Study on the Relationship between Groundwater and Land Subsidence in Bangladesh Combining GRACE and InSAR

Abstract

Due to a heavy reliance on groundwater, Bangladesh is experiencing a severe decline in groundwater storage, with some areas even facing land subsidence. This study aims to investigate the relationship between groundwater storage changes and land subsidence in Bangladesh, utilizing a combination of GRACE and InSAR technologies. To clarify this relationship from a macro perspective, the study employs GRACE data merged with GLDAS to analyze changes in groundwater storage and SBAS-InSAR technology to assess land subsidence. The Dynamic Time Warping (DTW) method calculates the similarity between groundwater storage and land subsidence time series, incorporating precipitation and land cover types into the data analysis. The findings reveal the following: (1) Groundwater storage in Bangladesh is declining at an average rate of −5.55 mm/year, with the most significant declines occurring in Rangpur, Mymensingh, and Rajshahi. Notably, subsidence areas closely match regions with deeper groundwater levels; (2) The similarity coefficient between the time series of groundwater storage and land subsidence changes exceeds 0.85. Additionally, land subsidence in different regions shows an average lagged response of 2 to 6 months to changes in groundwater storage. This study confirms a connection between groundwater dynamics and land subsidence in Bangladesh, providing essential knowledge and theoretical support for further research.

1. Introduction

Land subsidence is often hidden and difficult to detect during its development. Once it occurs, it poses significant threats to surface infrastructure, urban buildings, and human safety. One primary cause of land subsidence is the over-extraction of groundwater [1]. The widespread pumping of groundwater, especially deep groundwater, leads to substantial depletion of aquifer storage and compaction of underlying sediments, resulting in land subsidence [2].

As an agriculture-dependent nation, Bangladesh uses approximately 80% of its water for agricultural purposes. Since the 1970s, around 10 million wells have been installed in the country [3]. Until now, about 80% of irrigation water during the dry season and nearly 98% of drinking water come from groundwater [4]. This heavy reliance on groundwater has led to depletion in various regions of Bangladesh, including the northwestern, central-northern, and southwestern areas, between 1985 and 2005 [5]. For instance, the groundwater level near the Buriganga River has decreased by over 1 m [6]. Three-quarters of Bangladesh’s land consists mainly of compressible alluvial deposits and exceptionally soft organic soils [7]. Syvitski et al. and Brownd et al. identify groundwater extraction as a significant factor contributing to land subsidence in the Ganges-Brahmaputra Delta, especially in the southwestern region [8,9]. Due to excessive groundwater extraction, the urban landscape of the Dhaka region now lies 60–70 m below the surface [10]. Consequently, land subsidence has become a pressing issue in Bangladesh.

The most prevalent method for studying land subsidence to date is via Interferometric Synthetic Aperture Radar (InSAR) technology. This technology finds extensive application in processing radar data from a variety of satellite missions, including Sentinel-1, ALOS, TerraSAR-X, and others, enabling continuous monitoring of land displacement with millimeter-level accuracy under all weather conditions [11]. It has been widely applied in terrain mapping and displacement monitoring, making it a crucial tool for current subsidence studies. Millimeter-scale subsidence has been observed in various regions, including Mexico [12], China [13], Italy [14], Bangladesh [15], and India [16], confirming the reliability of InSAR. In Bangladesh, InSAR monitoring in Dhaka and Khulna has detected land subsidence rates of 0 to 10 mm/year in Dhaka and a maximum rate of 16 mm/year in Khulna, validating the efficacy of InSAR technology in the country [15,17]. Although InSAR can accurately sense land subsidence signals, high resolution for large areas often generates vast data volumes, requiring significant time and storage capacity for processing [18]. The Looking Inside the Continents from Space (LiCS) project, led by the UK’s Centre for the Observation and Modelling of Earthquakes, Volcanoes and Tectonics (COMET), monitors global tectonic and volcanic belts using Sentinel-1. This project provides processed unwrapped and interferometric images through software like SNAP and GAMMA, saving researchers substantial time on data preprocessing [19]. The launch of the LiCS project has enabled large-scale monitoring of land displacement.

Since groundwater is often a primary factor in land subsidence, monitoring it is crucial. Traditional groundwater monitoring typically involves periodic observations of groundwater level changes in multiple wells, estimating groundwater storage changes around these wells. However, the high construction and maintenance costs of observation wells, along with their uneven distribution, make it challenging to monitor large-scale groundwater storage changes [20]. The launch of the Gravity Recovery and Climate Experiment (GRACE) satellite has facilitated long-term, large-scale monitoring of groundwater changes. Shammsudduha et al. validated the reliability of GRACE satellite data for studying groundwater in Bangladesh by using in-situ well data from a network of groundwater monitoring stations [21]. With advancements in remote sensing technology and satellite data, research combining GRACE and InSAR to study groundwater storage and land subsidence has gained importance. Castellazzi et al. were the first to quantitatively monitor groundwater storage changes in complex aquifers using GRACE and InSAR. They applied this analysis to central Mexico to explore the effects of climate change and human activity on groundwater storage variations [22]. Subsequently, many researchers have conducted correlational studies between groundwater and land subsidence using GRACE and InSAR in various locations, including London [23], Iran [24], the North China Plain [25], and Saudi Arabia [26]. These studies have confirmed the relationship between groundwater storage and land subsidence, observable through GRACE and InSAR. In Bangladesh, relevant studies primarily use groundwater well data and InSAR technology for qualitative analyses of specific regions. For example, studies examining the relationship between land subsidence and groundwater storage in Chapai Nawabganj have revealed a corresponding trend between changes in groundwater storage and land subsidence [27]. Other analyses focus on groundwater loading variations in Bangladesh [28]. Current research tends to focus on point-based subsidence analysis in select regions of Bangladesh and lacks direct large-scale analyses of the relationships between groundwater and land subsidence. These limitations hinder a comprehensive understanding of groundwater resources and land subsidence issues in Bangladesh.

