lithological units across Attica.
The Normalized Difference Vegetation Index (NDVI) was calculated from Sentinel-2 images using Google Earth Engine, selecting images with less than 5% cloud coverage that align temporally with the InSAR dataset from 2015 to 2021. Land use data were obtained from the ESRI Sentinel-2 land use/land cover dataset with a 10 m spatial resolution. Additionally, distance-based parameters (distance to streams, distance to roads, and distance to faults) were computed using the Euclidean distance tool from ArcGIS Spatial Analyst. Stream networks were initially generated from DEM-based flow accumulation analysis, where areas with flow accumulation values under 2000 were identified as low accumulation zones, and those exceeding 2000 were categorized as high accumulation zones. These high accumulation areas were converted into a binary raster, from which stream order was subsequently derived. The resulting stream network was then transformed into vector data. Road networks and fault lines vectors were obtained from the National Observatory of Athens (NOA). The classification thresholds for continuous numerical variables were derived using quantile-based discretization to ensure a balanced distribution of data across classes, as shown in Table 1.

3. Methodology

3.1. SBAS-InSAR Technology

The SBAS approach forms multiple differential interferograms between images with small temporal and spatial baselines. Interferometric pairs were generated by selecting image combinations that minimize decorrelation effects, ensuring optimal coherence throughout the dataset. A total of 809 interferometric pairs were generated for SBAS analysis. A short spatiotemporal baseline subset combination can suppress the influence of decorrelation caused by random errors, unlike D-InSAR and PS-InSAR, making SBAS-InSAR more repeatable, scalable, and precise. The SBAS-InSAR processing workflow began with the co-registration of Sentinel-1 SLC images to a selected reference scene to ensure geometric alignment across the stack. Following this, differential interferograms were generated, and the topographic phase contribution was removed using the DEM. To improve phase quality and minimize noise, adaptive filtering based on coherence was applied. Coherence maps were subsequently generated to evaluate the quality of each interferogram and identify stable scatterers suitable for long-term deformation monitoring. Phase unwrapping was then performed using the Delaunay Minimum Cost Flow (MCF) algorithm, which provided strong results by minimizing ambiguity errors, especially in areas with variable coherence. To enhance the reliability of the displacement estimates, manual refinement (polynomial refinement) was performed using more than 49 ground control points (GCPs) distributed across the AOI. These GCPs were selected over stable, high-coherence urban features, such as buildings and roads, and were used to constrain phase unwrapping and improve baseline estimation.
The time series analysis involved a two-step inversion. In the first step, preliminary displacement velocities and residual DEM errors were estimated from the interferometric stack. This step helped isolate deformation trends from elevation-related residuals. In the second step, the displacement time series was refined by removing atmospheric phase delays. This was achieved through iterative filtering using a combination of temporal high-pass and spatial low-pass filters. Threshold parameters for atmospheric correction were fine-tuned based on multiple processing trials to obtain stable and realistic deformation patterns. Finally, the refined Line of Sight (LOS) deformation velocities and time series were geocoded into the WGS84 coordinate system, allowing spatial analysis and interpretation across the Attica region. This SBAS-InSAR output was used as the primary input for classifying LSS in the ML models. The complete methodology used for LSS mapping is presented in Figure 3.

3.2. InSAR Velocity Classification Criteria

The classification criteria used in our analysis are founded on established scientific methodologies that link ground movement velocity to land subsidence risk. Recognizing the dynamic nature of land subsidence across different geological and environmental conditions, we divided the study area into four susceptibility classes, based on velocity measurements derived from SBAS-InSAR (Table 2).
The classification ranges from Class 0, representing stable areas with minimal or negligible deformation, to Class 3, which highlights regions experiencing significant subsidence and high susceptibility to structural or environmental impacts. Our adaptation of velocity ranges is based on the literature [43,60,61] for categorizing SBAS-InSAR velocities, and we opted for a four-class scheme as it is more suitable for our study’s context.
The comprehensive and methodological approach ensures that the development of the LSS map is guided not only by empirical data but also by a deep understanding of the ground deformation characteristics and their impact on structural and environmental stability.
The performed classification was based on multiple previous studies that used a velocity threshold to assess subsidence susceptibility; for instance, others utilized the Jenks Natural Breaks method to classify SBAS-InSAR velocities into five susceptibility classes [61].
Similarly, others provided additional classification schemes, further supporting our selection of threshold values. For example, previous works classified velocities into very low, low, medium, high, and very high susceptibility ranges, with a cutoff of −20 mm/y for extreme subsidence [44], while others established thresholds at −6 mm/year, −8 mm/year, and −12 mm/year for progressive susceptibility levels [45]. These thresholds provide an optimal balance between statistical accuracy and field applicability, ensuring that areas prone to subsidence are clearly delineated. The classification criteria from the relevant literature are presented in Table 3, in comparison with the current study’s thresholds.
After establishing classification criteria based on ground movement velocity, the methodology further incorporated an in-depth analysis of environmental, geological, and hydrological factors, as outlined in Section 2.2. For each location identified through SBAS-InSAR data, these factors were systematically examined to assess their contribution to the assigned susceptibility class. This integrated approach facilitates a comprehensive study of correlations, aiming to identify key predictors of land subsidence susceptibility across the study area. By combining velocity thresholds with causal factors, the analysis enhances the reliability of susceptibility predictions and improves the understanding of subsidence dynamics in vulnerable regions.

3.3. Ensemble Learning Models

Ensemble learning is a robust ML approach that has shown significant advantages in various applications, including geospatial and remote sensing studies. An ensemble combines multiple individual models to generate a single predictive output. This technique enhances generalization capability and improves model performance compared to using a single model alone [62]. For the present work’s LSS model, we applied two ensemble learning algorithms, namely (RF) and (XGB), both of which are known for their accuracy and reliability in handling complex datasets.

3.3.1. Random Forest (RF)

RF is a supervised learning algorithm that operates through an ensemble of decision trees. Each tree is trained on a different bootstrap sample of the data, with random subsets of features selected at each split. This randomness reduces correlation between the individual trees, which improves the generalization performance of the model. The final prediction is made by aggregating outputs from all trees, typically using majority voting for classification tasks or averaging for regression tasks [63]. This ensemble structure mitigates overfitting and enhances the model’s robustness against noise and outliers, making it suitable for large, heterogeneous datasets. RF is also relatively insensitive to hyperparameter tuning, often yielding high accuracy with minimal adjustments to parameters [64]. Due to these characteristics, RF has been successfully applied to InSAR-based deformation studies, like land subsidence and landslide studies [33,43,45], where complex interactions between multiple environmental and geospatial factors must be modelled.

3.3.2. Extreme Gradient Boosting (XGB)

XGB is another powerful supervised learning algorithm that builds on the principles of gradient boosting. Like RF, XGB can handle both classification and regression tasks. Boosting involves training multiple weak learners sequentially, with each learner correcting the errors of the previous ones. XGB stands out for its scalability, efficiency, and ability to handle large datasets by running boosted trees in parallel [65]. A key advantage of XGB is its ability to optimize memory usage and hardware resources, enabling faster execution and better performance. The algorithm also incorporates L1 (Lasso) and L2 (Ridge) regularization techniques, which help to minimize overfitting by penalizing complex models [66]. This makes XGB an ideal choice for subsidence susceptibility modelling, where precise classification is critical to identify areas at high risk of ground deformation.

3.3.3. Model Parameter Tuning

Hyperparameters in ML algorithms play a crucial role in optimizing model performance and need careful adjustment before training [67]. In this study, both the RF and XGB models were fine-tuned to improve their ability to classify land subsidence susceptibility. We considered using Grid Search to perform hyperparameter tuning, evaluating various configurations using the F1 macro score as the scoring metric. This process was conducted on the training data after applying oversampling techniques, namely SMOTE and Adaptive Synthetic Sampling (ADASYN), to address class imbalance. These methods are widely used to enhance model performance by generating synthetic samples for underrepresented classes. SMOTE achieves this by interpolating between existing minority samples, while ADASYN adapts the sampling process based on the density distribution of the minority class, creating more synthetic samples in regions with higher class imbalance. This helps improve the models’ ability to predict minority classes accurately without discarding valuable data.
For the random forest model, the key parameters, such as the number of decision trees (n_estimators), tree depth (max_depth), minimum samples for node splitting (min_samples_split), and minimum samples required at leaf nodes (min_samples_leaf), were adjusted. The grid search tested different combinations of these parameters to find an optimal balance between model complexity and generalization. RF’s ensemble nature and built-in feature selection mechanisms contribute to its robustness in handling large, heterogeneous datasets, which is essential in subsidence studies.
Similarly, hyperparameter tuning for the XGB model focused on enhancing both scalability and accuracy. Parameters such as the number of trees, tree depth, learning rate, and the proportion of features sampled for each tree were optimized. The learning rate controlled the step size in updating the model, while tree depth regulated the complexity of the individual trees. The model was configured with use_label_encoder = False, which disables XGBoost’s legacy label encoding system to ensure compatibility with scikit-learn’s preprocessing methods for categorical data. Additionally, eval_metric = ‘mlogloss’ was used to evaluate classification performance using multi-class logarithmic loss, a metric that effectively quantifies prediction errors in subsidence susceptibility mapping involving multiple classes.
A Principal Component Analysis (PCA) step was incorporated to reduce feature dimensionality and mitigate multicollinearity. After aggregating the pre-processed batches, PCA was applied, and components explaining 95% of the variance were retained. This step helped streamline the dataset, improving both computational efficiency and the interpretability of the models. The PCA components were later mapped back to the original features, which facilitated a detailed analysis of feature importance during the evaluation phase.

