Integrating TS-InSAR and machine learning for spatiotemporal analysis and influencing factors identification of land subsidence in Nansha, Guangzhou, China

Rapid urban expansion and construction on compressible coastal soils have induced significant land subsidence across Nansha District. To quantify this subsidence, we applied Time-Series Interferometric Synthetic Aperture Radar (TS-InSAR) on Sentinel-1A data collected between 2019 and 2023, utilizing the RapidSAR method to identify high-quality points for precise deformation monitoring. On this basis, a multi-source database was developed that incorporated not only common factors (e.g., soft soil thickness, elevation, and construction activity) but also aquaculture areas and most notably a detailed building-type classification that specifically accounts for widespread custom rural houses. A Random Forest regression model was applied to quantitatively identify the main drivers of subsidence. The results reveal a maximum subsidence rate of 283.7 mm/year in Nansha, with 4.5% of monitored points exceeding 40 mm/year and 10.3% exceeding 20 mm/year. Significant subsidence was predominantly concentrated near major engineering construction sites and coastal soft soil zones. The model results indicate that construction activity is the most important factor driving significant subsidence. Additionally, soft soil thickness, aquaculture, and custom rural houses in riverine zones were also identified as leading contributing factors. The findings of this study can provide a scientific basis for future urban planning and flood prevention strategies in Nansha.

1 Introduction

With the rapid advancement of coastal urbanization, land subsidence has emerged as a critical concern, particularly in regions characterized by widespread soft soils and intensive construction activities. Subsidence not only compromises foundation stability but also leads to infrastructure damages, including pipeline ruptures and road cracking, thereby posing significant risks to urban safety. Meanwhile, the global sea level is rising at a rate of 2–5 mm per year (Hamlington et al., 2024), and land subsidence further amplifies saline water intrusion and exacerbates the flood exposure risk of coastal cities. Accurate monitoring of subsidence is thus essential for the management of soft soil foundations, infrastructure planning, and disaster risk mitigation in coastal urban areas.

Nansha District is a critical development zone within the Guangdong–Hong Kong–Macao Greater Bay Area (GBA), situated on extensive, thick layers of soft soil. Over the past decades, large-scale land reclamation has formed the foundation for its urban expansion. However, these reclaimed soils are highly compressible and susceptible to deformation. Recent rapid urban construction, coupled with prevalent groundwater extraction and other anthropogenic activities, has intensified the risk of land subsidence. The interplay of these natural geological conditions and human factors makes Nansha a representative and critical area for studying coastal subsidence. Therefore, a comprehensive investigation into its surface deformation patterns and underlying driving mechanisms is imperative.

Traditional subsidence monitoring techniques, such as GNSS and leveling, offer high precision but are limited by high costs and sparse spatial coverage. In contrast, Interferometric Synthetic Aperture Radar (InSAR) has emerged as a powerful tool for monitoring surface deformation due to its broad coverage and high temporal resolution (Hooper et al., 2012; Wang et al., 2012; Simons and Rosen, 2015; Yang et al., 2024). In particular, time-series InSAR techniques (TS-InSAR), which perform spatiotemporal analysis on stacks of multi-temporal interferograms, effectively mitigate errors such as atmospheric delay (Elliott et al., 2008; Jolivet et al., 2011; Bekaert et al., 2015) and orbital inaccuracies (Biggs et al., 2007). TS-InSAR can achieve millimeter-level deformation monitoring accuracy with advanced analysis methods, notably Small Baseline Subset (SBAS) (Berardino et al., 2002), Persistent Scatter Interferometry (PSI) (Ferretti et al., 2001; Hooper et al., 2004; Hooper et al., 2007), Distributed Scatter InSAR (DS-InSAR) (Ferretti et al., 2011), Π-RATE (Biggs et al., 2007; Wang et al., 2012), GEOS-ATSA (Ng et al., 2015; Ng et al., 2023) and RapidSAR (Spaans and Hooper, 2016).

Numerous prior studies have employed InSAR to monitor subsidence in the Pearl River Delta (PRD). For instance, Wang et al. (2012) used multi-track Envisat satellite data from 2007 to 2010 to construct a three-dimensional deformation field over the PRD region. Their research revealed widespread subsidence hazards induced by human activities in soft soil areas (with thickness >10 m) of the delta. Similarly, Ma et al. (2019) integrated multi-source SAR data from 2011 to 2017 to identify significant subsidence in the soft soil zones of the GBA, especially in the Zhuhai region, where the maximum subsidence rate reached 112.3 mm/year. As a core hub of the GBA, Nansha has undergone rapid development and construction in recent years. Many existing have studies the surface deformation in the PRD relying on PSI and DS-InSAR (Chen et al., 2012; Ng et al., 2018; Ma et al., 2019; Liu et al., 2023). However, these methods exhibit notable limitations in rapidly urbanizing areas, due to the lack of stable scatters in newly developed districts and construction sites. Wang et al. (2024) employed the Small Baseline Subset (SBAS-InSAR) technique to derive land subsidence results in Nansha between December 2015 and June 2019, which revealed subsidence rates ranging primarily from 5 to 40 mm/year. They further conducted overlay analysis combining factors such as soft soil distribution to investigate the causes of subsidence. Although conventional SBAS-InSAR effectively improve coherence through the use of short spatial and temporal baselines, it remains prone to significant noise in areas with high vegetation coverage or widespread aquaculture. As a result, stronger multi-looking processing is often required (Goldstein and Werner, 1998; Benoit et al., 2020), which reduces the spatial resolution of the derived deformation results.

Spaans and Hooper (2016) proposed a method for selecting homogeneous points that not only extracts high-quality points to ensure high-resolution results, but also does not require stable scattering characteristics of ground objects, making it particularly suitable for rapidly developing urban areas. Therefore, this work adopts the method introduced by Spaans and Hooper (2016) for the selection of high-quality points and employs TS-InSAR technology with π-RATE software (Wang et al., 2012) to monitor deformation in Nansha. Although numerous studies have investigated ground deformation in the Nansha region, most have focused on single-factor analyses and seldom systematically integrated or modeled multiple contributing factors. Liu et al. (2023) applied machine learning to systematically analyze and model subsidence in Zhuhai City, achieving promising results. Beyond subsidence, machine learning approaches have also proven effective in assessing other geohazards, such as landslides, using techniques like Artificial Neural Networks (Ebadati et al., 2025) and logistic regression (Cemiloglu et al., 2023). In this work, therefore, focuses on Nansha to integrate TS-InSAR technology with machine learning methods to conduct spatiotemporal analysis and influencing factors identification of surface subsidence. We further aim to explore the driving mechanisms of influencing factors and develop a predictive model, thereby providing data support and decision-making references for urban planning and disaster prevention and control.