This study aims to investigate the relationship between variations in groundwater storage and land subsidence in Bangladesh, providing a theoretical foundation for further research. Using the LiCSBAS package (v1.14.2), we conducted SBAS-InSAR analyses on Sentinel-1 satellite data to obtain land subsidence information in Bangladesh from 2014 to 2023. Concurrently, we acquired data on groundwater storage changes for the same period with GRACE and GLDAS. We extracted time-series curves from both datasets and used Dynamic Time Warping (DTW) similarity coefficients to assess the correlation between groundwater storage and land displacement variations. This work aims to provide a macro-level understanding of land–water interactions in Bangladesh and scientifically elucidate the patterns of changes in groundwater storage and land subsidence. This study offers a new, cost-effective research approach by utilizing freely available data to enhance understanding of the current state of groundwater storage and land subsidence in Bangladesh, providing foundational knowledge for future studies.

2. Materials and Methods

2.1. Study Area

Bangladesh is situated on the delta formed by the Ganges and Brahmaputra rivers in the northeastern part of the South Asian subcontinent. It is bordered by India to the east, west, and north; Myanmar to the southeast; and the Bay of Bengal to the south. Bangladesh has a subtropical monsoon climate, with annual precipitation ranging from about 1500 to 2000 mm in the west and 3000 to 4500 mm in the east [29]. The terrain is mostly flat, with an average elevation of 9 m. Covering a total area of 147,500 km², Bangladesh has a coastline of 550 km. The country is divided into eight administrative divisions, which include 64 districts. Bangladesh lies between latitudes 20°34′N and 26°38′N, and longitudes 88°01′E and 92°41′E. Figure 1 shows the geographical location of Bangladesh, with tectonic zoning adapted from the USGS’s global energy project [30].

The territory of Bangladesh is predominantly covered by a large amount of Quaternary sediments, while Neogene strata are mainly found in the eastern and northern border regions as well as in Chittagong. From the foothills to the coastline of the Bay of Bengal, the Quaternary sediments primarily consist of alluvial deposits, residual deposits, and deltaic sediments. The eastern fold zone of Bangladesh is part of the Indo-Burma fold belt, with a fold axis primarily oriented in a north–south direction [31]. Nearby Neogene strata are extensively exposed, and the Chittagong-Cox’s Bazar fault is located in the western part of Chittagong. In the central region, east of Jamalpur and Jessore, there is a depression completely covered by Quaternary sediments that generally trends northeastward, forming a deep-sea trough known as the Bangal Foredeep. North of Mymensingh lies the Shillong Plateau. The area north of Rangpur and Dinajpur is part of the Himalayan foreland basin, where late Cretaceous and Pleistocene molasse formations have developed extensively, with maximum thickness exceeding 2000 m. The western region is also predominantly covered by Quaternary sediments and features well-developed Pre-Cambrian and Gondwana formations [32]. Additionally, the Dauki Fault develops along the northern border, trending approximately east–west [33]. This fault is located at the southern margin of India’s Shillong Plateau and is characterized by uplift in the upper (northern) block and subsidence in the lower (southern) block, with relatively frequent Pleistocene activity [34,35,36].

Bangladesh is one of the most river-dense countries in the world, with three major river systems—the Ganges, Brahmaputra, and Meghna—merging here. The country has over 230 rivers of varying sizes. Surface water is abundant during the monsoon season, leading to seasonal floods. However, during the dry season, water scarcity becomes an issue, complicating irrigation, industrial, and other water needs. Consequently, much of the water during the dry season is sourced from groundwater. The groundwater in Bangladesh mainly consists of pore water found within the Quaternary sediments. The shallow aquifer (<70 m) is primarily composed of silt, clay, and fine sandstone. The intermediate aquifer (70–180 m) mainly includes fine sandstone and very fine sandstone. The deep aquifer (>180 m) consists of fine sandstone and coarse sandstone [29,37]. Groundwater levels are deeper in the northwest but shallower in the southern and low-lying areas [28]. In Bangladesh, aquifers are recharged by surface water during the dry season (November to March) and by precipitation from March to June. Due to the flat terrain and low hydraulic gradient, which averages around 0.0001, groundwater is rarely recharged during flood events. The primary discharge areas for groundwater are river valleys and lowlands [35].

2.2. Data Sources

2.2.1. GRACE and GLDAS

The GRACE data used in this study consist of mascons data provided by the Center for Space Research (CSR) [38], Jet Propulsion Laboratory (JPL) [39], and Goddard Space Flight Center (GSFC) [40] from 2014 to 2023. The processing of mascons data includes the removal of non-surface mass change signals, such as atmospheric and oceanic tidal effects, as well as the application of various filtering techniques to enhance signal quality. The final mascons product provides a global grid, with each grid cell containing a time series that displays mass changes in that region. To ensure data compatibility due to differing resolutions from the three institutions, cubic spline interpolation was employed to unify the data to a spatial resolution of 0.25°. Additionally, data from the Global Land Data Assimilation System (GLDAS) NOAH Land Surface Model for the years 2014–2023 was obtained, with a temporal resolution of one month and a spatial resolution of 0.25°. This dataset includes information on surface runoff, soil moisture, canopy water, and snow water.