3.3.4. Model Training

We followed a systematic approach to model training and development, beginning with data partitioning. The dataset was split into 70% for training and 30% for testing to evaluate model performance consistently.
Once the data were resampled, RF and XGB models were trained on the original class distribution, as well as on the oversampled datasets. The data were divided into batches for SMOTE and ADASYN resampling to ensure efficient handling of large datasets. The models were evaluated on the testing set after being trained on different combinations of features derived from PCA-reduced and frequency-encoded steps. The training process was designed to balance computational efficiency and predictive performance. This allowed the models to focus on key features while reducing noise and redundancy. Additionally, frequency encoding was used to transform categorical variables, improving model interpretability and scalability.

3.3.5. Model Validation

The model validation was conducted using multiple metrics to evaluate the predictive performance of the RF and XGB models. Model evaluation was performed using receiver operating characteristic (ROC) curves, confusion matrices, and multiple performance metrics, including accuracy, precision, recall, and F1 score. The ROC curve serves as a key indicator of the model’s ability to discriminate between classes by plotting the true positive rate against the false positive rate across various thresholds. Area Under the Curve (AUC) was computed for each class, with values approaching 1 indicating superior classification performance and discrimination.
In addition to the ROC analysis, confusion matrices were used to quantify classification outcomes by breaking down predictions into true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). These results were critical for calculating key evaluation metrics. Accuracy represents the overall percentage of correct predictions, while precision measures the proportion of correctly identified positive instances out of all predicted positives. Recall (or sensitivity) evaluates the model’s ability to detect all relevant positive instances, and the F1 score gives the mean of precision and recall, providing a balanced metric, especially useful for datasets with class imbalance.

4. Results and Discussion

4.1. SBAS-InSAR

In this study, SBAS-InSAR analysis was implemented to assess ground deformation across the Attica region by examining six selected distributed scatterer points (A, B, C, D, E, and F). These points exhibit varying displacement rates, influenced by geological conditions, tectonic activity, and land use. The time series displacement data from November 2015 to December 2021, together with the LOS velocity, provided important insights into the rate and progression of land subsidence over the study period.
To better understand the site-specific conditions influencing deformation, Table 4 summarizes the causal factor values and their classification categories for the six representative SBAS points (A to F) selected for detailed analysis.
Point A is located near Spata (Eastern Attica), within a rangeland landcover area, and lies near Athens International Airport. It demonstrates a significant LOS velocity of −16.82 mm/year and a cumulative displacement of over −109 mm during the observation period (~6 years). The Pyrina Hill fault, situated within this rangeland zone, lies approximately 4 km from the airport to the northeast (Figure 4), underscoring the close spatial relationship between tectonic structures and observed ground deformation. While the airport itself represents a built-up environment, the surrounding landscape on both sides consists predominantly of rangeland, within which Point A is situated. The observed localized subsidence at this location may result from a combination of soil compaction, fault-related tectonic activity, and possibly anthropogenic influences, such as groundwater extraction [56].
Point B is located in the Penteli surrounding area (just to the North of point A), characterized by a geotectonically complex environment influenced by multiple active faults, including the Penteli, Drafi, Voula Hill, and Pyrina Hill faults (Figure 4). The LOS velocity recorded at this point is −4.14 mm/year, with a cumulative displacement of approximately −23 mm (Figure 5) over the monitoring period. The time series shows relatively stable ground movement with minor seasonal fluctuations, possibly linked to varying groundwater recharge and extraction cycles. Penteli and the surrounding areas have experienced localized ground movements in the past, likely associated with co-seismic deformation resulting from regional earthquakes and fault activity. According to Papoutsis et al. [68], extensive studies using multi-stack persistent scatterer interferometry revealed that nearby areas, such as Kifisia and Kryoneri, which are 7 and 10 km away from Penteli, experienced significant subsidence, with <−6 mm/yr during the 1992–1999 period due to excessive groundwater extraction. Once water extraction was reduced, these areas underwent a physical restoration phase, resulting in an uplift in subsequent years. While specific data for Penteli in that period are limited, it is reasonable to infer that similar mechanisms may influence current deformation patterns in the area. The close proximity of multiple faults also adds tectonic stress, potentially exacerbating ground movement.
The observed subsidence at Point B is likely a result of the combined influences of tectonic stress from local active faults like Drafi and Voula and anthropogenic activities, particularly groundwater extraction. The findings from [68] provide important context for understanding the patterns of deformation observed in the Kifisia and Kryoneri areas located near Penteli, supporting our interpretation of SBAS results for this area. The classification results further enabled the identification of areas prone to significant subsidence due to excessive groundwater exploitation. Notable regions within the Attica prefecture, including those with high agricultural and industrial water demand, exhibited a high potential for extraction-related subsidence.
Point C is located in Kallithea, a densely urbanized part of southern Athens, exhibiting an LOS velocity of −5.95 mm/year with a cumulative subsidence of nearly −47mm (Figure 5). The displacement trend is largely consistent, but shows minor seasonal variations, likely influenced by shifts in urban water management practices and infrastructure activities. Ground subsidence in urban environments often results from a combination of building loads, groundwater fluctuations, and soil consolidation. While these local conditions are significant, broader environmental factors, such as slope, precipitation, and geology, also play a role, as indicated by our susceptibility model’s multi-factor approach.
Point D is located in the seismically active Megara region, near the Gulf of Corinth (southwestern Attica), which is characterized by high seismicity and rapid extensional tectonics. The SBAS analysis at this point reveals a significant LOS velocity of −13.13 mm/year, with a cumulative displacement of approximately −89 mm (Figure 5). The time series indicates a predominantly linear trend, with some fluctuations. This notable subsidence aligns with the tectonic framework outlined in studies of the Megara basin, indicating that both natural and anthropogenic factors play a role. Other studies confirm that the region experiences fault reactivation and block rotations, contributing to long-term ground motion [69]. These tectonic dynamics affect ground deformation, including both uplift and subsidence in fault-bounded regions.
These findings offer a robust context to interpret the subsidence rate observed in our SBAS movements. While the data period differs, the consistent patterns of deformation validate that the ground motion detected in Megara is part of a long-term geodynamic process.
Point E, located in the Corinthos region, exhibits an LOS velocity of −10.48 mm/year and a cumulative displacement of −63 mm (Figure 5). The displacement trend is steady, indicating persistent tectonic subsidence influenced by the region’s dynamic extensional tectonics and fault structures, including the Kalamaki-Isthmia and Akrokorinthos faults. Tectonically, the Corinthos area is characterized by a dynamic extensional regime associated with the Gulf of Corinth, one of the most active rift zones in Europe. According to previous studies, the Kalamaki-Isthmia fault in the eastern part of the canal exhibits low but persistent slip rates of around 0.05 mm/year; this fault, along with others in the region, offsets Pleistocene sediments and contributes to significant vertical ground deformation [70]. Further work highlighted that differential vertical motions and fault-controlled deformation patterns, such as subsidence on the southern coast near Kalamaki and uplift toward Loutraki, are prevalent across the region [71]. The cumulative subsidence observed in the Corinthos area could also be influenced by anthropogenic activities. Intensive irrigation practices, common in the surrounding agricultural areas, may cause soil compaction, exacerbating natural tectonic subsidence [72,73]. Additionally, studies of sedimentary environments near the fault zones indicate alternating phases of fluvial, marine, and lagoonal deposits, which are impacted by both tectonic and sea-level changes. The combination of fault-driven subsidence and anthropogenic influences helps explain the significant displacement detected in the SBAS analysis.
Point F, situated near Velo to the west of the Corinth Canal (Figure 4), exhibits a relatively low LOS velocity of −3.03 mm/year, with a cumulative displacement of approximately −20 mm. The time series analysis reveals a gradual, linear subsidence trend, suggesting ongoing ground deformation likely influenced by a combination of local and regional factors.
While land use practices—particularly agricultural activity and irrigation—may contribute to soil compaction and groundwater extraction-related subsidence, tectonic influences are also evident. Upon consulting the Earthquake Catalog of Athens University, two local seismic events (with magnitudes between 2.0 and 2.5 between 2016 and 2021) were identified in the vicinity of Point F, with an additional six events occurring in the broader Gulf of Corinth region. Although such low-magnitude earthquakes may not produce significant displacement detectable in InSAR time series, their occurrence affirms the presence of active tectonic processes and ongoing crustal adjustments, thus further supporting the role of nearby faults and regional stress regimes in contributing to the observed ground motion.
Our model highlights that, to fully understand the spatial variability in LSS, it is essential to integrate not only tectonic and anthropogenic influences, but also topographic, environmental, and hydrological factors; these elements collectively shape the deformation patterns in the area, and their interplay must be considered to ensure accurate geohazard assessment and mitigation planning.

