Ground subsidence evolution and groundwater variation in the middle route of south to north water diversion project

Abstract

The South-to-North Water Diversion Middle Route Project (SNWD-MRP) traverses two major agricultural production areas and populous regions in northern China, from Henan to Beijing-Tianjin-Hebei (BTH). It is of profound significance in alleviating water resource scarcity and improving the local ecological environment in these areas. This paper introduces a wide-area subsidence monitoring technology that combines the FS-InSAR method with ground measurement data correction. Based on this, a correlation analysis is conducted using measured groundwater level data along the route, exploring the intrinsic link between groundwater level fluctuations and uneven subsidence dynamics. Finally, a machine learning algorithm is employed to quantify the impact of groundwater changes on ground subsidence. The results indicate a high correlation between subsidence and groundwater level fluctuations (R: 0.6 ~ 0.9). The machine learning training results show that changes in groundwater storage are the main factor affecting subsidence, accounting for 75.1%. Further analyzing the reasons, groundwater over-exploitation triggers a drop in the groundwater level, which leads to a decrease in the pore pressure and an increase in the effective stress in the sand layer, and secondary consolidation and subsidence in the clay layer due to the drainage lag, whereas SNWD-MRP contributes to the rebound of the groundwater level through the reduction of groundwater extraction and ecological water recharge, which reduces the pressure release from the aquifer, and leads to the stabilization of the compression process of the strata. This study indicates that groundwater losses remain the primary controlling factor of subsidence. After the SNWD’s initiation, subsidence in the Beijing and Tianjin areas has noticeably slowed and stabilized. Overall, this major water project has effectively mitigated uneven subsidence, underscoring its positive role in regional ground stabilization and environmental improvement.

Introduction

For decades, the North China Plain has grappled with persistent water scarcity challenges, attributed to inadequate spatial allocation of water resources, limited rainfall, and burgeoning demand from urban, agricultural, and industrial sectors^1,2,3,4. In response to this challenge, China has proactively undertaken the construction of the South-to-North Water Diversion (SNWD) Project, aimed at transferring abundant water resources from the humid southern regions to the arid northern regions. As a significant engineering initiative to optimize water resource allocation in China, the SNWD project embodies the earnest expectations of the Chinese people for alleviating water scarcity in the Beijing-Tianjin-Hebei (BTH) region, improving urban water supply quality, and conserving water sources. However, since the Henan and BTH segments traversed by the SNWD-MRP are located in major crop-producing areas and population centers in northern China, the demand for agricultural and domestic water use in this region is exceptionally high. Under the combined influence of local geological structures and groundwater over-exploitation, significant uneven subsidence and localized subsidence funnels are observed along certain segments of the route. Therefore, conducting long-term subsidence monitoring in the areas along the route holds crucial significance and reference value for evaluating the effectiveness of the SNWD project implementation.

In the field of wide-area subsidence monitoring, various methods are employed, including ground-based measurements and satellite InSAR monitoring. Traditional ground-based subsidence measurement methods offer high observation accuracy but require significant manpower, resources, and long-term maintenance of monitoring points to ensure integrity. Moreover, traditional methods struggle to comprehensively capture the spatial distribution patterns and subsidence trends of the entire SNWD-MRP at a macro level. InSAR measurement technology, as a significant branch of radar remote sensing, possesses advantages such as wide coverage, rapid data acquisition, all-weather and all-day monitoring capabilities, and high precision⁵making it widely applicable across various domains^6,7,8,9. However, conventional Differential InSAR (D-InSAR) technology suffers from limitations in monitoring accuracy due to issues such as spatial-temporal decorrelation and atmospheric delay. Time-Series InSAR^10,11 has emerged to overcome these limitations, significantly enhancing subsidence monitoring accuracy and demonstrating great potential in detecting long-term accumulated slow subsidence^12,13. Zhang Shuangcheng¹⁴ et al. utilized SBAS-InSAR technology to extract subsidence information along the entire Middle Route of the SNWD project from April 2015 to November 2020. The results showed that ground subsidence along the route was mainly distributed in the southeastern part of Hebei Province, with two subsidence areas along the Tianjin branch. Ling Jun¹⁵ et al. observed the Xichuan County section of the SNWD using SBAS-InSAR technology, finding minor subsidences at the Taocha outlet of the Danjiangkou Reservoir and in the Zhanggou and Ligou areas from May 2019 to May 2020. Wang Nan¹⁶ et al. monitored the canal within the Henan Province of the SNWD Middle Route using PS technology. They discovered a subsidence funnel in the Yuzhou-Changge canal section with a maximum land subsidence rate of 20 mm/yr from 2017 to 2019. Furthermore, through comparison with leveling measurement results, the effectiveness of InSAR technology in large-scale water conservancy project monitoring was validated. Yan Xia¹⁷ et al. employed Time-Series InSAR technology to extract subsidence information from the Beijing Plain area from January 2017 to December 2019. The maximum annual subsidence rates in Chaoyang District and Tongzhou District were 115 mm/yr and 120 mm/yr, respectively. Compared to the period before water diversion from the SNWD, significant subsidence rates in Beijing’s notable subsidence areas have been somewhat alleviated. Zhang et al. detected the hydraulic and mechanical properties of hydrogeologic units at different depths by using stress-strain analyses and oedometer tests based on the data from leveling surveys, borehole extensometers, and multilayer monitoring of groundwater levels¹⁸. These results can help to improve the assessment and the prediction of land subsidence. SBAS InSAR has been applied in Xi’an, Shaanxi to study the relationships between land subsidence and fissures¹⁹. They detected four main land subsidence zones in the Xi’an Hi-tech Zone, Xi’an Qujiang New Zone, Chanba River Economic Zone, and Hujiamiao, with an average land subsidence rate of about 50 mm/year from 2005 to 2012. Good correlations between the ground water-level decline and the land subsidence were found in their studies. Meanwhile, Zhu et al. applied PSI on the ENVISAT ASAR images from 2003 to 2010 to investigate the land subsidence in the northern Beijing plain²⁰. They reported that the maximum deformation rate reached 52 mm/year, with a cumulative maximum sinking 342 mm in Houshayu, located in the southwestern part of the Beijing plain. Luo et al. analyzed the historical and current data of the land subsidence in the Beijing plain to reveal that the land subsidence center moved from the west to the east and also moved from the north to the south with the increased amount of land subsidence²¹.