2.2.2. InSAR Data

This study uses the ascending LiCSAR product data from 2014 to 2023. Each interferogram generated by the LiCSAR processor is composed of three sub-swaths based on the Interferometric Wide Swath (IWS) mode of Sentinel-1 data. Each sub-swath consists of 13 bursts corresponding to a 250 km × 250 km area, achieving a spatial resolution of 0.001°. The Sentinel-1A satellite was launched in April 2014, with a revisit cycle of 12 days. After the launch of Sentinel-1B in April 2016, the revisit cycle was reduced to 6 days. However, on 23 December 2021, Sentinel-1B experienced an anomaly related to the satellite platform’s instrument power supply, rendering it unable to provide radar data. As a result, the European Space Agency and the European Commission announced the termination of the Sentinel-1B mission. The LiCSAR data covers the entire country of Bangladesh and was processed using the LiCSBAS package. Specific information regarding the imagery is provided in

Table 1. The information from LiCSAR data for the study area.

Frame ID	Date		Period (Year)	Products
	Start	End
012A_06687_181919	2020.11.27	2023.04.30	2.4	1364
012A_06441_131313	2015.12.12	2020.07.24	4.6	386
114A_06391_131313	2014.10.25	2023.08.15	8.8	635
041A_06628_131313	2014.10.08	2023.11.26	9.1	1657
114A_06590_131313	2014.10.25	2023.12.01	9.1	1611
114A_06765_080913	2015.04.11	2023.06.28	8.2	1227
041A_06428_131313	2014.10.08	2023.11.26	9.1	699
041A_06828_131313	2014.10.08	2023.10.21	9	1848

.

2.2.3. GNSS

We used GNSS station data from the Nevada Geodetic Laboratory (NGL) [41] to validate the InSAR results. Currently, there are nine GNSS stations located within Bangladesh and near the subsidence areas. However, only five stations have data that cover the entire duration of this study with continuity. These are VKLA station in the Mymensingh district, DHA2 station in the Dhaka district, COML station in the Comilla district, BNTL station in the Khulna district, and BNGM station in the Rangpur district.

2.2.4. Precipitation

The 2014–2023 precipitation data were obtained from the Monitoring Version 2022 product of the Global Precipitation Climatology Centre (GPCC), which has a spatial resolution of 1°. To align with the GRACE data, we applied cubic spline interpolation to increase the resolution to 0.25°. Established in 1989 to meet the needs of the World Meteorological Organization (WMO), GPCC is operated by the Deutscher Wetterdienst (German Weather Service). The center provides precipitation datasets based on approximately 86,100 rain gauges and meteorological stations, primarily using data from the WMO and the Global Telecommunication System (GTS), including the SYNOP and CLIMAT databases. GPCC’s mission is to analyze global daily and monthly precipitation based on observed rainfall data, making it the largest precipitation database in the world, suitable for large-scale studies. These precipitation data were used in this study to investigate their relationship with groundwater and land subsidence.

2.2.5. Land Cover Data

The land cover data used were obtained from the Dynamic World dataset, published by Christopher F. Brown et al. in Nature Scientific Data, and can be accessed using the GEE tool [42]. This dataset is generated from Sentinel-2 L1C satellite data, starting from 27 June 2015, and utilizes deep learning to produce land cover types at a 10 m resolution. It categorizes land into nine types: water, trees, grass, flooded vegetation, crops, shrub and scrub, built, bare, and snow and ice. These data were used to analyze changes in groundwater storage and their relation to land subsidence. The details of all the data utilized in this study can be found in

Table 2. The information on all the data utilized in this study (comparison of spatial and temporal resolution before and after data processing).

Data	Data Source	Spatial Resolution		Temporal Resolution		Date
		Original	Processed	Original	Processed
GRACE	CSR	0.25°	0.25°	Monthly	Monthly	2014–2023
	JPL	0.5°	0.25°			2014–2023
	GSFC	0.5°	0.25°			2014–2023
GLDAS	GEE	0.25°	0.25°	Monthly	Monthly	2014–2023
InSAR	COMET-LiCS	~100 m	\	24 d/12 d	Monthly	2014–2023
GNSS	NGL	\	\	24 h	\	2018–2021
Land Cover	Sentinel-2 Dynamic World	10 m	\	2–5 d	Annual	2015–2023
Precipitation	CPCC	1°	0.25°	Monthly	Monthly	2014–2023

.

2.3. Methods

First, GRACE and GLDAS data were used to extract both spatial and temporal variations in groundwater storage. Subsequently, the LiCSBAS package was employed to obtain the spatial and temporal variations of land displacement rates in Bangladesh. The results were integrated using Dynamic Time Warping (DTW) to assess the similarity of their time series curves. Finally, this analysis revealed overarching patterns between changes in groundwater storage and land subsidence in Bangladesh. The detailed workflow is shown in Figure 2.