4.2. Model Accuracy Validation

The predictive performance of the developed LSS model was evaluated using standard classification metrics, including accuracy, precision, recall, F1-score, and the ROC and AUC curve. The evaluation was conducted across different model configurations, incorporating both RF and XGB classifiers under three data processing conditions: original class distribution, SMOTE, and ADASYN. These configurations were compared to determine the most effective model for accurately classifying LSS while minimizing misclassification errors. The confusion metrics (Figure 6) provide a detailed breakdown of classification performance across susceptibility levels. The RF model with original data exhibited the highest classification accuracy for the stable class (99%). The high susceptibility class was classified at 88%, representing the lowest accuracy among the four susceptibility categories. However, the moderate and low susceptibility categories demonstrated minor misclassifications, with 12% of moderate cases being assigned to the high susceptibility class. The RF with SMOTE improved recall, achieving 93% classification accuracy in the high susceptibility category, while also reducing false negatives in moderate areas. Similarly, the RF model with ADASYN further enhanced recall, correctly identifying 96% of high-susceptibility cases. However, a slight reduction in precision was observed, suggesting an increased likelihood of false positives in moderate susceptibility regions.
The XGB model exhibited similar trends but with distinct classification behaviours. The XGB model with original data demonstrated a high accuracy of 97%, particularly in the stable class, but with lower moderate (77%) and high (77%) susceptibility categories. This indicates that while the model confidently predicted stable regions, some subsidence-prone areas were not optimally classified. The SMOTE-enhanced XGB model showed an improvement in recall, increasing classification accuracy to 83% for moderate and 91% for high susceptibility zones, respectively. Notably, the ADASYN-enhanced XGB model provided the most balanced classification across all susceptibility levels, achieving 94% accuracy for high susceptibility, while maintaining precision comparable to the SMOTE configuration.
A comparison of overall model performance reveals key distinctions. The RF model with original data achieved the highest classification accuracy (0.96) (Figure 7) but displayed moderate levels of misclassification in subsidence-prone areas. The RF model with SMOTE effectively reduced false negatives, improving recall for high susceptibility regions. The ADASYN configuration further increased sensitivity, particularly in the high susceptibility class, though at the cost of slightly lower precision. In contrast, the XGB model demonstrated a different trade-off. The original XGB classifier exhibited strong precision but lower recall, indicating that while it minimized false positives, it occasionally failed to identify high-risk areas. The SMOTE-enhanced XGB model improved recall and reduced false negatives, while the ADASYN variant balanced both recall and precision, making it the most robust configuration for land subsidence susceptibility classification.
Further validation of model performance is provided by the ROC curve analysis. The AUC values across all configurations ranged between 0.96 and 1.00, confirming strong classification capabilities (Figure 8). The XGB model with ADASYN exhibited the highest AUC, reinforcing its superior predictive capability. While RF with original data displayed near-perfect classification for stable areas, it exhibited lower recall in moderate and high susceptibility zones. The incorporation of SMOTE and ADASYN significantly improved recall, leading to better detection of high-risk regions.
The overall findings confirm that ML-based susceptibility mapping effectively characterizes LSS, with the model choice influencing the classification balance. While RF with original data achieved the highest overall accuracy, its relatively lower recall suggests that it may not be optimal for identifying high-risk areas. The XGB model with ADASYN, despite a minor reduction in precision, demonstrated superior recall and AUC values, making it the most balanced configuration. Given the trade-off between precision and recall, the XGB model with ADASYN is considered the most suitable for LSS classification in this study. These results emphasize the importance of oversampling techniques in enhancing model sensitivity, particularly for capturing high-susceptibility zones.

4.3. Land Subsidence Susceptibility (LSS) Mapping

LSS mapping serves as an essential tool to identify regions prone to ground deformation, classifying them by different levels of susceptibility. LSS maps play a crucial role in urban planning, infrastructure development, and environmental risk mitigation, providing valuable information for government agencies, urban planners, and geotechnical experts. By identifying high-susceptibility zones, decision-makers can implement preventive strategies, such as groundwater extraction regulation, geotechnical stabilization, and land-use planning, to mitigate potential risks. Additionally, these maps support disaster response efforts, allowing for timely intervention, infrastructure resilience planning, and long-term environmental management strategies.
The feature importance analysis from both the RF and XGB models under different data configurations provides a detailed understanding of the dominant features influencing land subsidence (Figure 9). Among the most influential factors, distance to streams, aspect, and land use emerged as the primary contributors, collectively accounting for approximately 45–55% of the total feature importance across all models. Aspect alone exhibited the highest significance, ranging between 15% and 20%, indicating its crucial role in terrain stability and water drainage patterns, which influence soil saturation and compaction processes. Land use contributed between 12% and 18%, highlighting the impact of urbanization, deforestation, and agricultural activities on ground stability. Distance to streams accounted for 10–16%, aligning with previous studies that emphasize the erosional and hydrological effects of proximity to water bodies on subsidence-prone areas.
Hydrogeological parameters, particularly rainfall and groundwater levels, also played a significant role, contributing approximately 12–18% to the model’s predictions. Rainfall alone accounted for 7–9%, reinforcing its role in soil moisture variation and potential consolidation of soft sediments, while groundwater levels contributed around 6–9%, corroborating the well-documented link between groundwater extraction and subsidence. Geological and infrastructural factors, including distance to roads and geology, showed moderate importance, contributing approximately 8–12% collectively. Distance to roads ranged between 4% and 7%, indicating the impact of construction-induced stress and compaction. Geology contributed around 4–6%, reflecting the variability in subsurface lithology and its susceptibility to settlement. The remaining factors, such as distance to faults, curvature, slope, and plan curvature, exhibited relatively lower importance, contributing less than 8% in total across all models. While these factors influence localized subsidence events, their overall contribution to large-scale subsidence patterns appears secondary compared to hydrological and anthropogenic drivers.
The consistency in feature importance rankings across the RF and XGB models, regardless of whether original, SMOTE, or ADASYN resampling techniques were applied, underscores the reliability of the identified factors. The comparable weighting of aspect, land use, and distance to streams across different ML configurations validates the robustness of the susceptibility model and highlights the dominant drivers controlling subsidence in Attica.
The results of the present study align with and extend previous findings on the drivers of land subsidence in Attica. Earlier research has highlighted the role of groundwater depletion and hydrogeological imbalance in urban and agricultural areas [74,75]. Our model confirms this at a regional scale, where groundwater level change, rainfall, and proximity to streams collectively contribute a substantial portion of the total feature importance, indicating their central role in controlling subsidence susceptibility. The observed spatial distribution of high and moderate susceptibility zones in our model, particularly in areas with Quaternary alluvial deposits and intensive land use, supports earlier geotechnical studies that documented ground deformation associated with excessive groundwater extraction and compressible soil layers [50,51].
The influence of tectonic structures, previously identified through fault mapping and seismic hazard assessments [76,77,78,79], is also reflected in our results, although with lower relative importance. While the “distance to fault” factor did not rank among the top predictors in the model, regions located near active faults still demonstrated elevated susceptibility, suggesting that tectonic activity may act as a secondary control, potentially facilitating subsurface water movement and associated soil compaction.
Additionally, topographic parameters such as slope and aspect—traditionally associated with slope stability—also emerged as influential in certain regions. North-facing and gently sloped terrains exhibited higher susceptibility, likely due to increased soil moisture retention and reduced drainage. This finding indicates a partial overlap in the conditioning factors of landslides and subsidence, particularly in geomorphologically complex areas of Attica.
By integrating SBAS-InSAR-derived velocities with geospatial predictors through ensemble learning, this study provides a comprehensive and scalable assessment of land subsidence susceptibility. The results offer new insights into spatial patterns of deformation and establish a robust foundation for informed land-use planning and groundwater resource management.
The present regional-scale susceptibility analysis, integrating critical parameters, such as groundwater extraction, geological settings, hydrological conditions (rainfall and stream proximity), and anthropogenic influences, quantitatively verifies and expands upon other localized studies. Notably, while previous research primarily focused on specific locations within Attica, such as Eleonas, Athens, and northeastern Attica, our analysis covers a significantly broader geographic extent, including the entirety of the Attica prefecture and extending westward toward the Corinth region. Our findings demonstrate that hydrological (e.g., groundwater extraction, rainfall, proximity to streams) and anthropogenic (land-use and urban development) factors emerged consistently as dominant contributors to subsidence susceptibility across the wider region. Distance to faults, despite being highlighted in previous localized tectonic studies, did not rank among the top influencing factors in our model. This can likely be attributed to regional variations in fault distribution and activity levels. Particularly, faults in southern Attica are spatially sparse and generally distant from densely populated urban areas compared to those in northern and central Attica. Consequently, their overall influence on regional-scale subsidence susceptibility appears reduced when integrated into a broader spatial analysis.
The LSS map generated classifies the study area into four distinct susceptibility levels—stable, low, moderate, and high—based on the integration of geotechnical, hydrological, and anthropogenic factors (Figure 10). This classification provided crucial insights into the spatial distribution of subsidence-prone regions, supporting urban planning, infrastructure development, and risk mitigation strategies. The susceptibility analysis revealed that stable zones, where subsidence risk is negligible, constituted the majority of the study area, encompassing 5,691,477 grid points (58.2%). Here, grid points refer to the individual raster cells, each representing a 10 m x 10 m spatial resolution, covering the entire study area. Each cell was classified into one of the susceptibility categories. Low susceptibility areas, where mild subsidence processes are likely to occur, accounted for 2,662,627 grid points (27.2%). Moderate susceptibility zones, indicative of notable subsidence potential, were identified in 1,095,278 grid points (11.2%). The high susceptibility class, representing regions experiencing the most severe subsidence activity, mapped over 334,273 grid points (3.4%), predominantly concentrated in areas undergoing hydrogeological stress, urbanization, and soft soil deposits.
The spatial distribution of high and moderate susceptibility zones aligns with the previous studies mentioned above, highlighting subsidence-prone regions in Attica. The Megara region, located west of Athens, exhibits moderate to high susceptibility primarily in the areas surrounding its urban core, where soft Quaternary sediments and sedimentary formations are present. These findings align with previous studies identifying subsidence-prone zones in western Attica due to soil compressibility, fault reactivation, block rotation, and hydrogeological characteristics [69]. In the Fyli region (northwestern Attica), localized zones of moderate susceptibility are observed, with limited high susceptibility areas. While the exact cause of subsidence remains uncertain, potential contributors include geomorphological characteristics and terrain instability. Unlike regions with documented groundwater over-extraction, Fyli’s susceptibility may be influenced by its proximity to elevated terrain and erosional processes rather than direct anthropogenic factors.
The areas near Kifisia (northern Attica) exhibit high susceptibility along the regions adjacent to streams, likely due to erosion and soil instability caused by hydrological processes. The study by [52] highlights that landslides and soil degradation in northern Attica, including Kifisia, are often linked to slope failures triggered by stream undercutting and erosion along steep terrains. This mechanism suggests that groundwater flow and surface runoff contribute to gradual terrain instability, exacerbating subsidence risks in these areas. Similarly, in the Penteli region (northeastern Attica), moderate to high susceptibility is noted, particularly in areas with weathered metamorphic formations and fractured lithologies, making them vulnerable to differential compaction and ground deformation. Within Athens, scattered zones of moderate to high susceptibility are observed because of rapid urbanization. Susceptibility in these areas is likely driven by a combination of land use transitions, groundwater usage fluctuations, and soil conditions. Notably, the area surrounding Athens International Airport, predominantly classified as rangeland, demonstrated extensive high susceptibility. Rangelands, which are open natural landscapes primarily used for grazing, are often characterized by loose, unconsolidated soils that are susceptible to compaction and subsidence when subjected to environmental and anthropogenic pressures.
In southern and eastern Attica, mixed moderate and high susceptibility zones are detected along the Saronikos Gulf and Glyfada, regions with varied topography and limited urbanization. These areas, located along coastal and elevated terrain, suggest a possible correlation between subsidence susceptibility and topographic factors, such as aspect and proximity to streams. Steeper slopes and specific aspect orientations can influence water drainage patterns, increasing soil saturation and accelerating subsidence processes in certain conditions. Additionally, regions adjacent to stream networks may experience fluvial erosion and sediment instability, contributing to localized subsidence. The coastal belt and regions extending from Megara to Corinth (west of Athens), where high susceptibility is predominant, align with prior geotechnical and hydrogeological studies highlighting tectonic-induced deformation and groundwater extraction for irrigation as key subsidence drivers in the region.