2 Materials and methods

2.1 Study area

Nansha District of Guangzhou, China, located at the geometric center of the Pearl River Delta (Figure 1), constitutes a core area of the Guangdong–Hong Kong–Macao Greater Bay Area (GBA), with a permanent population exceeding one million and a land area of approximately 803 km² (Yu et al., 2024). Geologically, soft soil layers cover nearly the entire district, accounting for 91.6% of its total area with an average thickness of 15–20 m (Liu et al., 2022). The district is crossed by a dense river network governed by the Pearl River Delta hydrological system, coupled with frequent changes in surface water recharge and groundwater levels, leading to complex hydrogeological conditions. Historically, extensive reclamation has been a major driver of land formation, with the reclaimed area reaching approximately 553.51 km² (Xiao et al., 2015). These reclaimed foundations are predominantly composed of saturated, silty soft clays with low bearing capacity and high compressibility (Song, 2023), making them highly susceptible to subsidence under engineering disturbances, overburden loads, or groundwater fluctuations. In recent years, rapid urban expansion, particularly in key zones like the Wanqingsha Bay Area, has been prominent. The coastal economy is dominated by aquaculture, where saline groundwater extraction is common (Wang et al., 2024). Additionally, along riverbanks, numerous private rural houses are often built with inadequate foundation treatment.

Figure 1

Figure 1. Location of study area.

2.2 Methods

This study integrates TS-InSAR time series analysis with machine learning-based regression modeling using multi-source geospatial data, to analyze the spatiotemporal evolution and dominant controlling factors of land subsidence in Nansha District, Guangzhou. The workflow of data processing is showed in Figure 2.

Figure 2

Figure 2. Workflow of data processing.

103 Sentinel-1A ascending orbit SAR images from September 2019 to March 2023 were used for the time series analysis. Given Nansha’s extensive aquaculture and agricultural activities, the area is characterized by widespread surface water, vegetation cover, and high soil moisture, which often lead to interferometric decorrelation. To ensure the coherence of interferograms, spatial baselines were strictly limited to less than 200 m, and interferometric pairs with temporal baselines shorter than 36 days were prioritized. The baseline distribution is shown in Figure 3. Data Preprocessing was conducted using the ISCE software (Rosen et al., 2012), where topographic phases were removed using the 1-arcsecond Shuttle Radar Topography Mission Digital Elevation Model (SRTM DEM) (Farr et al., 2007). The selection of high-quality measurement points and coherence estimation were performed following the method by Spaans and Hooper (2016). In contrast to Persistent Scatterer InSAR (PS-InSAR), which requires phase stability over the entire time series, this method employs a stack of short-temporal-baseline interferograms. Moreover, it diverges from conventional SBAS-InSAR techniques that typically apply a boxcar ensemble average for coherence estimation. Instead, the RapidSAR processing chain begins by generating a full-resolution average intensity image from the entire interfrograms subset. Coherence is then estimated for each interferogram within a 15 × 15 moving window using these pixels, and a threshold of >0.25 is applied to select high-quality points on a per-interferogram basis. This approach allows for the inclusion of points that exhibit only temporary coherence, thereby increasing point density in challenging areas. Following this selection, additional multi-looking (2 in azimuth, 8 in range) was applied to the phase data to further enhance quality for per-interferogram. Phase unwrapping was carried out using the branch-cut method (Goldstein et al., 1988; Wang et al., 2012; Jiang and Lohman, 2021). Time series InSAR processing was implemented with π-RATE software (Wang et al., 2012). The Minimum Spanning Tree (MST) algorithm was used to select independent observations (Wang et al., 2012), and the reference phase was derived from the average of all pixels. Polynomial fitting was applied to correct orbital (Biggs et al., 2007) and DEM errors, while stratified delay fitting with a linear function of the SRTM DEM (Elliott et al., 2008) and APS was estimated using spatiotemporal filtering (Ferretti et al., 2001; Wang et al., 2012). Finally, Laplacian smoothing and Tikhonov regularization were employed to reconstruct the time series (Schmidt and Bürgmann, 2003; Wang et al., 2012), with a smoothing factor of −2. The regularization parameter λ ranged from 10^–3 to 10¹ with a logarithmic step size of 0.2. Linear displacement rates were calculated by an iterative weighted least squares method (Wang et al., 2009).

Figure 3

Figure 3. Temporal and perpendicular baseline of Sentinel-1 data used. Red dots indicate the SAR acquisition epochs.

To investigate the driving mechanisms of subsidence, Our study integrated multi-source geospatial factors, including geology, topography, land use, building information, and groundwater-related variables. All data were uniformly projected, resampled, and standardized. Soft soil distribution was derived from a regionally interpolated map based on engineering borehole data from Liu et al. (2022). Topographic information was obtained from the 1-arcsecond SRTM DEM (Farr et al., 2007). Land use data were used to identify construction sites and custom rural houses near rivers. Construction areas were manually interpreted from multi-year high-resolution remote sensing images, while custom rural houses were characterized by small area, low rise, and clustered distribution along rivers. Building height was used as a proxy for structural load, with floor count data sourced from Che et al. (2024). As the unavailability of groundwater monitoring data in the region, and considering that groundwater extraction mainly originates from saline aquaculture, coastal aquaculture zones were adopted as a proxy to assess the potential influence of groundwater on subsidence. After variable preparation, a Random Forest regression model (Breiman, 2001) was applied to model subsidence rates and attribute contributing factors (Strobl et al., 2008; Groemping, 2009). The response variable was the subsidence rate derived from InSAR, and the explanatory variables consisted of the standardized spatial factors. The model was configured with 500 trees, a maximum tree depth of 10, and a minimum of 15 samples per leaf. Seventy percent of the data were used for training, and the remaining 30% were used for validation. The model performance was evaluated via cross-validation, with overall accuracy assessed using the coefficient of determination (R²) and root mean square error (RMSE).