2.3.1. Obtaining Groundwater Storage via GRACE and GLDAS

In the absence of seismic activity or significant geological movements, the redistribution of surface mass is primarily influenced by the atmosphere and water. The Mascon data from both JPL and GSFC are generated using GRACE Level-1b inter-satellite distance and GPS data. For JPL’s Mascon data, we used the product processed through the Coastline Resolution Improvement (CRI) filter, while GSFC’s product employed a least squares estimator to calculate land values, ensuring the quality of each region. Both datasets ensure the integrity of terrestrial signals, effectively eliminating leakage effects. The CSR mascons are based on GSM Level-2 data, which have already had atmospheric and oceanic influences removed through relevant models prior to packaging. Therefore, it can be considered that mass changes represent changes in terrestrial water storage, expressed as equivalent water height [43,44].

According to the principle of water balance, the changes in terrestrial water storage ($\Delta TWS$) encompass various hydrological components, including surface runoff equivalent ($\Delta SW$), soil moisture storage ($\Delta SMS$), snow water equivalent storage ($\Delta SWES$), groundwater storage ($\Delta GWS$), and canopy water storage ($\Delta CANS$). Based on this water balance principle, the following water balance equation is derived:

$$\Delta TWS=\Delta SW+\Delta SMS+\Delta SWES+\Delta GWS+\Delta CANS$$

(1)

Thus, the groundwater storage ($\Delta GWS$) can be obtained from Equation (1):

$$\Delta GWS=\Delta TWS-\Delta SW-\Delta SMS-\Delta SWES-\Delta CANS$$

(2)

The total change in terrestrial water storage is derived using GRACE. $\Delta SW$, $\Delta SMS$, $\Delta CANS$, and ∆SWES are sourced from GLDAS. It is important to note that the mascons data provided by JPL, GSFC and CSR have had the average from 2004 to 2009 removed. Therefore, the same operation should be applied when processing the GLDAS data.

Due to technical issues and instrument failures, some months of GRACE data are missing. The GRACE satellite ended its mission in May 2017, while its successor, GRACE Follow-On, was not launched until June 2018, resulting in a one-year data gap. To ensure the completeness of the groundwater storage time series, we needed to supplement the missing data. We employed the Singular Spectrum Analysis (SSA) method, which is an improved version of the original SSA method and includes two cycles [45]. This method iteratively updates the missing values. When the estimated values of the missing data stabilize, the inner loop stops, while the outer loop gradually increases the complexity of the reconstructed sequence, controlled by the K-th modulus. The initial missing values are set to zero, and the updated values are retained for the next rounds of inner and outer loop cycles, ultimately yielding suitable values for the missing data.

As the resolution of GRACE data is 0.25°, the groundwater storage obtained at this scale reflects a large-area anomaly rather than the true groundwater storage at specific locations. Therefore, caution must be taken when interpreting groundwater storage results. The results derived from the processing of GRACE and GLDAS data should be viewed as representing the overall changes in groundwater storage for the region, serving as an indicator of groundwater storage change signals.

2.3.2. Monitoring Land Subsidence via SBAS-InSAR

LiCSBAS is an open-source InSAR time series analysis package developed by Yu Morishita et al., based on LiCSAR data. LiCSAR is an automated Sentinel-1 InSAR processing system. As of June 2024, approximately 1.69 million interferograms processed with GAMMA SAR software were available for free. When the study area is within the monitoring region of the LiCS project, users can directly access processed unwrapped interferograms, without the need for data preprocessing steps. This allows users to obtain InSAR time series analysis results while saving processing time and disk space. The LiCSBAS package utilizes free LiCSAR data to perform time series analysis based on the SBAS-InSAR method [18,46].

The principle of LiCSBAS starts with downloading LiCSAR products for the location of interest. It uses external Generic Atmospheric Correction Online Service (GACOS) products to correct atmospheric noise and applies overall interferometric quality checks and loop closure to eliminate errors in the interferogram. Factors that could degrade the results are identified and discarded. Interferograms that have undergone atmospheric correction and unwrapping error removal are then used for small baseline inversion (SBAS Inversion) to generate displacement time series and rates. Subsequently, the standard deviation (STD) of the estimated rates and a noise pixel masking index based on multiple noise sources are calculated. Finally, temporal and spatial filtering is applied to the time series to reduce residual noise, and the filtered time series and velocity results are exported.

The analysis method of LiCSBAS differs from general time series analysis methods in its processes for eliminating unwrapping phase errors and achieving loop closure. Since LiCSAR products are processed using GAMMA SAR software, phase unwrapping employs the SNAPHU method solely in the spatial domain, without considering phase information in the temporal dimension. Additionally, radar imaging effects, such as shadowing and overlapping, hinder phase continuity in some areas, resulting in unwrapping errors. If pixels with significant unwrapping phase errors are not removed, they will inevitably affect the accuracy of displacement calculations. In LiCSBAS, the check for unwrapping errors is performed on an interferogram-by-interferogram basis rather than on a pixel-by-pixel basis to calculate the root mean square error (RMSE) of the loop closure. This approach may allow unwrapping errors to remain in some pixels within interferogram pairs that pass the RMSE threshold. Therefore, an additional count is conducted for the number of pixels where the phase closure error exceeds the threshold. This count is then used as a criterion for masking [18].