5. Conclusions

This study developed an ML-based LSS Model for Attica, Greece, integrating SBAS-InSAR-derived ground deformation data with geospatial factors using RF and XGB classifiers. The results demonstrated that SBAS-InSAR techniques provide reliable subsidence velocity measurements, allowing for high-resolution susceptibility mapping at a regional scale. By classifying land subsidence into four susceptibility classes (stable/negligible susceptibility, low, moderate, and high), the model effectively delineated areas prone to subsidence-related hazards.
The feature importance analysis identified proximity to streams, aspect, and land use as influential factors, collectively accounting for 45–55% of the model’s predictive capacity. Hydrogeological parameters, including rainfall and groundwater levels, played a secondary but significant role, reinforcing the well-documented link between groundwater depletion, surface compaction, and land subsidence. The classification results indicated that 27.2% of the study area falls under low susceptibility, while 3.4% is categorized as highly susceptible, primarily located in zones experiencing hydrogeological stress, the presence of compressible sediments, and recent changes in land use.
In conclusion, the proposed ML-based approach offers a scalable and efficient framework for mapping land subsidence, enhancing urban resilience, and supporting sustainable land-use planning and infrastructure management. Future research should incorporate multi-temporal hydrological datasets and continuous monitoring to improve model precision and address the evolving nature of subsidence-related hazards.
Despite the model’s demonstrated effectiveness, certain aspects remain open for further enhancement. For instance, while the current framework reliably utilizes EO and geospatial data, incorporating in situ ground truth measurements such as GNSS or levelling benchmarks in future studies could offer added validation of the observed subsidence patterns. Additionally, integrating dynamic factors such as real-time groundwater fluctuations, seismic activity, and high-frequency precipitation data could improve the model’s temporal responsiveness. Regularly updating the model with the latest Earth Observation data and newly identified subsidence zones will be crucial for maintaining its accuracy and effectiveness over time. Furthermore, extending the approach to additional regions will support its transferability and broader application in land subsidence monitoring.

Author Contributions

Conceptualization, V.R.Y. and E.O.; data curation, V.R.Y.; formal analysis, V.R.Y.; methodology, V.R.Y. and E.O.; resources, E.O.; software, V.R.Y.; supervision, D.S.V. and E.O.; validation, V.R.Y.; visualization, V.R.Y., D.S.V. and E.O.; writing- original draft, V.R.Y.; Writing—review and editing, V.R.Y., D.S.V. and E.O. All authors have read and agreed to the published version of the manuscript.

Funding

This work is supported by the Special Accounts for Research Funds (ELKE), University of West Attica, Athens, Greece, under grant number 80781.

Data Availability Statement

The data are available upon request from the corresponding author.

Conflicts of Interest

The authors declare no conflicts of interest. Author Divya Sekhar Vaka was employed by the company Geofem. The remaining authors declare that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