3 Results

3.1 Spatiotemporal subsidence characteristics

3.1.1 Spatial distribution characteristics

This study employed the TS-InSAR technique to derive time-series Subsidence results in Nansha District. The spatial distribution of the subsidence rate is illustrated in Figure 4, where positive values indicate ground subsidence. Subsiding areas are primarily concentrated in Dongchong Town, Wanqingsha, Nansha Street, and parts of Dagang Town. Notably, regions with greater soft soil thickness and high construction intensity exhibit significantly higher subsidence rates compared to surrounding areas.

Figure 4

Figure 4. Map of deformation rate in Nansha (Negative values represent subsidence): (A) Subsidence rate from 2019 to 2023, White solid polygons represent the primary subsidence bowls; (B) Distribution of subsidence rates; (C) standard deviation distribution.

A total of 225,370 measurement points were obtained in the subsidence rate map. The maximum subsidence rate reached 283.7 mm/year. Among all points, 10,114 (4.5%) exhibited subsidence rates exceeding 40 mm/year, indicating severe localized subsidence. Within these, a critical subset of 2,106 points (0.93% of all points) exhibited extreme rates surpassing 100 mm/year. These points are predominantly concentrated within several pronounced subsidence bowls—primarily located in the intensive construction area of Wanqingsha and the deep soft soil/groundwater exploitation zone of Longxue Island in the south—which are outlined as the key risk zones in Figure 4A. Furthermore, 10.3% of the points showed notable subsidence (≥20 mm/year), while 18.8% experienced mild subsidence (>5 mm/year and <20 mm/year). The majority of points (53.7%) remained relatively stable or showed slight uplift, with subsidence rates ranging between −5 mm/year and 5 mm/year.

Firstly, the subsidence rates in the intensive construction zones of Nansha District are significantly elevated. In recent years, Nansha District has experienced rapid development and construction. As shown in Figure 5A, the major construction zones spatially coincide with areas of significant localized subsidence during the study period, indicating that construction activities are one of the key drivers of land subsidence. The spatial distribution of subsidence rates within the construction areas is presented in Figure 5C, where pronounced subsidence is clearly observed. Within these zones, the maximum subsidence rate reaches 283.7 mm/year, with 76.6% of monitoring points exceeding 10 mm/year, 56% exceeding 20 mm/year, and 43.1% exceeding 30 mm/year. These high subsidence rates are concentrated in the core construction areas, forming distinct subsidence funnels.

Figure 5

Figure 5. Distribution of major construction sites in Nansha: (A,B) Spatial distribution of construction areas, superimposed on remote sensing images from 2019 and 2023, respectively; (C) Subsidence rate distribution within construction zones.

Secondly, the subsidence rates in aquaculture areas of Nansha District are significantly higher than those in the surrounding non-aquaculture zones. Large numbers of fish ponds are distributed in regions such as Wanqingsha and Hengli. Figure 6 illustrates the aquaculture distribution in Nansha District. The southern part is dominated by freshwater aquaculture, where groundwater extraction is relatively limited, resulting in comparatively stable subsidence. In contrast, the coastal areas, particularly around Wanqingsha and Longxue Island, are characterized by intensive saline water extraction (Wang et al., 2024). Combined with the long-term alternation of land submergence and drainage, the underlying saturated soft soils maintain in a state of high pore water pressure (Ghobadi-Far et al., 2023), reducing effective stress and leading to much greater subsidence than in non-aquaculture areas. In Figure 6A, the light blue shading represents the main coastal aquaculture-affected zones, delineated by applying a 2 km buffer around concentrated aquaculture areas while excluding construction sites and inland aquaculture regions. Figure 6B shows the spatial distribution of subsidence rates within this main aquaculture-affected area, where 51.8% of monitoring points exhibit rates exceeding 5 mm/year, 38.9% exceed 10 mm/year, and 23% even surpass 20 mm/year. This pattern reveals a distinct subsidence gradient, indicating a spatial correlation between aquaculture activities and land subsidence.

Figure 6

Figure 6. Distribution of aquaculture areas in Nansha (2019–2023): (A) Spatial distribution of aquaculture zones; (B) Subsidence rates within these areas.

In addition, in the southern part of Nansha, soft soils are widely distributed, with thicknesses generally exceeding 15 m and locally surpassing 25 m, as shown in the soft soil thickness map in Figure 7A. The subsidence rate map of Nansha District (Figure 4A) reveals that widespread mild subsidence is present in the southern areas, the spatial pattern is similar to the distribution of thick soft soil layers. After excluding monitoring points within construction and aquaculture zones, subsidence rates across six soft soil thickness intervals were compared with the corresponding soil thickness, as shown in Figure 7B. The measured average subsidence rates exhibited a strong positive correlation with soft soil thickness. For thickness ranges of <10 m, 10–15 m, 15–20 m, 20–25 m, 25–30 m and >30 m, the corresponding average subsidence rates were −1.5 mm/year, 0.8 mm/year, 2.0 mm/year, 6.1 mm/year, 8.1 mm/year and 7.6 mm/year, respectively. The results demonstrate a clear positive correlation between average subsidence rates and soft soil thickness. However, when the soft soil thickness exceeds 30 m, the correlation becomes weaker, mainly because areas with thickness greater than 30 m are largely aquaculture zones with fewer monitoring points. Therefore, in this work, the fitting was performed using average subsidence rates corresponding to soft soil thicknesses below 30 m. A linear fit yielded a high coefficient of determination (R² = 0.959), while a second-order polynomial fit provided an even better approximation of the data, with an R² value of 0.974, as illustrated in Figure 7C.

Figure 7

Figure 7. Relationship between soft soil and subsidence: (A) Soft soil thickness; (B) Distribution of subsidence rates across different soft soil thickness zones; (C) Relationship between soft soil thickness and average subsidence rate; (D–F) Subsidence rate, building height, and soft soil thickness in urban built-up areas and custom rural housing areas, respectively.