We used GNSS station data to validate the InSAR results. Since the GNSS data obtained from the Nevada Geodetic Laboratory represent eastward, northward, and vertical displacements, while the InSAR results are in the line of sight (LOS) direction, a direct comparison is often not feasible. Therefore, we projected the GNSS data onto the LOS direction displacement:

$$L_{GNSS}^{\mathrm{los}}=S_x\times D_{\left(EW\right)}+S_y\times D_{\left(NS\right)}+S_{\mathrm{z}}\times D_{\left(UD\right)}$$

(3)

where $L_{GNSS}^{\mathrm{los}}$ represents the LOS direction displacement of GNSS. $D_{\left(EW\right)}$, $D_{\left(NS\right)}$ and $D_{\left(UD\right)}$ represent the east–west, north–south, and vertical displacements of GNSS. $S_x$, $S_y$ and $S_{\mathrm{z}}$ are the unit projection vectors of the InSAR. $\alpha$ is the satellite’s azimuth angle (measured clockwise from north), and $\theta$ is the radar incidence angle.$S_x=-\sin \theta \times \sin \left(\alpha -{\displaystyle \frac{3\pi }{2}}\right)$, $S_{\mathrm{y}}=-\sin \theta \times \cos \left(\alpha -{\displaystyle \frac{3\pi }{2}}\right)$, and $S_{\mathrm{z}}=\cos \theta$, all of which can be obtained from the respective frame’s E.geo, N.geo, and U.geo in liCSBAS.

2.3.3. Similarity Analysis Based on Dynamic Time Warping

Due to differences in groundwater depth, aquifer lithology, and other factors, the response of land displacement to changes in groundwater is often not immediate. Consequently, directly using correlation coefficients (such as Pearson’s) for evaluation may not obtain satisfactory results [47]. Dynamic Time Warping (DTW) is an algorithm based on dynamic programming that aligns two time series non-linearly in the time domain to accurately compute their similarity [48]. When comparing two time series, direct point-by-point comparisons often fail to provide an accurate similarity assessment, especially when there are temporal offsets or differences in speed between the two sequences. DTW employs dynamic programming to find an optimal alignment path that minimizes the distance between the aligned time series.

The principle of calculating the shortest distance using Dynamic Time Warping (DTW) involves defining two time series, $X$ and $Y$, with lengths $n$ and $m$, respectively. The process begins by computing the warping path $W=w_1,w_2,\dots ,w_k,\dots ,w_K$, where $\max \left(m,n\right)\le K\le m+n-1$. Each element $w_k$ takes the form of ${\left(i,j\right)}_k$, where $i$ represents the index in $X$ and $j$ represents the index in $Y$. The warping path $W$ must start at $w_1\left(1,1\right)$ and end at $w_K=\left(\mathrm{n},\mathrm{m}\right)$ to ensure that every index from both $X$ and $Y$ appears in $W$. Furthermore, the indices $i$ and $j$ in $W$ must be monotone increasing, which ensures that the connections between corresponding points in the two time series do not intersect. Monotone increasing means that [49]

$$w_k=\left(i,j\right),w_{k+1}=\left(i^{\prime },j^{\prime }\right)i\le i^{\prime }\le i+1,j\le j^{\prime }\le j+1$$

(4)

The final desired warping path is the one that minimizes the distance between the two time series:

$$D\left(i,j\right)=Dist\left(i,j\right)+\min \left[D\left(i-1,j\right),D\left(i,j-1\right),D\left(i-1,j-1\right)\right]$$

(5)

where the $Dist\left(i,j\right)$ represents the distance between $X\left(i\right)$ and $Y\left(j\right)$ (typically the Euclidean distance), while $D\left(i,j\right)$ denotes the accumulated distance matrix, which represents the shortest path to reach $\left(i,j\right)$, computed using dynamic programming [50].

The advantage of DTW lies in its effectiveness in aligning time series with temporal offsets and its robustness against noise and local mismatches [51]. Thus, it is well-suited for evaluating periodic data such as groundwater and land displacement. When using DTW to compute time series, the result is often the shortest distance. To make the similarity analysis more intuitive, the results are processed using normalization based on Euclidean distance:

$$S=1-{\displaystyle \frac{d}{L\times d_{\max }}}$$

(6)

where $S$ represents the similarity coefficient, which ranges from $[0,1]$. A value closer to 1 indicates greater similarity, while a value closer to 0 indicates greater differences. $d$ is the shortest distance computed by DTW. $L$ refers to the length of the sequence, and $d_{\max }$ is the maximum possible distance between each pair of elements.

In Similarity analysis, filtered and masked InSAR data are utilized to avoid selecting points with unwrapping errors when extracting time series curves. Meanwhile, GRACE data are provided on a monthly scale. To ensure compatibility and clarity, it is necessary to standardize the time series curves of land displacement to a monthly scale for comparison with groundwater storage.

3. Results

3.1. Groundwater Storage Change in Bangladesh

We utilized mascons product data from JPL, GSFC, and CSR. Due to the discrepancy in their original spatial resolutions, we applied cubic spline interpolation to standardize all three datasets to a resolution of 0.25°. Based on the water balance equation, we derived the groundwater storage variation in Bangladesh. The spatial distribution (Figure 3) and time series curve (Figure 4) of groundwater storage changes were obtained by averaging the data from the three institutions. Additionally, we extracted the annual mean variation and the long-term trend of groundwater storage.

Figure 3a illustrates the spatial distribution of multi-year groundwater storage variations in Bangladesh from 2014 to 2023. The northwestern regions, including Rangpur, Rajshahi, and Mymensingh, show significant signals of groundwater depletion, with Rangpur experiencing the most severe reduction at a rate of −10 to −15 mm/year. In contrast, coastal areas near the Bay of Bengal display an upward trend in groundwater storage variations.