References

  1. Galloway, D.L.; Burbey, T.J. Review: Regional Land Subsidence Accompanying Groundwater Extraction. Hydrogeol. J. 2011, 19, 1459–1486. [Google Scholar] [CrossRef]
  2. Holzer, T.L.; Galloway, D.L. Impacts of Land Subsidence Caused by Withdrawal of Underground Fluids in the United States. In Humans as Geologic Agents; Ehlen, J., Haneberg, W.C., Larson, R.A., Eds.; Geological Society of America: Boulder, CO, USA, 2005; Volume 16, pp. 87–99. ISBN 978-0-8137-4116-1. [Google Scholar]
  3. Tomás, R.; Romero, R.; Mulas, J.; Marturià, J.J.; Mallorquí, J.J.; Lopez-Sanchez, J.M.; Herrera, G.; Gutiérrez, F.; González, P.J.; Fernández, J.; et al. Radar Interferometry Techniques for the Study of Ground Subsidence Phenomena: A Review of Practical Issues through Cases in Spain. Environ. Earth Sci. 2014, 71, 163–181. [Google Scholar] [CrossRef]
  4. Bianchini, S.; Moretti, S. Analysis of Recent Ground Subsidence in the Sibari Plain (Italy) by Means of Satellite SAR Interferometry-Based Methods. Int. J. Remote Sens. 2015, 36, 4550–4569. [Google Scholar] [CrossRef]
  5. Pradhan, B.; Abokharima, M.H.; Jebur, M.N.; Tehrany, M.S. Land Subsidence Susceptibility Mapping at Kinta Valley (Malaysia) Using the Evidential Belief Function Model in GIS. Nat. Hazards 2014, 73, 1019–1042. [Google Scholar] [CrossRef]
  6. Solari, L.; Ciampalini, A.; Raspini, F.; Bianchini, S.; Moretti, S. PSInSAR Analysis in the Pisa Urban Area (Italy): A Case Study of Subsidence Related to Stratigraphical Factors and Urbanization. Remote Sens. 2016, 8, 120. [Google Scholar] [CrossRef]
  7. Lee, S.; Park, I. Application of Decision Tree Model for the Ground Subsidence Hazard Mapping near Abandoned Underground Coal Mines. J. Environ. Manag. 2013, 127, 166–176. [Google Scholar] [CrossRef]
  8. Ng, A.H.-M.; Ge, L.; Li, X.; Abidin, H.Z.; Andreas, H.; Zhang, K. Mapping Land Subsidence in Jakarta, Indonesia Using Persistent Scatterer Interferometry (PSI) Technique with ALOS PALSAR. Int. J. Appl. Earth Obs. Geoinf. 2012, 18, 232–242. [Google Scholar] [CrossRef]
  9. Ittycheria, N.; Vaka, D.S.; Rao, Y.S. Time Series Analysis of Surface Deformation of Bengaluru City Using Sentinel-1 Images. ISPRS Ann. Photogramm. Remote Sens. Spat. Inf. Sci. 2018, 4, 473–477. [Google Scholar] [CrossRef]
  10. Notti, D.; Mateos, R.M.; Monserrat, O.; Devanthéry, N.; Peinado, T.; Roldán, F.J.; Fernández-Chacón, F.; Galve, J.P.; Lamas, F.; Azañón, J.M. Lithological Control of Land Subsidence Induced by Groundwater Withdrawal in New Urban AREAS (Granada Basin, SE Spain). Multiband DInSAR Monitoring. Hydrol. Process. 2016, 30, 2317–2331. [Google Scholar] [CrossRef]
  11. Pulido-Bosch, A.; Delgado, J.; Sola, F.; Vallejos, Á.; Vicente, F.; López-Sánchez, J.M.; Mallorquí, J.J. Identification of Potential Subsidence Related to Pumping in the Almería Basin (SE Spain). Hydrol. Process. 2012, 26, 731–740. [Google Scholar] [CrossRef]
  12. Bozzano, F.; Esposito, C.; Franchi, S.; Mazzanti, P.; Perissin, D.; Rocca, A.; Romano, E. Analysis of a Subsidence Process by Integrating Geological and Hydrogeological Modelling with Satellite InSAR Data. In Engineering Geology for Society and Territory—Volume 5; Lollino, G., Manconi, A., Guzzetti, F., Culshaw, M., Bobrowsky, P., Luino, F., Eds.; Springer International Publishing: Cham, Switzerland, 2015; pp. 155–159. [Google Scholar]
  13. Julio-Miranda, P.; Ortíz-Rodríguez, A.J.; Palacio-Aponte, A.G.; López-Doncel, R.; Barboza-Gudiño, R. Damage Assessment Associated with Land Subsidence in the San Luis Potosi-Soledad de Graciano Sanchez Metropolitan Area, Mexico, Elements for Risk Management. Nat. Hazards 2012, 64, 751–765. [Google Scholar] [CrossRef]
  14. Abidin, H.Z.; Andreas, H.; Gumilar, I.; Sidiq, T.P.; Fukuda, Y. Land Subsidence in Coastal City of Semarang (Indonesia): Characteristics, Impacts and Causes. Geomat. Nat. Hazards Risk 2013, 4, 226–240. [Google Scholar] [CrossRef]
  15. Motagh, M.; Djamour, Y.; Walter, T.R.; Wetzel, H.-U.; Zschau, J.; Arabi, S. Land Subsidence in Mashhad Valley, Northeast Iran: Results from InSAR, Levelling and GPS. Geophys. J. Int. 2007, 168, 518–526. [Google Scholar] [CrossRef]
  16. Cian, F.; Blasco, J.M.D.; Carrera, L. Sentinel-1 for Monitoring Land Subsidence of Coastal Cities in Africa Using PSInSAR: A Methodology Based on the Integration of SNAP and StaMPS. Geosciences 2019, 9, 124. [Google Scholar] [CrossRef]
  17. Hu, B.; Chen, J.; Zhang, X. Monitoring the Land Subsidence Area in a Coastal Urban Area with InSAR and GNSS. Sensors 2019, 19, 3181. [Google Scholar] [CrossRef] [PubMed]
  18. Kouhartsiouk, D.; Perdikou, S. The Application of DInSAR and Bayesian Statistics for the Assessment of Landslide Susceptibility. Nat. Hazards 2021, 105, 2957–2985. [Google Scholar] [CrossRef]
  19. Vaka, D.S.; Sharma, S.; Rao, Y. Comparison of HH and VV Polarizations for Deformation Estimation Using Persistent Scatterer Interferometry. In Proceedings of the 38th Asian Conference on Remote Sensing—Space Applications: Touching Human Lives, ACRS, New Delhi, India, 23–27 October 2017. [Google Scholar]
  20. Chen, X.; Achilli, V.; Fabris, M.; Menin, A.; Monego, M.; Tessari, G.; Floris, M. Combining Sentinel-1 Interferometry and Ground-Based Geomatics Techniques for Monitoring Buildings Affected by Mass Movements. Remote Sens. 2021, 13, 452. [Google Scholar] [CrossRef]
  21. Yaragunda, V.R.; Oikonomou, E. Monitoring the Los (Line of Sight) Deformation of the Corinth Canal, Greece Using SBAS and GNSS Techniques. In Proceedings of the IGARSS 2024—2024 IEEE International Geoscience and Remote Sensing Symposium, Athens, Greece, 7–12 July 2024; pp. 11056–11061. [Google Scholar]
  22. Ferretti, A.; Prati, C.; Rocca, F. Permanent Scatterers in SAR Interferometry. IEEE Trans. Geosci. Remote Sens. 2001, 39, 8–20. [Google Scholar] [CrossRef]
  23. Hooper, A.; Zebker, H.; Segall, P.; Kampes, B. A New Method for Measuring Deformation on Volcanoes and Other Natural Terrains Using InSAR Persistent Scatterers. Geophys. Res. Lett. 2004, 31, L23611. [Google Scholar] [CrossRef]
  24. Ramirez, R.A.; Abdullah, R.E.E.; Rubio, C.J.P. S1-PSInSAR Monitoring and Hyperbolic Modeling of Nonlinear Ground Subsidence in Naga City, Cebu Island in The Philippines. GEOMATE J. 2022, 23, 102–109. [Google Scholar] [CrossRef]
  25. Chen, G.; Zhang, Y.; Zeng, R.; Yang, Z.; Chen, X.; Zhao, F.; Meng, X. Detection of Land Subsidence Associated with Land Creation and Rapid Urbanization in the Chinese Loess Plateau Using Time Series InSAR: A Case Study of Lanzhou New District. Remote Sens. 2018, 10, 270. [Google Scholar] [CrossRef]
  26. Devaraj, S.; Yarrakula, K.; Martha, T.R.; Murugesan, G.P.; Vaka, D.S.; Surampudi, S.; Wadhwa, A.; Loganathan, P.; Budamala, V. Time Series SAR Interferometry Approach for Landslide Identification in Mountainous Areas of Western Ghats, India. J. Earth Syst. Sci. 2022, 131, 133. [Google Scholar] [CrossRef]
  27. Vaka, D.S.; Rao, Y.S.; Bhattacharya, A. Time Series Analysis of the Pre-Seismic and Post-Seismic Surface Deformation of the 2017 Iran–Iraq Earthquake Derived from Sentinel-1 InSAR Data. J. Earth Syst. Sci. 2023, 132, 64. [Google Scholar] [CrossRef]
  28. Berardino, P.; Fornaro, G.; Lanari, R.; Sansosti, E. A New Algorithm for Surface Deformation Monitoring Based on Small Baseline Differential SAR Interferograms. IEEE Trans. Geosci. Remote Sens. 2002, 40, 2375–2383. [Google Scholar] [CrossRef]
  29. Arabameri, A.; Saha, S.; Roy, J.; Tiefenbacher, J.P.; Cerda, A.; Biggs, T.; Pradhan, B.; Thi Ngo, P.T.; Collins, A.L. A Novel Ensemble Computational Intelligence Approach for the Spatial Prediction of Land Subsidence Susceptibility. Sci. Total Environ. 2020, 726, 138595. [Google Scholar] [CrossRef]
  30. Jin, B.; Yin, K.; Li, Q.; Gui, L.; Yang, T.; Zhao, B.; Guo, B.; Zeng, T.; Ma, Z. Susceptibility Analysis of Land Subsidence along the Transmission Line in the Salt Lake Area Based on Remote Sensing Interpretation. Remote Sens. 2022, 14, 3229. [Google Scholar] [CrossRef]
  31. Ghorbanzadeh, O.; Feizizadeh, B.; Blaschke, T. An Interval Matrix Method Used to Optimize the Decision Matrix in AHP Technique for Land Subsidence Susceptibility Mapping. Environ. Earth Sci. 2018, 77, 584. [Google Scholar] [CrossRef]
  32. Ranjgar, B.; Razavi-Termeh, S.V.; Foroughnia, F.; Sadeghi-Niaraki, A.; Perissin, D. Land Subsidence Susceptibility Mapping Using Persistent Scatterer SAR Interferometry Technique and Optimized Hybrid Machine Learning Algorithms. Remote Sens. 2021, 13, 1326. [Google Scholar] [CrossRef]
  33. Mohammady, M.; Pourghasemi, H.R.; Amiri, M. Land Subsidence Susceptibility Assessment Using Random Forest Machine Learning Algorithm. Environ. Earth Sci. 2019, 78, 503. [Google Scholar] [CrossRef]
  34. Oh, H.-J.; Syifa, M.; Lee, C.-W.; Lee, S. Land Subsidence Susceptibility Mapping Using Bayesian, Functional, and Meta-Ensemble Machine Learning Models. Appl. Sci. 2019, 9, 1248. [Google Scholar] [CrossRef]
  35. Bui, D.T.; Tsangaratos, P.; Nguyen, V.-T.; Liem, N.V.; Trinh, P.T. Comparing the Prediction Performance of a Deep Learning Neural Network Model with Conventional Machine Learning Models in Landslide Susceptibility Assessment. CATENA 2020, 188, 104426. [Google Scholar] [CrossRef]
  36. Cao, J.; Zhang, Z.; Du, J.; Zhang, L.; Song, Y.; Sun, G. Multi-Geohazards Susceptibility Mapping Based on Machine Learning—A Case Study in Jiuzhaigou, China. Nat. Hazards 2020, 102, 851–871. [Google Scholar] [CrossRef]
  37. Daviran, M.; Shamekhi, M.; Ghezelbash, R.; Maghsoudi, A. Landslide Susceptibility Prediction Using Artificial Neural Networks, SVMs and Random Forest: Hyperparameters Tuning by Genetic Optimization Algorithm. Int. J. Environ. Sci. Technol. 2023, 20, 259–276. [Google Scholar] [CrossRef]
  38. Huang, F.; Yan, J.; Fan, X.; Yao, C.; Huang, J.; Chen, W.; Hong, H. Uncertainty Pattern in Landslide Susceptibility Prediction Modelling: Effects of Different Landslide Boundaries and Spatial Shape Expressions. Geosci. Front. 2022, 13, 101317. [Google Scholar] [CrossRef]
  39. Kadavi, P.R.; Lee, C.-W.; Lee, S. Landslide-Susceptibility Mapping in Gangwon-Do, South Korea, Using Logistic Regression and Decision Tree Models. Environ. Earth Sci. 2019, 78, 116. [Google Scholar] [CrossRef]
  40. Arabameri, A.; Saha, S.; Roy, J.; Chen, W.; Blaschke, T.; Tien Bui, D. Landslide Susceptibility Evaluation and Management Using Different Machine Learning Methods in The Gallicash River Watershed, Iran. Remote Sens. 2020, 12, 475. [Google Scholar] [CrossRef]