It is noteworthy that within each soft soil thickness interval, some points exhibit relatively small subsidence (Figure 7B), which may be related to building foundation treatments. Our analysis indicates that in most deep soft soil areas of Wanqingsha, subsidence in mature urban construction zones is lower than that observed in custom rural housing areas. A comparative analysis was conducted between one urban construction zone and one rural custom rural housing zone, both with an average soft soil thickness of approximately 25 m, as shown in Figure 7D. In Area B (blue box in Figure 7D), with an average building height of 7.1 m and an average floor area of 81 m², the mean subsidence rate is 9.4 mm/year. In contrast, in Area A (white box in Figure 7D), with an average building height of 9.5 m and an average floor area of 200 m², the mean subsidence rate is only 4.0 mm/year. This observation diverges from conventional building load–subsidence theory, suggesting that building loads are not the dominant factor controlling land subsidence. The underlying reason may be that high-rise buildings generally adopt pile foundations or other forms of deep foundations, whereas custom rural houses mostly rely on shallow foundations, which impose their loads directly on shallow soft soil layers, thereby leading to more pronounced surface subsidence.

3.1.2 Temporal evolution characteristics

To better quantify the temporal characteristics of the subsidence process, two representative points located within construction zones were selected for time-series analysis, which were further compared with historical remote sensing images of the corresponding areas, as illustrated in Figure 8. Point P1 is located within an industrial park construction site. During the study period, this area was mainly in the phases of earth filling and compaction (Figures 8D–F). The time-series results (Figure 8C) show that subsidence at P1 occurred at a relatively high rate before August 2020, after which the rate gradually decreased. Point P2 is situated at the site of a metro station construction project. Figures 8G–I depict the progression of this site from land leveling to the completion of the main structure between 2020 and 2023, with the line officially opening in September 2021. The time-series results indicate that P2 experienced substantial subsidence before June 2021, after which the ground surface stabilized. As observed from the 2021 remote sensing imagery (Figure 8H), the basic framework of the site had already been completed. The InSAR derived subsidence is consistent with the engineering construction timeline. The post-construction subsidence rate at this point remained below 10 mm/year.

Figure 8

Figure 8. Time series of subsidence in construction areas: (A) Subsidence rate at point P1; (B) Subsidence rate at point P2; (C) Time series of subsidence at points P1 and P2; (D–F) Remote sensing images of point P1 in 2019, 2021, and 2023, respectively; (G–I) Remote sensing images of point P2 in 2019, 2021, and 2023, respectively.

3.2 Field investigation and validation

To further validate the subsidence monitoring results derived from TS-InSAR, several representative sites were selected within the deep soft soil areas of Nansha, covering zones of severe, moderate, and minor subsidence, as shown in Figures 9A–C. Field investigations and on-site comparisons were conducted accordingly. Site a, located in a construction area of Zhujiang Subdistrict, exhibited severe localized surface subsidence, consistent with the high subsidence rates identified by InSAR. This subsidence is inferred to be primarily caused by soft soil consolidation induced by construction activities. Sites b and c are situated in Jiaomen, the central urban area of Nansha, surrounded by high-rise buildings. The area is characterized by deep soft soil deposits, and significant ground subsidence has been observed in the vicinity, but no structural cracks in the buildings were observed. These observations are consistent with the InSAR results, which indicate that while high-rise structures remain relatively stable, but pronounced ground subsidence has occurred in the surrounding area of the high-rise building. Sites d–f are located in Dongchong Town. Site d, adjacent to a construction area, showed structural damages such as wall cracks and uneven surface subsidence. Site e exhibited slight ground subsidence, while Site f, a custom rural house with minimal foundation treatment, showed more severe ground subsidence. Field photographs confirmed that these areas of structural damage correspond well with the subsidence zones identified by TS-InSAR (Figures 9D–I), thereby validating the spatial accuracy of the monitoring results.

Figure 9

Figure 9. Subsidence rates and field survey results of field survey points: (A–C) located in Zhujiang, Jiaomen, Dongchong, respectively: The white circles (A–F) mark the field investigation areas showed in Figures 9D–I, respectively. (D–I) Investigation results corresponding to site a-f, respectively.

Point P3 (near the Guangzhou Foreign Language School) provides a clear illustration of construction-related subsidence (Figure 10). Following the initiation of the school’s second-phase construction in December 2020, the spatial InSAR results show localized subsidence around the site. Critically, the time-series data (Figure 10B) reveals a sustained deformation trend beginning in 2021, demonstrating a strong temporal coincidence with the construction period. This correlation strongly suggests that the ground settlement was likely triggered or accelerated by the construction activities.

Figure 10

Figure 10. Time series of subsidence in construction areas: (A) Subsidence rate at point P3; (B) Time series of subsidence at points P3.

3.3 Modeling and analysis of surface subsidence

In Section 3.1, potential influencing factors of land subsidence in Nansha were preliminarily identified based on its spatial distribution characteristics. To further clarify the driving mechanisms of these factors and quantify their relative contributions, this study constructed both multiple linear regression and random forest regression models. The input variables included six environmental factors: distribution of construction activities (showed in Figure 5A), soft soil thickness distribution (showed in Figure 7A), distribution of custom rural houses relative to river proximity, aquaculture distribution (showed in Figure 6A), building height, and topography. A total of 70% of the sample points were used for model training, while the remaining 30% were reserved for independent validation.

The modelling results demonstrate that the random forest model was able to capture the spatial distribution of subsidence more stably. The coefficient of determination (R²) reached 0.73 for the training set and 0.61 for the validation set, indicating reasonable explanatory and predictive capability. The results are presented in Figure 11. Figure 11A shows the ranking of variable importance, where construction activities emerged as the dominant driver of significant subsidence in Nansha, with a normalized importance score of 0.56. In regions with intensive construction activities, the model predicted significantly higher subsidence rates, highlighting the combined effects of foundation disturbance, groundwater level changes, and soil consolidation due to surcharge loads. Within construction areas, 30% of predicted values exceeded 15 mm/year. The relatively large discrepancies between models and observations in these areas may stem from the absence of detailed construction records, as average subsidence rates were used as proxies for construction activity.

Figure 11

Figure 11. Machine learning modeling results: (A) Variable importance ranking; (B) InSAR-derived subsidence rate at 100 m resolution; (C) Model subsidence rate from the machine learning; (D) Difference between InSAR-observed and model subsidence rates; (E) Distribution of the differences.