Figure 3b presents the trend of total surface water storage in Bangladesh based on GLDAS. There is a gradual decrease from southwest to northeast, with surface water storage in the northeastern part of Rangpur showing a declining trend, while the surface water storage in the southwestern region of Khulna is on the rise.

The precipitation distribution in Bangladesh is characterized by more precipitation in the east and less in the west. Figure 3c indicates that the southern parts of Sylhet and Chittagong receive more precipitation, whereas Rajshahi, the northwest part of Khulna, and the western part of Dhaka receive the least precipitation.

Figure 4 presents the long-term changes in groundwater in Bangladesh extracted from GRACE. A comprehensive comparison of mascons data from the CSR, GSFC, and JPL institutions shows that the original GRACE total water storage (TWS) for the Bangladesh region is largely consistent (Figure 4a), with values from GSFC being slightly lower than those from the other two institutions. The long-term groundwater results indicate that, over the years, groundwater in Bangladesh is decreasing at an overall rate of −5.5 mm/year (Figure 4d). Observable seasonal variations in groundwater are evident, with peak amplitudes typically occurring in June and the lowest amplitudes generally observed in October. Meanwhile, soil moisture, canopy water, surface runoff, and snow water in Bangladesh have remained relatively stable over the years (Figure 4b).

3.2. Changes in Land Displacement in Bangladesh

In processing the data using LiCSBAS, this study applied Generic Atmospheric Correction Online Service (GACOS) data to correct the atmosphere for all interferograms utilized, using phase standard deviation (STD) as the evaluation metric for atmospheric correction. Taking 012A_06441_131313 as an example, a total of 371 pairs of interferograms were generated. After atmospheric correction, the phase standard deviation was reduced in most cases, with a maximum reduction of approximately 65%. Subsequently, an overall quality check was conducted on the interferograms used, setting the average coherence threshold at 0.05 and the unwrapped coverage threshold at 0.3, excluding the interferograms that failed the check. Interferograms meeting the quality criteria were further subjected to a loop closure phase test. To ensure temporal continuity, the threshold was set to fluctuate between 1.5 and 3 rad, adjusted according to the quality of the interferograms.

Figure 5 shows the land subsidence situation in Bangladesh from 2014 to 2023. Some areas have unwrapping errors, resulting in missing information, such as the southern part of Chittagong. However, the subsidence information in those regions is still quite evident. It also shows the correlation between areas experiencing land subsidence in Bangladesh and regions with deeper groundwater levels, particularly in Dhaka and Mymensingh. Based on the groundwater depth in Bangladesh, the subsidence areas are divided into six subsidence zones for further analysis. The reported deformations in this study reflect the Line of Sight (LOS) displacement. Overall, land subsidence mainly occurs in urban areas and farmlands. In delta plains, such as Khulna and Barisal, there have been extensive subsidence incidents, with considerable subsidence rates, the maximum reaching −389 mm/year.

Zone I is located in the central part of Rangpur, which has an average land displacement rate of −4.7 mm/year. This area is a major rice-growing region, with displacement mainly occurring in agricultural fields, while more significant subsidence is observed in urban areas.

Zone II is located in the central part of Mymensingh, which has an average land displacement rate of −21.6 mm/year, featuring three distinct subsidence areas. A linear subsidence zone was observed in northern Mymensingh, aligning with the spatial position of the Dauki fault located near the northern border of Bangladesh. Given the focus of this study on the relationship between subsidence and groundwater, subsiding areas within the region were selected for further investigation to avoid interference from tectonic movements.

Zone III is located in the east part of Rajshahi, which shows an average land displacement rate of −20.6 mm/year. Over the study period, land subsidence in this area was more pronounced compared to the others. The water system in this subsidence area is densely distributed near the Ganges River, Mahananda River, and Pagla River.

Zone IV is located in the south part of Dhaka, which has an average land displacement rate of −15.4 mm/year, with displacement predominantly occurring within Dhaka city.

Zone V is located in the north part of Chittagong, which exhibits an average land displacement rate of −14.9 mm/year, with severer subsidence mainly occurring near Cumilla city.

Zone VI is located in the central part of Khulna, which shows an average land displacement rate of −12.2 mm/year, with subsidence primarily occurring in urban and agricultural areas.

Figure 6 illustrates the time series of cumulative displacement in the line-of-sight (LOS) direction, comparing InSAR results with GNSS station data. The findings reveal that the overall trend of the InSAR displacement curve closely aligns with that of the GNSS displacement curve. Further analysis was conducted by aligning the GNSS and InSAR data with temporal and spatial references. To assess the reliability of the results, we projected the GNSS cumulative displacement onto the LOS direction and compared it with the cumulative displacement from InSAR, calculating the root mean square error (RMSE,

Table 3. Comparison between InSAR and GNSS.

GNSS Station	InSAR	GNSS	Difference (mm/year)	RMSE (mm)
VLKA (Mymensingh)	−12.48	−11.43	−1.05	2.92
DHA2 (Dhaka)	−6.21	−15.22	9.01	9.69
COML (Comilla)	−6.93	−3.41	−3.52	7.31
BNTL (Khulna)	−11.92	−7.9	−4.02	14.06
BNGM (Rangpur)	−1.05	−3.16	2.11	3.44

). The results showed that the GNSS data at the VLKA station were the most consistent with the InSAR results, with a difference of −1.05 mm/year and an RMSE of 2.92 mm. In contrast, the BNTL station exhibited the largest discrepancy, with a difference of −4.02 mm/year and an RMSE of 14.06 mm. Due to the BNTL station’s location in a coastal mangrove area, coherence was relatively poor, and some data were missing at this site (Figure 6d), for which NGL provided no explanation. Excluding the BNTL station, the RMSE values for all other stations were below 1 cm, indicating a high degree of reliability in these results.