However, when applying time-series InSAR for surface deformation monitoring in areas with large topographic relief and lush vegetation cover, there are problems of interference from atmospheric delay effects, insufficient monitoring density, temporal decoherence, and signal destabilization²². To address the above problems, the full-scatterer synthetic aperture radar interferometry (FS-InSAR) technique provides a new technical means for ground subsidence monitoring by innovatively fusing the phase information of permanent scatterers (PS) and distributed scatterers (DS)²³. With the advanced scatterer identification algorithm, FS-InSAR can realize high-density coverage of thousands of monitoring points per square kilometer, which greatly improves the spatial resolution of subsidence monitoring²⁴. The unique dual-scale time-domain low-pass filtering technology of FS-InSAR effectively suppresses the atmospheric phase and noise interference and maintains millimeter-level monitoring accuracy in low-coherence areas such as vegetation cover. FS-InSAR has excellent FS-InSAR has excellent time series analysis capability, which can accurately capture the non-linear evolution of ground subsidence and provide a reliable basis for the study of subsidence mechanism and disaster early warning.

In the previous studies, most of them are limited to the analysis and discussion of some cities or sections of the Central Route Project and have not been analyzed in depth for the settlement evolution and settlement control factors after the opening of the South-to-North Water Diversion Project. For this reason, this paper firstly extracts the time series surface deformation results of the whole South-to-North Water Diversion Project from 2017 to 2021 by using FS-InSAR technology combined with the measured level correction method. Subsequently, correlation analysis is carried out in combination with the measured groundwater level data to further reveal the spatial and temporal evolution of ground subsidence and groundwater recovery in the cities along and around the opening of the South-to-North Water Diversion Middle Line Project. Finally, in order to quantify the influence of groundwater on ground subsidence, population density, rainfall, NDVI, and groundwater storage changes are introduced as subsidence-inducing factors, and the contribution of each inducing factor is calculated by the machine learning algorithm. The output of this study will not only provide scientific data references for future water resource management and ground subsidence prevention, but will also help to better understand the impact of the South-to-North Water Diversion Project on the regional surface stability, and provide support for research and policy formulation related to sustainable development and ecological balance.

Materials and methods

The MRP is designed to traverse remote areas, following the piedmont plains of the Taihang Mountains. However, dense vegetation in these regions often leads to challenges such as the absence of monitoring points during interferometric processing. To address this, this paper introduces the FS-InSAR technology proposed by Wu and others²⁵ and combines it with the technical approach of overall correction of long strip and adjacent track partition monitoring results based on ground-level measurement data, proposing a multi-track long-strip time-series InSAR monitoring method tailored for precise extraction of uneven ground subsidence over a wide area and long-time period. Firstly, FS-InSAR time-series ground subsidence information extraction is carried out separately for different Sentinel-1 A images; subsequently, the InSAR results are corrected using ground-level measurement data within the MRP corridor, and then the overall ground subsidence information of the SNWD-MRP is obtained by stitching together multi-track long-strip data²⁶. Finally, combined analysis with measured groundwater level data is conducted to comprehensively analyze the evolving trend of ground subsidence after the operation of the MRP. Figure 1 is illustrated in the following diagram.

Fig. 1

MTLS-InSAR technology flowchart.