  41. Chen, W.; Xie, X.; Wang, J.; Pradhan, B.; Hong, H.; Bui, D.T.; Duan, Z.; Ma, J. A Comparative Study of Logistic Model Tree, Random Forest, and Classification and Regression Tree Models for Spatial Prediction of Landslide Susceptibility. CATENA 2017, 151, 147–160. [Google Scholar] [CrossRef]
  42. Deng, H.; Wu, X.; Zhang, W.; Liu, Y.; Li, W.; Li, X.; Zhou, P.; Zhuo, W. Slope-Unit Scale Landslide Susceptibility Mapping Based on the Random Forest Model in Deep Valley Areas. Remote Sens. 2022, 14, 4245. [Google Scholar] [CrossRef]
  43. Vaka, D.S.; Yaragunda, V.R.; Perdikou, S.; Papanicolaou, A. InSAR Integrated Machine Learning Approach for Landslide Susceptibility Mapping in California. Remote Sens. 2024, 16, 3574. [Google Scholar] [CrossRef]
  44. Chai, L.; Xie, X.; Wang, C.; Tang, G.; Song, Z. Ground Subsidence Risk Assessment Method Using PS-InSAR and LightGBM: A Case Study of Shanghai Metro Network. Int. J. Digit. Earth 2024, 17, 2297842. [Google Scholar] [CrossRef]
  45. Zhao, F.; Miao, F.; Wu, Y.; Xiong, Y.; Gong, S.; Sun, D. Land Subsidence Susceptibility Mapping in Urban Settlements Using Time-Series PS-InSAR and Random Forest Model. Gondwana Res. 2024, 125, 406–424. [Google Scholar] [CrossRef]
  46. Bammou, Y.; Benzougagh, B.; Ouallali, A.; Kader, S.; Raougua, M.; Igmoullan, B. Improving Landslide Susceptibility Mapping in Semi-Arid Regions Using Machine Learning and Geospatial Techniques. DYSONA Appl. Sci. 2025, 6, 269–290. [Google Scholar] [CrossRef]
  47. Mohammady, M.; Pourghasemi, H.R.; Amiri, M. Assessment of Land Subsidence Susceptibility in Semnan Plain (Iran): A Comparison of Support Vector Machine and Weights of Evidence Data Mining Algorithms. Nat. Hazards 2019, 99, 951–971. [Google Scholar] [CrossRef]
  48. Yu, H.; Arabameri, A.; Costache, R.; Crăciun, A.; Arora, A. Land Subsidence Susceptibility Assessment Using Advanced Artificial Intelligence Models. Geocarto Int. 2022, 37, 18067–18093. [Google Scholar] [CrossRef]
  49. Gounaridis, D.; Symeonakis, E.; Chorianopoulos, I.; Koukoulas, S. Incorporating Density in Spatiotemporal Land Use/Cover Change Patterns: The Case of Attica, Greece. Remote Sens. 2018, 10, 1034. [Google Scholar] [CrossRef]
  50. Kaitantzian, A.; Loupasakis, C. Preliminary Investigation of the Land Subsidence Phenomena Occurring at the Industrial-Commersial Area of Eleonas, Athens, Greece. Bull. Geol. Soc. Greece 2017, 50, 1703. [Google Scholar] [CrossRef]
  51. Parcharidis, I.; Lagios, E.; Sakkas, V.; Raucoules, D.; Feurer, D.; Mouélic, S.; King, C.; Carnec, C.; Novali, F.; Ferretti, A.; et al. Subsidence Monitoring within the Athens Basin (Greece) Using Space Radar Interferometric Techniques. Earth Planets Space 2006, 58, 505–513. [Google Scholar] [CrossRef]
  52. Tavoularis, N.; Papathanassiou, G.; Ganas, A.; Argyrakis, P. Development of the Landslide Susceptibility Map of Attica Region, Greece, Based on the Method of Rock Engineering System. Land 2021, 10, 148. [Google Scholar] [CrossRef]
  53. Zhou, Z.; Whittaker, A.C.; Bell, R.E.; Hampson, G.J. Unravelling Tectonic and Lithological Effects on Transient Landscapes in the Gulf of Corinth, Greece. Basin Res. 2024, 36, e12901. [Google Scholar] [CrossRef]
  54. Photiades, A.; Zananiri, I. Digital Geological Map of Greece in Scale 1:1000 000, Harmonized To One Geology Europe. In Proceedings of the 8th EUREGEO Conference—Geological 3D Modelling, Barcelona, Spain, 15–17 June 2015; pp. 117–118. Available online: http://www.researchgate.net/publication/311885407 (accessed on 20 June 2025).
  55. Ganas, A. NOAFAULTS KMZ Layer, Version 5.0; Zenodo: Athens, Greece, 2023. [Google Scholar]
  56. Ashraf, T.; Yin, F.; Liu, L.; Zhang, Q. Land Subsidence Detection Using SBAS- and Stacking-InSAR with Zonal Statistics and Topographic Correlations in Lakhra Coal Mines, Pakistan. Remote Sens. 2024, 16, 3815. [Google Scholar] [CrossRef]
  57. Guo, J.; Hu, J.; Li, B.; Zhou, L.; Wang, W. Land Subsidence in Tianjin for 2015 to 2016 Revealed by the Analysis of Sentinel-1A with SBAS-InSAR. J. Appl. Remote Sens. 2017, 11, 026024. [Google Scholar] [CrossRef]
  58. Hakim, W.L.; Achmad, A.R.; Lee, C.-W. Land Subsidence Susceptibility Mapping in Jakarta Using Functional and Meta-Ensemble Machine Learning Algorithm Based on Time-Series InSAR Data. Remote Sens. 2020, 12, 3627. [Google Scholar] [CrossRef]
  59. Zhao, R.; Arabameri, A.; Santosh, M. Land Subsidence Susceptibility Mapping: A New Approach to Improve Decision Stump Classification (DSC) Performance and Combine It with Four Machine Learning Algorithms. Environ. Sci. Pollut. Res. 2024, 31, 15443–15466. [Google Scholar] [CrossRef]
  60. Moretto, S.; Bozzano, F.; Mazzanti, P. The Role of Satellite InSAR for Landslide Forecasting: Limitations and Openings. Remote Sens. 2021, 13, 3735. [Google Scholar] [CrossRef]
  61. Yao, Z.; Chen, M.; Zhan, J.; Zhuang, J.; Sun, Y.; Yu, Q.; Yu, Z. Refined Landslide Susceptibility Mapping by Integrating the SHAP-CatBoost Model and InSAR Observations: A Case Study of Lishui, Southern China. Appl. Sci. 2023, 13, 12817. [Google Scholar] [CrossRef]
  62. Huang, F.; Xie, G.; Xiao, R. Research on Ensemble Learning. In Proceedings of the 2009 International Conference on Artificial Intelligence and Computational Intelligence, Shanghai, China, 7–8 November 2009; Volume 3, pp. 249–252. [Google Scholar]
  63. Belgiu, M.; Drăguţ, L. Random Forest in Remote Sensing: A Review of Applications and Future Directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  64. Nam, K.; Kim, J.; Chae, B.-G. Exploring Class Imbalance with Under-Sampling, over-Sampling, and Hybrid Sampling Based on Mahalanobis Distance for Landslide Susceptibility Assessment: A Case Study of the 2018 Iburi Earthquake Induced Landslides in Hokkaido, Japan. Geosci. J. 2024, 28, 71–94. [Google Scholar] [CrossRef]
  65. Zhang, Y.; Liang, S.; Zhu, Z.; Ma, H.; He, T. Soil Moisture Content Retrieval from Landsat 8 Data Using Ensemble Learning. ISPRS J. Photogramm. Remote Sens. 2022, 185, 32–47. [Google Scholar] [CrossRef]
  66. Ahmad, M.; Al-Mansob, R.A.; Kashyzadeh, K.R.; Keawsawasvong, S.; Sabri Sabri, M.M.; Jamil, I.; Alguno, A.C. Extreme Gradient Boosting Algorithm for Predicting Shear Strengths of Rockfill Materials. Complexity 2022, 2022, 9415863. [Google Scholar] [CrossRef]
  67. Sharma, N.; Saharia, M.; Ramana, G.V. High Resolution Landslide Susceptibility Mapping Using Ensemble Machine Learning and Geospatial Big Data. CATENA 2024, 235, 107653. [Google Scholar] [CrossRef]
  68. Papoutsis, I.; Charalabos, K.; Paradissis, D. Multi-Stack Persistent Scatterer Interferometry Analysis in Wider Athens, Greece. Remote Sens. 2017, 9, 276. [Google Scholar] [CrossRef]
  69. Mattei, M.; D’Agostino, N.; Zananiri, I.; Kondopoulou, D.; Pavlides, S.; Spatharas, V. Tectonic Evolution of Fault-Bounded Continental Blocks: Comparison of Paleomagnetic and GPS Data in the Corinth and Megara Basins (Greece). J. Geophys. Res. Solid. Earth 2004, 109, B02106. [Google Scholar] [CrossRef]
  70. Pallikarakis, A.; Papanikolaοu, I.; Triantaphyllou, M.; Grützner, C.; Migiros, G.; Reicherter, K.; Mason, J.; Schneiderwind, S. Magnetic Susceptibility Changes Correlated with Paleoenvironmental and Lithological Changes, Within Borehole Cores on Either Sides of The Kalamaki-Isthmia Fault in Corinth Canal. Bull. Geol. Soc. Greece 2017, 50, 104. [Google Scholar] [CrossRef]
  71. Mariolakos, I.; Stiros, S. Quaternary Deformation of the Isthmus and Gulf of Corinthos (Greece). Geology 1987, 15, 225–228. [Google Scholar] [CrossRef]
  72. Daskalaki, P.; Voudouris, K. Groundwater Quality of Porous Aquifers in Greece: A Synoptic Review. Environ. Geol. 2008, 54, 505–513. [Google Scholar] [CrossRef]
  73. Kallioras, A.; Pliakas, F.-K.; Marinos, P. Groundwater Resources Assessment of Granular Aquifers in Greece. Bull. Geol. Soc. Greece 2017, 50, 701. [Google Scholar] [CrossRef]
  74. Bathrellos, G.; Skilodimou, H.; Kelepertsis, A.; Alexakis, D.; Chrisanthaki, I.; Archonti, D. Environmental Research of Groundwater in the Urban and Suburban Areas of Attica Region, Greece. Environ. Geol. 2008, 56, 11–18. [Google Scholar] [CrossRef]
  75. Kourgialas, N.N. A Critical Review of Water Resources in Greece: The Key Role of Agricultural Adaptation to Climate-Water Effects. Sci. Total Environ. 2021, 775, 145857. [Google Scholar] [CrossRef]
  76. Deligiannakis, G.; Papanikolaou, I.D.; Roberts, G. Fault Specific GIS Based Seismic Hazard Maps for the Attica Region, Greece. Geomorphology 2018, 306, 264–282. [Google Scholar] [CrossRef]
  77. Ganas, A.; Pavlides, S.; Karastathis, V. DEM-Based Morphometry of Range-Front Escarpments in Attica, Central Greece, and Its Relation to Fault Slip Rates. Geomorphology 2005, 65, 301–319. [Google Scholar] [CrossRef]
  78. Grützner, C.; Schneiderwind, S.; Papanikolaou, I.; Deligiannakis, G.; Pallikarakis, A.; Reicherter, K. New Constraints on Extensional Tectonics and Seismic Hazard in Northern Attica, Greece: The Case of the Milesi Fault. Geophys. J. Int. 2016, 204, 180–199. [Google Scholar] [CrossRef]
  79. Papanikolaou, D.; Papanikolaou, I. Geological, Geomorphological and Tectonic Structure of NE Attica and Seismic Hazard Implications for the Northern Edge of the Athens Plain. Bull. Geol. Soc. Greece 2007, 40, 425–438. [Google Scholar] [CrossRef]
Figure 1. Study area map of Attica and the Corinth region. The area of interest (AOI) was defined based on the Sentinel-1 frame coverage to ensure complete inclusion of SBAS-InSAR deformation data. As a result, the AOI was delineated by applying a direct cut to the shapefile, rather than using the full extent of the Attica region. The geological map of the study area shows various lithologies, such as schist, flysch, limestone, ophiolites, metamorphic rocks, and Tertiary and Quaternary deposits. The geological map was obtained from the Department of Surveying and Geoinformatics Engineering, University of West Attica (Greece), originally developed by the Institute of Geological and Mineral Exploration (IGME) of Greece [54]. Active faults featured in the map were collected from the National Observatory of Athens [55].
Figure 1. Study area map of Attica and the Corinth region. The area of interest (AOI) was defined based on the Sentinel-1 frame coverage to ensure complete inclusion of SBAS-InSAR deformation data. As a result, the AOI was delineated by applying a direct cut to the shapefile, rather than using the full extent of the Attica region. The geological map of the study area shows various lithologies, such as schist, flysch, limestone, ophiolites, metamorphic rocks, and Tertiary and Quaternary deposits. The geological map was obtained from the Department of Surveying and Geoinformatics Engineering, University of West Attica (Greece), originally developed by the Institute of Geological and Mineral Exploration (IGME) of Greece [54]. Active faults featured in the map were collected from the National Observatory of Athens [55].