4. Discussion

4.1. Impact of Precipitation on Groundwater Storage

Precipitation primarily affects groundwater and land subsidence by influencing the recharge amounts of groundwater [52,53]. Figure 7a illustrates the precipitation data from 2014 to 2023. The original spatial resolution of the precipitation data was 1°. To make it comparable with the GRACE groundwater storage data, we refined it to 0.25° using cubic spline interpolation. Both datasets have a monthly temporal resolution, allowing for direct comparison. We removed the long-term trend from the precipitation data (Figure 7b) to emphasize seasonal variations, which were then compared with the groundwater storage data. Bangladesh has a subtropical monsoon climate, with precipitation typically occurring from June to September, displaying distinct seasonal variations. This precipitation is generally sourced from moist monsoons brought in by weak tropical depressions from the Bay of Bengal [54,55]. In recent years, there has been a declining trend in winter precipitation in Bangladesh, consistent with previous research [56].

In conjunction with the detrended precipitation data (Figure 7b), it is evident that the response of groundwater to precipitation exhibits a significant lag, with groundwater peaks generally occurring 1–2 months after precipitation. This indirectly confirms that precipitation is a major source of groundwater recharge in Bangladesh. Consequently, the main recharge period occurs from June to September, typically peaking in June or July.

Spatial distribution analysis of precipitation and total surface water storage indicates that Sylhet receives abundant precipitation, resulting in adequate year-round groundwater recharge, with no significant subsidence zones observed. In contrast, Rangpur shows clear signs of groundwater depletion, with lower total surface water storage and precipitation. Furthermore, due to the flat terrain and low hydraulic gradients in Bangladesh, it is challenging for groundwater to receive surface water recharge. This results in even less recharge from surface water and precipitation in Rangpur than in other regions, potentially contributing to groundwater depletion in that area. The highest annual precipitation and the most significant fluctuations of groundwater storage in Bangladesh were observed in 2016.

During the study period, four cyclones that affected precipitation occurred (indicated by yellow dotted lines in Figure 7b: Cyclone Fani in May 2019, Cyclone Amphan in May 2020, Cyclone YAAS on 26 May 2021, and the “Great Depression” in the Bay of Bengal in September 2022). Each event was accompanied by increased precipitation, resulting in a rise in groundwater storage compared to the previous month. Coastal areas are more affected by tropical cyclone storms than inland areas [56]. Thus, during the peak rainy season, the coastal region of Khulna receives preferential groundwater recharge.

4.2. Similarity Analysis of Groundwater and Land Subsidence

The similarity analysis was not conducted on all points but rather on a carefully selected set of representative locations. These points were selected based on several critical criteria. Firstly, we prioritized areas where the InSAR land displacement data exhibited both continuity and high quality over the time series, as a clear and complete displacement record is crucial for precise similarity analysis. Secondly, we focused on regions with deeper groundwater levels and urban areas, where subsidence is more strongly correlated with groundwater extraction, producing land displacement time series that more distinctly capture changes in groundwater storage. Lastly, through preliminary assessments, we identified points where the time series curves of land displacement and groundwater storage exhibited a strong trend correlation, while avoiding areas influenced by other natural disturbances, such as tectonic activity or flooding. These points best reflect the dynamic relationship between groundwater extraction and land subsidence. Due to the inconsistency in the temporal resolution between land displacement and groundwater storage, both were unified into monthly data for further comparison.

We obtained land cover data for 2015–2023 through Google Earth Engine and compared it with areas of land subsidence. Based on the representative points selected in the previous section, we extracted the land cover changes at those points from 2015 to 2023 and integrated these results into the time series curves of groundwater storage and land displacement, making the analysis more intuitive for further investigation. Taking the Mymensingh District as an example, Figure 8 shows the comparison between the multi-year land cover changes and land displacement at a representative point in the zone (II). In 2016, the land cover type at the representative point was categorized as trees, but with rapid urbanization, the tree cover gradually diminished and was replaced by crops. Meanwhile, buildings proliferated, and by 2023, the amount of cropland had further decreased.

Figure 9 shows the time series curves of groundwater storage changes and cumulative land displacement changes in the six previously mentioned subsidence zones. It is evident that the overall trend of groundwater storage variation in Bangladesh aligns with land deformation trends in different regions. The similarity calculated using Dynamic Time Warping (DTW) exceeds 0.85, indicating a strong correlation between groundwater storage and land subsidence from 2014 to 2023. This means that as groundwater decreases, land deformation also declines, while the deformation rate slows when groundwater levels rise. The average lag time between changes in groundwater storage and surface deformation is approximately 2 to 6 months. Detailed data can be found in

Table 4. Similarity coefficients and lag time between groundwater and land subsidence in subsidence areas.

Area	I	II	III	IV	V	VI
Similarity coefficient	0.8705	0.9171	0.9055	0.8507	0.8997	0.8668
Lag time (month)	2.25	5.24	2.14	5.47	5.54	2.24

.