The second most important factor was soft soil distribution, with a score of 0.15, underscoring the strong geological control on subsidence. Areas underlain by soft soils exhibited significantly higher subsidence rates than other soil types. Soft soils further interacted with construction activities, leading to amplified subsidence in areas where both factors were present. The model further identified the distribution of custom rural houses near rivers and aquaculture zones as notable contributors, each with an importance score of 0.11. This finding corroborates the observations in Section 3.1, where dense clusters of custom rural houses were found along natural riverbanks, coinciding with mild subsidence zones, reflecting the combined effects of weak geological foundations and poor ground conditions. In contrast, topography and building height contributed relatively little, with building height scoring only 0.01, suggesting that structural loading plays a negligible role in controlling subsidence.

Figures 11B–D compare the model results with actual monitoring data. The overall RMSE between models and observations was 11.3 mm/year, as showed in Figure 11E. The proportions of points with absolute errors below 5 mm/year, 10 mm/year, and 15 mm/year were 55.7%, 80.1%, and 89.5%, respectively. The spatial prediction maps show that the most severe subsidence is concentrated in areas of intensive construction, while widespread mild subsidence occurs across regions of thick soft soils. Overall, the model rate spatial patterns align well with the InSAR-derived monitoring results.

4 Discussion

This study, based on TS-InSAR time-series techniques combined with random forest modeling, systematically reveals the spatiotemporal characteristics and primary driving factors of land subsidence in Nansha District, Guangzhou. The major contribution of this research lies in validating the application of the Rapid-InSAR high-quality point selection technique in rapidly developing urban areas. This method not only maintains high spatial resolution but also effectively preserves subsidence information in unstable scattering regions, thereby facilitating the monitoring of subsidence evolution during construction periods. In addition, by incorporating custom rural housing areas along riverbanks into the factor analysis, the study further confirms the significant correlation between densely distributed custom rural houses and mild subsidence zones. Under conditions of soft soils and shallow overburden, such areas emerge as potential high-risk subsidence hotspots. To more clearly elaborate on the internal mechanisms of these findings and provide external comparisons, the following discussion is structured around four aspects: comparison with existing studies, future engineering applications, limitations of this study, and directions for further research.

4.1 Comparative analysis with previous studies

Compared with the spatial distribution of subsidence in Nansha derived by Wang et al. (2024) using 2016–2019 data, the subsidence patterns identified in this study show strong overall consistency, with the most significant subsidence concentrated in construction areas and coastal aquaculture zones. The main difference lies in the frequent changes in construction activities in recent years, which have altered the distribution of newly developed areas, leading to discrepancies in the most recent subsidence patterns compared to those reported by Wang et al. (2024). In addition, the widespread subsidence caused by thick soft soil deposits shows both similar spatial distribution and comparable subsidence magnitude to those observed by Wang et al. (2024).

In terms of subsidence analysis, a key distinction from previous studies is that this research explicitly incorporates “custom rural houses along riverbanks” as an influencing factor. This inclusion reveals marked differences in subsidence responses between different building types (e.g., rural custom rural houses versus urban high-rise buildings) under similar soft soil thickness conditions. Furthermore, by applying machine learning, this study quantifies the relative contributions of multiple subsidence drivers in Nansha. Regarding the quantitative relationship between soft soils and subsidence, Wang et al. (2024),employed spatial correlation analysis based on a subset of sample points some of which overlapped with active construction zones. This overlap amplified the adjustment parameter a in their model, suggesting higher sensitivity to load variations. In contrast, this study adopted a regression approach based on pixel-level averages across the entire study area, eliminating sampling bias. After controlling for construction-related disturbances, the resulting parameter showed high consistency with the soft soil–subsidence relationship reported by Liu et al. (2023) in Zhuhai. Additionally, Ma et al. (2019) reported a mean subsidence rate of 7.2 mm/year in the Zhongshan–Nansha region, which closely matches the fitted value for deep soft soil zones calculated from the adjustment parameter a in this study, further validating its reliability.

While Liu et al. (2023) modelled subsidence in Zhuhai using machine learning with factors such as soft soil thickness, groundwater extraction intensity, land use type, lithology, and topography, this study employed a different set of predictors—construction distribution, soft soil distribution, proximity of custom rural houses to rivers, aquaculture distribution, building height, and topography. The overall model performance is comparable; however, the RMSE of the predicted values in this study was 11.3 mm/year, while Liu et al. (2023) achieved an error below 3 mm/year. Several factors may explain this difference. First, the study area of Liu et al. was broader, with a much smaller proportion of thick soft soil zones compared to Nansha, where prediction errors are generally larger in deep soft soil regions. Second, in parts of Dongchong Town, this study’s predictions were higher than observed, mainly due to thick soft soil layers in the area, although no corresponding subsidence was detected in the monitoring results, consistent with earlier observations (e.g., Wang et al., 2024). Since the modelling covered the entire Nansha District, this mismatch resulted in larger prediction errors. Third, Liu et al. (2023) employed PS-InSAR, which focuses on stable scatters, while this study included numerous points in active construction areas. The lack of detailed field data for these sites likely reduced prediction accuracy.

4.2 Practical engineering applications

The findings of this study provide valuable references for future urban development, underground engineering design, and risk management in Nansha District. On one hand, the TS-InSAR time-series method demonstrates the capability to monitor subsidence throughout the entire construction process, offering continuous and quantifiable deformation information before, during, and after construction. This supports dynamic construction management and the development of subsidence early-warning systems. On the other hand, the variable importance results from the random forest model indicate that the distribution of custom rural houses exerts a significant influence on land subsidence in Nansha. In particular, custom rural houses located along riverbanks and lacking foundation reinforcement are more prone to pronounced subsidence induced by soft soil consolidation or hydrological disturbances. This insight can serve as a reference for building risk assessment in the district.

4.3 Limitations

Although this study provides a comprehensive framework for the methodological approach and factor analysis, certain limitations remain. The primary limitation is the lack of dynamic groundwater level data for construction areas in Nansha, which prevents the development of a coupled “water–soil–load” model for the region. Consequently, the model cannot fully capture the true temporal variations during the construction period, affecting the prediction accuracy in local areas. This is particularly critical for large construction sites, where groundwater extraction and recharge can induce short-term, significant subsidence events, which have not yet been effectively incorporated into the current modelling framework.