Earth 06 00061 g001
Figure 2. Different land subsidence causative factors are used as input for the machine learning model. Topographic derived parameters include the following: (A) slope, (B) aspect, (C) curvature, (D) plan curvature, (E) profile curvature, (F) flow direction, and (G) distance from streams. Other causal factors include: (H) distance from roads, (I) vegetation, (J) precipitation, (K) distance from faults, (L) groundwater extraction, and (M) land use.
Figure 2. Different land subsidence causative factors are used as input for the machine learning model. Topographic derived parameters include the following: (A) slope, (B) aspect, (C) curvature, (D) plan curvature, (E) profile curvature, (F) flow direction, and (G) distance from streams. Other causal factors include: (H) distance from roads, (I) vegetation, (J) precipitation, (K) distance from faults, (L) groundwater extraction, and (M) land use.
Earth 06 00061 g002
Figure 3. Methodological flowchart of the proposed model.
Figure 3. Methodological flowchart of the proposed model.
Earth 06 00061 g003
Figure 4. Mean LOS velocity map of Attica prefecture. Selected distributed scatterer points (A, B, C, D, E, and F) are highlighted in black boxes for detailed assessment. Active faults highlighted in red were collected from the National Observatory of Athens [55].
Figure 4. Mean LOS velocity map of Attica prefecture. Selected distributed scatterer points (A, B, C, D, E, and F) are highlighted in black boxes for detailed assessment. Active faults highlighted in red were collected from the National Observatory of Athens [55].
Earth 06 00061 g004
Figure 5. SBAS-InSAR displacement time series trends for selected distributed scatterer points from A to F in Figure 4. (a) Point A is located near Spata, showing consistent subsidence with a velocity of −16.82 mm/year; (b) Point B is situated in Penteli with moderate deformation at −4.14 mm/year; (c) Point C is located in the urban area of Kallithea, showing velocity of −5.95 mm/year; (d) Point D is located in the Megara region with significant deformation at −13.13 mm/year; (e) Point E is located in the Corinth region with linear subsidence at −10.48 mm/yr; (f) Point F is located in Velo area showing low rate deformation with velocity −3.03 mm/year.
Figure 5. SBAS-InSAR displacement time series trends for selected distributed scatterer points from A to F in Figure 4. (a) Point A is located near Spata, showing consistent subsidence with a velocity of −16.82 mm/year; (b) Point B is situated in Penteli with moderate deformation at −4.14 mm/year; (c) Point C is located in the urban area of Kallithea, showing velocity of −5.95 mm/year; (d) Point D is located in the Megara region with significant deformation at −13.13 mm/year; (e) Point E is located in the Corinth region with linear subsidence at −10.48 mm/yr; (f) Point F is located in Velo area showing low rate deformation with velocity −3.03 mm/year.
Earth 06 00061 g005
Figure 6. Confusion metrics for all the models (RF and XGB) with original and oversampled data.
Figure 6. Confusion metrics for all the models (RF and XGB) with original and oversampled data.
Earth 06 00061 g006
Figure 7. Comparative metrics for RF and XGB models with original and oversampled data (SMOTE and ADASYN).
Figure 7. Comparative metrics for RF and XGB models with original and oversampled data (SMOTE and ADASYN).
Earth 06 00061 g007
Figure 8. ROC curve and AUC values for each class (from Class 1 to Class 4) using RF and XGB models with original and oversampled data (SMOTE and ADASYN).
Figure 8. ROC curve and AUC values for each class (from Class 1 to Class 4) using RF and XGB models with original and oversampled data (SMOTE and ADASYN).
Earth 06 00061 g008
Figure 9. Importance of factors for the used models (RF and XGB) with different data configurations (original data, SMOTE, and ADASYN).
Figure 9. Importance of factors for the used models (RF and XGB) with different data configurations (original data, SMOTE, and ADASYN).
Earth 06 00061 g009
Figure 10. Land subsidence susceptibility (LSS) model of Attica prefecture, Corinth region using XGB Adasyn model. The map classifies subsidence into four categories: Stable/Negligible susceptibility, Low, Moderate, and High susceptibility.
Figure 10. Land subsidence susceptibility (LSS) model of Attica prefecture, Corinth region using XGB Adasyn model. The map classifies subsidence into four categories: Stable/Negligible susceptibility, Low, Moderate, and High susceptibility.
Earth 06 00061 g010
Table 1. Geospatial factor classification.
Table 1. Geospatial factor classification.
FactorClassification CriteriaCategories
Slope≤2°Flat
2–10°Moderate
10–20°Steep
>20°Very Steep
Aspect−1Flat
0–22.5North
22.5–67.5Northeast
67.5–112.5East
112.5–157.5Southeast
157.5–202.5South
202.5–247.5Southwest
247.5–292.5West
292.5–337.5Northwest
337.5–360North
Curvature≤−20Concave
−20 to −5Slightly Concave
−5 to 5Flat
5 to 20Slightly Convex
>20Convex
Plan Curvature≤−15Strongly Concave
−15 to −5Moderately Concave
−5 to 5Flat
5 to 15Moderately Convex
>15Strongly Convex
Profile Curvature≤−15Strongly Concave
−15 to −5Moderately Concave
−5 to 5Flat
5 to 15Moderately Convex
>15Strongly Convex
Flow DirectionClassified using eight directional values 1, 2, 4, 8, 16, 32, 64, 128 (Directional flow values)
Distance to Streams (m)≤100Very close
100–500Close
500–1000Moderate
1000–2000Far
>2000Very far
Distance to Roads (m)≤250Very close
250–500Close
500–1000Moderate
1000–1500Far
>1500Very far
NDVI<0.0Bare/Sparse
0.0–0.2Low
0.2–0.4Moderate
0.4–0.6High
>0.6Very High
GeologyCategorical data with geological formations1. Limestones
2. Limestone and Marbles
3. Hornstones
4. Tertiary and Quaternary deposits
5. Limestones and Dolomites
6. Flysch
7. Marine Deposits
8. Ophiolites
9. Schists and Limestones
10. Clastic and conglomerates
11. Metamorphic Rocks
Groundwater Level Change≤50Minimal
50–100Low
100–150Moderate
150–200High
>200Very High
Land Use Categorical land use typesWater
Trees
Built-up
Croplands
Rangelands
Bare Ground
Rainfall (mm/y)≤100Low
100–200Moderate
200–300High
>300Very High
Distance to Faults (m)≤2000Very Close
2000–5000Close
5000–10000Moderate
10000–20000Far
>20000Very far
Table 2. Land subsidence susceptibility (LSS) classification criteria based on velocity and literature.
Table 2. Land subsidence susceptibility (LSS) classification criteria based on velocity and literature.
ClassificationSusceptibility Class
Velocity Interval: from 0 to 2 mm/yStable/Negligible susceptibility—No significant ground motion detectable at this scale
Velocity Interval: from −2 to −4 mm/yLow susceptibility
Velocity Interval: from −4 to −7 mm/yModerate susceptibility
Velocity Interval: <−7 mm/yHigh susceptibility
Table 3. The land subsidence/deformation classification criteria of this study are compared with the previous literature.
Table 3. The land subsidence/deformation classification criteria of this study are compared with the previous literature.
StudyVelocity ClassificationSusceptibility
Current Study0 to −2.0 mm/y Stable/Negligible—No significant ground motion detectable at this scale
−2.0 to −4.0 mm/yLow
−4.0 to −7.0 mm/yModerate
<−7.0 mm/yHigh
Yao et al. [61]>−2.46 mm/yVery low
−2.46 to −5.39Low
−5.39 to −9.14Moderate
−9.14 to −14.88High
<−14.88Very high
Chai et al. [44]>0 mm/yVery low
0 to −5 mm/ylow
−5 to −10 mm/yMedium
−10 to −20 mm/yHigh
<−20 mm/yVery high
Zhao et al. [45]−6 mm/yLow
−8 mm/yModerate
−10 mm/yHigh
−12 mm/yVery high
Vaka et al. [43]0 to −2 mm/yNo
2 to −5 mm/yLow
5 to −15 mm/yModerate
<−15 mm/yHigh
Table 4. Summary of the causal factor values and their corresponding classification categories at the six selected SBAS-InSAR points (A–F). These values reflect the local geospatial, geological, environmental, and anthropogenic conditions that may contribute to observed deformation patterns.
Table 4. Summary of the causal factor values and their corresponding classification categories at the six selected SBAS-InSAR points (A–F). These values reflect the local geospatial, geological, environmental, and anthropogenic conditions that may contribute to observed deformation patterns.
FactorsPoint A (−16.32 mm/yr)Point B (−4.14 mm/yr.)Point C (−5.95 mm/yr.)Point D (−13.13 mm/yr.)Point E (−10.48 mm/yr.)Point F (−3.03 mm/yr.)
Slope1.39
(≤2°—Flat)
17.71
(>20°—Very Steep)
0
(≤2°—Flat)
8.29
(2 to 10°—Moderate)
2.42
(2 to 10°—Moderate)
0.65
(≤2°—Flat)
Aspect115.7
(Southeast)
59.51
(Northeast)
−1
(Flat)
344.85
(North)
172.4
(South)
74.8
(East)
Curvature−0.9
(−5 to 5 Flat)
0.1
(−5 to 5 Flat)
0
(−5 to 5 Flat)
−0.44
(−5 to 5 Flat)
−0.19
(−5 to 5 Flat)
0.22
(−5 to 5 Flat)
Plan Curvature−0.72
(−5 to 5 Flat)
0.19
(−5 to 5 Flat)
0
(−5 to 5 Flat)
−0.12
(−5 to 5 Flat)
−0.09
(−5 to 5 Flat)
0.17
(−5 to 5 Flat)
Profile Curvature0.18
(−5 to 5 Flat)
0.08
(−5 to 5 Flat)
0
(−5 to 5 Flat)
0.32
(−5 to 5 Flat)
0.09
(−5 to 5 Flat)
−0.44
(−5 to 5 Flat)
Flow Direction1
(East)
128
(Northeast)
4
(South)
64
(North)
2
(Southeast)
32
(Northwest)
Dist. to Streams201.24
(100 to 500—Close)
92.19
(100 to 500—Close)
20
(100 to 500—Close)
60.82
(100 to 500—Close)
150
(100 to 500—Close)
142.1
(100 to 500—Close)
Dist. to Roads120
(≤250—Very close)
0
(≤250—Very close)
41.23
(≤250—Very close)
41.23
(≤250—Very close)
0
(≤250—Very close)
0
(≤250—Very close)
NDVI0.13
(Low)
0.2
(Low)
0.07
(Low)
0.09
(Low)
0.23
(Moderate)
0.05
(Low)
GeologyTertiary and Quaternary DepositsTertiary and Quaternary DepositsTertiary and Quaternary DepositsTertiary and Quaternary DepositsMarine DepositsTertiary and Quaternary Deposits
Groundwater Level Change15.35
(≤50—Minimal)
165.6
(150 to 200—High)
8.96
(≤50—Minimal)
67.75
(50 to 100—Low)
54.3
(50 to 100—Low)
29.04
(50 to 100—Low)
Land Use11 (Rangeland)7 (Built Area)7 (Built Area)11 (Rangeland)5 (Crops)7 (Buit Area)
Rainfall298.61
(200 to 300—High)
202.4
(200 to 300—High)
213.8
(200 to 300—High)
289.4
(200 to 300—High)
250.6
(200 to 300—High)
248
(200 to 300—High)
Distance to Faults5737.8
(5000 to 10,000—Moderate)
461.4
(≤2000—Very Close)
13761.8
(10,000 to 20,000—Far)
4226.8
(2000 to 5000—Close)
1058.4
(≤2000—Very Close)
7678
(5000 to 10,000—Moderate)
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.