In Figure 9, the land cover types in the subsidence zones of (I) Rangpur, (II) Mymensingh, (III) Rajshahi, and (VI) Khulna were primarily crops from 2014 to 2023. The average lag time of land displacement in relation to changes in groundwater storage in Rangpur, Rajshahi, and Khulna was approximately 2 months. The same land cover type demonstrates consistent responsiveness to changes in groundwater storage. From 2015 to 2017, the land cover type in the subsidence zone of (II) Mymensingh was predominantly trees, with no significant land subsidence observed during this period. Land subsidence began once the area was converted to crops.

According to the information shown in Figure 5 and geological data [30], the strata in zone (I) primarily consist of soil and Quaternary alluvial deposits, while zone (II) mainly comprises alluvial sediments. Zone (III) contains both residual and alluvial deposits, and there is a dense surface water system in the area. The response of land displacement to groundwater storage changes in this area is the quickest, likely due to the abundance of surface water systems, which leads to more stable groundwater dynamics. As a result, land displacement is more closely aligned with variations in groundwater. Zone (VI) includes deltaic sediments and alluvial silt deposits. The groundwater in these areas is predominantly shallow pore water. During the monsoon season, precipitation replenishes the groundwater. In the dry season (December to February), the lack of precipitation makes groundwater the primary irrigation source. The time series curves show clear seasonal variations in land displacement: during the monsoon, the displacement rate slows with groundwater recharge, while in the dry season, it increases as groundwater levels diminish. Additionally, factors such as silt deposits affect the infiltration rate of precipitation, resulting in a lag effect of land displacement relative to changes in groundwater.

In Figure 9, the land cover type in the subsidence zones of (IV) Dhaka and (V) Chittagong were built from 2014 to 2023. The average lag time between land displacement and changes in groundwater storage was approximately 5 months. Human activity and urban land consolidation extend the time required for groundwater recharge compared to other areas. According to the information shown in Figure 5 and geological data, subsidence in zones (IV) and (V) occurred in urban settings.

The primary lithology in zone (IV) consists of residual and paludal deposits. Excessive extraction of groundwater in Dhaka and its surrounding areas has resulted in the slow drainage and compaction of paludal sediments, leading to subsidence [13]. Land subsidence initially occurred slowly, but by 2021, the rate of subsidence began to increase significantly. It is speculated that this is due to the rising population density in Dhaka in recent years [57], leading to a substantial increase in the number of buildings. The presence of numerous buildings often results in soil consolidation. Additionally, the pumping of concrete during construction greatly reduces surface infiltration in the city, which in turn affects groundwater recharge. According to the theory of effective stress, as surface loads increase, a reduction in groundwater means that greater pressure is borne by the soil skeleton. Once this pressure exceeds the soil skeleton’s capacity, it can accelerate land subsidence. The time series curve of deformation illustrates this subsidence process.

In zone (V), the dominant lithology comprises alluvial and residual deposits, with significant subsidence occurring in Comilla city and its surroundings. From 2016 to 2018, the land cover type was trees, and the rate of land displacement slowed as groundwater storage increased, demonstrating clear seasonal variations and a rapid response to groundwater changes. However, after the land cover was changed to built-up areas, the delay in the response of land displacement to groundwater changes began to increase. Therefore, it is inferred that land cover primarily affects land subsidence by controlling the timing of groundwater recharge.

5. Conclusions

Using GRACE and GLDAS to analyze changes in groundwater storage in Bangladesh, along with LiCSBAS to monitor large-scale land subsidence, the following conclusions were drawn:

(1) From 2014 to 2023, overall groundwater storage in Bangladesh decreased at a rate of −5.5 mm/year. Groundwater depletion is primarily concentrated in Rangpur, Mymensingh, and Rajshahi.

(2) Land subsidence occurs mainly in areas where the rate of groundwater storage decline is significant and often overlaps with regions where groundwater is found at greater depths.

(3) Precipitation changes have been relatively stable, but there is significant variability within the year. Changes in groundwater are largely influenced by variations in precipitation, supporting the notion that precipitation is one of the primary sources of groundwater recharge in Bangladesh. Moreover, using the Dynamic Time Warping (DTW) method, the similarity coefficient between the time series of surface deformation and groundwater storage variations was found to be above 0.85. The response of surface deformation to groundwater changes occurs within a timeframe of 2 to 6 months, confirming a correlation between land subsidence and groundwater from a macro perspective in Bangladesh.

This study explored the feasibility of combining GRACE and InSAR technologies in Bangladesh. It is similarly suitable for related research in other underdeveloped regions, providing preliminary validation before further studies. However, several limitations were identified regarding this research. The data provided by GRACE satellites have low spatial resolution, making it difficult to accurately capture dynamic changes in groundwater at smaller scales. Additionally, while InSAR technology provides high-resolution surface deformation information, signal attenuation or distortion may occur in cloudy, rainy, or densely vegetated areas, impacting data quality and continuity. Therefore, future research could focus on extracting higher resolution groundwater change data and exploring the intrinsic mechanisms linking groundwater storage changes to land subsidence, thereby deepening the understanding of the relationship between groundwater dynamics and land subsidence in Bangladesh. Moreover, since LiCSBAS utilizes pre-processed interferograms, the merging of overlapping areas between orbits becomes a notable issue. Current work primarily focuses on along-track merging; however, when the duration of data acquisition from adjacent tracks varies, resulting in discontinuities, cross-track merging becomes especially critical. Consequently, conducting further research on cross-track merging based on LiCSBAS results could greatly enhance the accuracy of wide-area InSAR in the future.