A second limitation is the spatial resolution of SAR imagery. During field surveys, significant ground subsidence was observed in certain residential areas with high-rise buildings; however, these subsidence was not detected in the InSAR results. This discrepancy may be attributed to the limited resolution of the SAR data. The InSAR measurements in such settings are predominantly influenced by the stable phase responses from elevated structures, which can mask ground surface subsidence. Consequently, higher-resolution SAR data are needed in the future to better distinguish between structural stability of high-rise buildings and ground surface subsidence. A second limitation is the spatial resolution and the inherent averaging effect of the InSAR processing. During field surveys, significant ground subsidence was observed in certain residential areas with high-rise buildings; however, this subsidence was not clearly captured in our InSAR results. This discrepancy arises because our processing chain, which applied multi-looking (2 in azimuth, 8 in range) to enhance phase stability in low-coherence areas, results in a ground resolution cell of approximately 40 m × 40 m. While the data was geocoded and resampled onto a 30 m grid (aligned with the SRTM DEM) for geographical referencing and integration with other GIS datasets, the effective spatial resolution—determined by the prior multi-looking step—remains at approximately 40 m. At this resolution, the signal from a building (a stable, dominant scatterer) is averaged with the surrounding ground within the same pixel. For future studies aiming to monitor deformation in complex built environments, two complementary approaches are recommended: (1) utilizing higher-resolution SAR data (e.g., TerraSAR-X, COSMO-SkyMed): to better isolate scattering sources, and (2) employing Persistent Scatterer InSAR (PS-InSAR) techniques on standard data. PS-InSAR can maintain the sensor’s native high resolution by focusing on permanently stable point targets, potentially yielding a denser measurement network around existing infrastructure for direct comparison and validation.

Another limitation concerns the classification of self-built rural houses. These structures were identified primarily through manual interpretation of high-resolution imagery rather than automated classification. As a result, no quantitative uncertainty assessment (e.g., classification accuracy) was performed for this category. While this may introduce local positional or categorical uncertainty, it does not alter the main conclusions of the study, which rely on broader, quantitatively validated factors such as geology and groundwater extraction. Future work would benefit from using standardized building footprint datasets or automated classification to reduce this subjectivity.

4.4 Future directions

Future research can be further expanded in two directions. First, by integrating groundwater monitoring well data and groundwater level models, a coupled groundwater–subsidence model can be developed to enhance the physical interpretability of construction-induced subsidence. Second, by combining SAR time-series data with building structural information (e.g., building age, structural type, foundation type), more engineering-oriented subsidence identification and prediction models can be established. Considering the current status of urban underground space development, future studies may also explore the potential applications of TS-InSAR in deep foundation pit projects, metro line subsidence control, and underground pipeline subsidence risk assessment, thereby supporting the development of regional urban safety and infrastructure management systems.

5 Conclusion

We employed TS-InSAR to monitor land subsidence in Nansha District from 2019 to 2023. The results reveal that maximum subsidence rate reached 283.7 mm/year. Subsidence rates exceeding 40 mm/year, ≥20 mm/year, and 5–20 mm/year accounted for 4.5%, 10.3%, and 18.8% of points, respectively. Subsidence areas are mainly concentrated in Dongchong Town, Wanqingsha, Nansha Street, and parts of Dagang Town. Despite the challenges of dense urban construction, the method preserved some coherent points within newly developed areas. Significant subsidence was concentrated in large-scale construction sites, with 76.6% of monitoring points exceeding 10 mm/year. In addition, widespread mild subsidence was spatially correlated with the distribution of soft soil layers, highlighting the geological control on deformation patterns. The strong correlation between soft soil thickness and ground subsidence is strongly demonstrated by our models, with polynomial regression (R² = 0.974) providing a marginally better fit than linear regression (R² = 0.959).

Based on the Random Forest model analysis, urban construction (importance score: 0.56) was identified as the primary driver of land subsidence in Nansha, followed by soft soil distribution (0.15), with custom rural houses near rivers and aquaculture zones also being significant contributing factors (0.11 each). Notably, subsidence associated with custom rural houses was greater than that observed in mature urban areas, suggesting a higher vulnerability of such buildings in the future. This finding emphasizes the necessity of integrating TS-InSAR with building-type information to establish predictive models and to track the progressive subsidence of different structures. Overall, the results demonstrate that TS-InSAR is a reliable tool for long-term monitoring of urban subsidence in coastal megacities such as Nansha, and it provides critical insights for risk assessment, infrastructure management, and sustainable urban development.

Data availability statement

Publicly available datasets were analyzed in this study. This data can be found here: Alaska Satellite Facility (ASF) DAAC at https://search.asf.alaska.edu/.

Author contributions

QH: Writing – original draft, Conceptualization, Methodology, Writing – review and editing, Validation. CZ: Data curation, Writing – original draft, Validation, Project administration, Investigation, Writing – review and editing. XY: Investigation, Methodology, Writing – review and editing, Formal Analysis. HW: Writing – review and editing, Software, Methodology.

Funding

The author(s) declared that financial support was received for this work and/or its publication. HW would like to acknowledge support from the ICTP through the Associates Programme and from the Simons Foundation through grant number 284558FY19.

Acknowledgements

The authors acknowledge the European Space Agency (ESA) Copernicus Program for providing the Sentinel-1 data and the ISCE team for the open-source software utilized for Differential Interferogram Generation processing.

Conflict of interest

Authors QH and CZ were employed by Guangdong Provincial Construction Engineering Quality & Safety Testing Head Station Co., Ltd.