Share and Cite

MDPI and ACS Style

Yaragunda, V.R.; Vaka, D.S.; Oikonomou, E. Land Subsidence Susceptibility Modelling in Attica, Greece: A Machine Learning Approach Using InSAR and Geospatial Data. Earth 2025, 6, 61. https://doi.org/10.3390/earth6030061

AMA Style

Yaragunda VR, Vaka DS, Oikonomou E. Land Subsidence Susceptibility Modelling in Attica, Greece: A Machine Learning Approach Using InSAR and Geospatial Data. Earth. 2025; 6(3):61. https://doi.org/10.3390/earth6030061

Chicago/Turabian Style

Yaragunda, Vishnuvardhan Reddy, Divya Sekhar Vaka, and Emmanouil Oikonomou. 2025. "Land Subsidence Susceptibility Modelling in Attica, Greece: A Machine Learning Approach Using InSAR and Geospatial Data" Earth 6, no. 3: 61. https://doi.org/10.3390/earth6030061

APA Style

Yaragunda, V. R., Vaka, D. S., & Oikonomou, E. (2025). Land Subsidence Susceptibility Modelling in Attica, Greece: A Machine Learning Approach Using InSAR and Geospatial Data. Earth, 6(3), 61. https://doi.org/10.3390/earth6030061

Article Metrics

Back to TopTop