The remaining author(s) declared that this work was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The author(s) declared that generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Bekaert, D. P. S., Hooper, A., and Wright, T. J. (2015). A spatially variable power law tropospheric correction technique for InSAR data. J. Geophys. Res. Solid Earth 120, 1345–1356. doi:10.1002/2014JB011558

CrossRef Full Text | Google Scholar

Benoit, A., Puyssegur, B., Jolivet, R., and Lasserre, C. (2020). CorPhU: an algorithm based on phase closure for the correction of unwrapping errors in SAR interferometry. Geophys. J. Int. 221, 1959–1970. doi:10.1093/gji/ggaa120

CrossRef Full Text | Google Scholar

Berardino, P., Fornaro, G., Lanari, R., and Sansosti, E. (2002). A new algorithm for surface deformation monitoring based on small baseline differential SAR interferograms. IEEE Trans. Geosci. Remote Sens. 40, 2375–2383. doi:10.1109/TGRS.2002.803792

CrossRef Full Text | Google Scholar

Biggs, J., Wright, T., Lu, Z., and Parsons, B. (2007). Multi-interferogram method for measuring interseismic deformation: denali fault, Alaska. Geophys. J. Int. 170, 1165–1179. doi:10.1111/j.1365-246X.2007.03415.x

CrossRef Full Text | Google Scholar

Breiman, L. (2001). Random forests. Mach. Learn. 45, 5–32. doi:10.1023/A:1010933404324

CrossRef Full Text | Google Scholar

Cemiloglu, A., Zhu, L., Mohammednour, A. B., Azarafza, M., and Nanehkaran, Y. A. (2023). Landslide susceptibility assessment for Maragheh County, Iran, using the logistic regression algorithm. Land 12, 1397. doi:10.3390/land12071397

CrossRef Full Text | Google Scholar

Che, Y., Li, X., Liu, X., Wang, Y., Liao, W., Zheng, X., et al. (2024). 3D-GloBFP: the first global three-dimensional building footprint dataset. Earth Syst. Sci. Data 16, 5357–5374. doi:10.5194/essd-16-5357-2024

CrossRef Full Text | Google Scholar

Chen, F., Lin, H., Zhang, Y., and Lu, Z. (2012). Ground subsidence geo-hazards induced by rapid urbanization: implications from InSAR observation and geological analysis. Nat. Hazards Earth Syst. Sci. 12, 935–942. doi:10.5194/nhess-12-935-2012

CrossRef Full Text | Google Scholar

Ebadati, M., Sun, F., Lee, Y., and Moghadam, M. J. (2025). Landslide susceptibility analysis using artificial neural networks for Chalus County, Iran. Civ. Geoeng. Lett. 2, e100039. doi:10.22034/CGEL.2.2.e100039

CrossRef Full Text | Google Scholar

Elliott, J. R., Biggs, J., Parsons, B., and Wright, T. J. (2008). InSAR slip rate determination on the Altyn Tagh Fault, northern Tibet, in the presence of topographically correlated atmospheric delays. Geophys. Res. Lett. 35, 1–5. doi:10.1029/2008gl033659

CrossRef Full Text | Google Scholar

Farr, T. G., Rosen, P. A., Caro, E., Crippen, R., Duren, R., Hensley, S., et al. (2007). The shuttle radar topography mission. Rev. Geophys. 45, RG2004. doi:10.1029/2005RG000183

CrossRef Full Text | Google Scholar

Ferretti, A., Prati, C., and Rocca, F. (2001). Permanent scatterers in SAR interferometry. IEEE Trans. Geosci. Remote Sens. 39, 8–20. doi:10.1109/36.898661

CrossRef Full Text | Google Scholar

Ferretti, A., Fumagalli, A., Novali, F., Prati, C., Rocca, F., and Rucci, A. (2011). A new algorithm for processing interferometric data-stacks: SqueeSAR. IEEE Trans. Geosci. Remote Sens. 49, 3460–3470. doi:10.1109/TGRS.2011.2124465

CrossRef Full Text | Google Scholar

Ghobadi-Far, K., Werth, S., Shirzaei, M., and Burgmann, R. (2023). Spatiotemporal groundwater storage dynamics and aquifer mechanical properties in the Santa Clara valley inferred from InSAR deformation over 2017-2022. Geophys. Res. Lett. 50, e2023GL105157. doi:10.1029/2023GL105157

CrossRef Full Text | Google Scholar

Goldstein, R. M., and Werner, C. L. (1998). Radar interferogram filtering for geophysical applications. Geophys. Res. Lett. 25, 4035–4038. doi:10.1029/1998GL900033

CrossRef Full Text | Google Scholar

Goldstein, R. M., Zebker, H. A., and Werner, C. L. (1988). Satellite radar interferometry: two-dimensional phase unwrapping. Radio Sci. 23, 713–720. doi:10.1029/RS023i004p00713

CrossRef Full Text | Google Scholar

Groemping, U. (2009). Variable importance assessment in regression: linear regression versus random forest. Am. Stat. 63, 308–319. doi:10.1198/tast.2009.08199

CrossRef Full Text | Google Scholar

Hamlington, B. D., Bellas-Manley, A., Willis, J. K., Fournier, S., Vinogradova, N., Nerem, R. S., et al. (2024). The rate of global sea level rise doubled during the past three decades. Commun. Earth Environ. 5, 601. doi:10.1038/s43247-024-01761-5

CrossRef Full Text | Google Scholar

Hooper, A., Zebker, H., Segall, P., and Kampes, B. (2004). A new method for measuring deformation on volcanoes and other natural terrains using InSAR persistent scatterers. Geophys. Res. Lett. 31, L23611. doi:10.1029/2004GL021737

CrossRef Full Text | Google Scholar

Hooper, A., Segall, P., and Zebker, H. (2007). Persistent scatterer interferometric synthetic aperture radar for crustal deformation analysis, with application to volcan Alcedo, Galapagos. J. Geophys. Res. Solid Earth 112, B07407. doi:10.1029/2006JB004763

CrossRef Full Text | Google Scholar

Hooper, A., Bekaert, D., Spaans, K., and Arikan, M. (2012). Recent advances in SAR interferometry time series analysis for measuring crustal deformation. TECTONOPHYSICS 514, 1–13. doi:10.1016/j.tecto.2011.10.013

CrossRef Full Text | Google Scholar

Jiang, J., and Lohman, R. B. (2021). Coherence-guided InSAR deformation analysis in the presence of ongoing land surface changes in the imperial valley, California. Remote Sens. Environ. 253, 112160. doi:10.1016/j.rse.2020.112160

CrossRef Full Text | Google Scholar

Jolivet, R., Grandin, R., Lasserre, C., Doin, M. P., and Peltzer, G. (2011). Systematic InSAR tropospheric phase delay corrections from global meteorological reanalysis data. Geophys. Res. Lett. 38, L17311. doi:10.1029/2011GL048757

CrossRef Full Text | Google Scholar

Liu, W., Chen, L., Zhang, Q., Huang, J., and Wu, L. (2022). Distribution and engineering characteristics of soft clay in Nansha, Guangzhou. J. Yangtze River Sci. Res. Inst. 39, 94–99. doi:10.11988/ckyyb.20200786

CrossRef Full Text | Google Scholar

Liu, Z., Ng, A. H. M., Wang, H., Chen, J., Du, Z., and Ge, L. (2023). Land subsidence modeling and assessment in the west pearl river Delta from combined InSAR time series, land use and geological data. Int. Int. Int. J. Appl. Earth Obs. Geoinf. 118, 103228. doi:10.1016/j.jag.2023.103228

CrossRef Full Text | Google Scholar

Ma, P., Wang, W., Zhang, B., Wang, J., Shi, G., Huang, G., et al. (2019). Remotely sensing large- and small-scale ground subsidence: a case study of the guangdong–Hong Kong–Macao greater Bay area of China. Remote Sens. Environ. 232, 111282. doi:10.1016/j.rse.2019.111282

CrossRef Full Text | Google Scholar

Ng, A. H. M., Ge, L. L., and Li, X. J. (2015). Assessments of land subsidence in the Gippsland Basin of Australia using ALOS PALSAR data. Remote Sens. Environ. 159, 86–101. doi:10.1016/j.rse.2014.12.003

CrossRef Full Text | Google Scholar

Ng, A. H. M., Wang, H., Dai, Y., Pagli, C., Chen, W., Ge, L., et al. (2018). InSAR reveals land deformation at Guangzhou and Foshan, China between 2011 and 2017 with COSMO-SkyMed data. Remote Sens. 10, 813. doi:10.3390/rs10060813

CrossRef Full Text | Google Scholar

Ng, A. H. M., Liu, Z. Y., Du, Z. Y., Huang, H. W., Wang, H., and Ge, L. L. (2023). A novel framework for combining polarimetric Sentinel-1 InSAR time series in subsidence monitoring-A case study of Sydney. Remote Sens. Environ. 295, 113694. doi:10.1016/j.rse.2023.113694

CrossRef Full Text | Google Scholar

Rosen, P. A., Gurrola, E., Sacco, G. F., and Zebker, H. (2012). “The InSAR scientific computing environment,” in Proceedings of the 2012 9th European conference on synthetic aperture radar (EUSAR) (Nuremberg, Germany: VDE Verlag GmbH), 730–733.

Google Scholar

Schmidt, D. A., and Bürgmann, R. (2003). Time-dependent land uplift and subsidence in the Santa Clara valley, California, from a large interferometric synthetic aperture radar data set. J. Geophys. Res. Solid Earth 108, 2416. doi:10.1029/2002JB002267

CrossRef Full Text | Google Scholar

Simons, M., and Rosen, P. A. (2015). “Interferometric synthetic aperture radar geodesy,” in Treatise on geophysics. Second Edition (Oxford: Elsevier press), 339–385.

CrossRef Full Text | Google Scholar

Song, X. (2023). Analysis of the deformation and failure mechanism of a utility tunnel excavation in the deep and soft clay area of Nansha, Guangzhou. Chin. J. Rock Mech. Eng. 42, 3629–3642. doi:10.13722/j.cnki.jrme.2021.1218

CrossRef Full Text | Google Scholar

Spaans, K., and Hooper, A. (2016). InSAR processing for volcano monitoring and other near-real time applications. J. Geophys. Res. Solid Earth 121, 2947–2960. doi:10.1002/2015JB012752

CrossRef Full Text | Google Scholar

Strobl, C., Boulesteix, A.-L., Kneib, T., Augustin, T., and Zeileis, A. (2008). Conditional variable importance for random forests. BMC Bioinforma. 9, 307. doi:10.1186/1471-2105-9-307

PubMed Abstract | CrossRef Full Text | Google Scholar

Wang, H., Wright, T. J., and Biggs, J. (2009). Interseismic slip rate of the northwestern Xianshuihe fault from InSAR data. Geophys. Res. Lett. 36, 1–5. doi:10.1029/2008gl036560

CrossRef Full Text | Google Scholar

Wang, H., Wright, T. J., Yu, Y., Lin, H., Jiang, L., Li, C., et al. (2012). InSAR reveals coastal subsidence in the Pearl river Delta, China. Geophys. J. Int. 191, 1119–1128. doi:10.1111/j.1365-246X.2012.05687.x

CrossRef Full Text | Google Scholar

Wang, S., Sun, H., Wei, L., Pi, P., Zeng, M., Pan, Y., et al. (2024). Integrated assessment of coastal subsidence in Nansha district, Guangzhou city, China: insights from SBAS-InSAR monitoring and risk evaluation. Remote Sens. 16, 248. doi:10.3390/rs16020248

CrossRef Full Text | Google Scholar

Xiao, R., Bai, J., Lu, Q., Zhao, Q., Gao, Z., Wen, X., et al. (2015). Fractionation, transfer, and ecological risks of heavy metals in riparian and ditch wetlands across a 100-year chronosequence of reclamation in an estuary of China. Sci. Total Environ. 517, 66–75. doi:10.1016/j.scitotenv.2015.02.052

PubMed Abstract | CrossRef Full Text | Google Scholar

Yang, X., Wang, H., Pagli, C., Ng, A. H. M., and He, Q. (2024). Closure-based correction of InSAR phase unwrapping errors by integrating block-wise tikhonov regularization and flux analysis. IEEE Trans. Geosci. Remote Sens. 62, 5219215–15. doi:10.1109/TGRS.2024.3437757

CrossRef Full Text | Google Scholar

Yu, H., Li, B., Xiao, Y., Sun, J., Chen, C., Jin, G., et al. (2024). Surface subsidence over a coastal city using SBAS-InSAR with Sentinel-1A data: a case of Nansha district, China. Remote Sens. 16, 55. doi:10.3390/rs16010055

CrossRef Full Text | Google Scholar