Study area

As shown in Fig. 2, the Middle Route Project of the SNWD spans approximately 1432 km. It originates from the eastern bank of the Danjiangkou Reservoir, passing through six cities in Henan Province and four cities in Hebei Province, before ultimately reaching Beijing and Tianjin. These cities hold significant positions in China’s economic and social development; however, due to geographical constraints, they face a relative scarcity of water resources. These cities play crucial roles in China’s economic and social advancement; nevertheless, they encounter challenges due to limited water resources caused by geographical constraints. The implementation of the SNWD-MRP ensures a steady water supply for Beijing and Tianjin, fulfilling the requirements of urban expansion and residents’ welfare. Moreover, it delivers consistent and equitable water support for agricultural and industrial activities in Hebei Province, thereby fostering economic growth. Additionally, it contributes to enhancing the ecological landscape in cities like Chengde and Langfang, rejuvenating the tourism sector and environmental conservation efforts. By July 2023, the initial phase of the SNWD-MRP had redirected over 57 billion cubic meters of water northward, benefiting 24 major and medium-sized urban centers as well as more than 130 counties and cities along its path. This initiative directly impacts a population exceeding 85 million. To delineate the study area, this paper selects a geographical range spanning from 108°E to 120°E in longitude and from 31°N to 41.5°N in latitude, aligning with the spatial distribution of the SNWD-MRP.

Fig. 2

Study area coverage map. (GMT, Version 6.5; https://www.generic-mapping-tools.org/).

Fig. 3

Median Corridor Elevation Profile. (Origin, Version2019; https://www.originlab.com/).

Elevation profiles were drawn based on the centerline corridor route, as shown in Fig. 3. Figure 3(a) represents the topographic change from Danjiangkou Reservoir in Henan Province to Beijing City, and Fig. 3(b) represents the elevation change from Danjiangkou in Henan Province to Tianjin City. Comprehensive coverage and elevation profiles of the study area show that: in terms of topography and geomorphology, the terrain in the study area is undulating, with a general trend of high in the south and low in the north, high in the west and low in the east. The western part of the cities along the route is higher, with greater relief, and is dominated by mountains and hills, while the eastern part is lower, relatively flat, and dominated by plains and hills.

Data resource

This study utilizes SAR data from the Sentinel-1 A satellite to investigate the spatial distribution of ground subsidence. We acquired a total of 1209 ascending orbit Sentinel-1 A radar images from 2017 to 2021. Specific details are provided in Table 1. To mitigate errors caused by terrain undulations in differential interferometric measurements, we introduced a 30 m resolution SRTM digital elevation model²². This approach enhances measurement accuracy and provides more precise spatial distribution information of ground subsidence.

Table 1 Sentinel-1 A data parameters furthermore, this study selected a total of 20 groundwater wells near the corridor of the SNWD-MRP and adjacent to severely subsided areas to conduct correlation analysis with InSAR-measured ground subsidence. The locations and Spatial distribution of the groundwater wells are provided in Table 2 (Doi: https://doi.org/10.17632/nmw536gnws.1).

Table 2 Measured groundwater well-specific information.

Among them, SY, CY, HD, TZ, and CP represent Shunyi, Chaoyang, Haidian, Tongzhou, and Changping districts in Beijing respectively, LF represents Langfang, TJ represents Tianjin and CZ represents Cangzhou. HDA represents Handan, and SZ represents Shenzhou.

FS-InSAR technology

The core idea of the FS-InSAR technique is to improve the density of InSAR monitoring points by utilizing dual-scale time-domain low-pass filtering and SAR average amplitude thresholding based on obtaining the de-entangled phase at a certain moment for small baseline differential interferometric image pairs²⁵. Since only the subsidence phase predominantly exhibits low-frequency characteristics in the differential interferometric phase, a dual-scale time-domain low-pass filtering method is employed to separate the low-frequency phase in the time domain. This involves utilizing small-scale time-domain low-pass filtering to preserve the subsidence phase to the greatest extent, subtracting the results of the small-scale time-domain low-pass filtering from the time-series phase, and subsequently conducting large-scale time-domain low-pass filtering to extract the residual subsidence phase. The filtering formula is as follows.

$${\phi _{lpf}}({t_m})=\sum\limits_{{z=1}}^{N} {\phi ({t_z})G(} {t_z} - {t_m})$$

(1)

$$G({t_z} - {t_m})=\frac{1}{{\sqrt {2\pi } W}}{e^{\frac{{ - {{({t_z} - {t_m})}^2}}}{{2{W^2}}}}}.$$

(2)

where, ${\phi _{lpf}}({t_m})$ represents the low-pass filtered phase at the output time(${t_m}$), $\phi ({t_z})$ is the temporal phase, $G( \cdot )$ is the Gaussian function, and W represents the small-scale/large-scale time window size.

Joint leveling calibration of MTLS-InSAR technology

The main idea of this method is to correct the FS-InSAR results by combining the measured level data first and then stitching the data of neighboring strips and tracks. Dividing the ground subsidence results by the cosine of the incidence angle yields the vertical subsidence results. This paper introduces ground-level measured elevation data for the correction of vertical subsidence: Taking benchmark points as the reference, within a certain distance range (80 m), the nearest InSAR point targets are selected. The subsidence values of the selected InSAR point targets are compared with their corresponding elevation data, and the differences between the elevation data and InSAR measurements are calculated. Subsequently, we construct a polynomial fitting equation and solve for the polynomial coefficients using the least squares method. Finally, the obtained polynomial coefficients are used to correct the initial subsidence results from InSAR. The polynomial fitting equation is given in the following equation:

$$f(x,y)={a_1}{y^2}+{a_2}{x^2}+{a_3}xy+{a_4}x+{a_5}y+{a_0}={v_{InSAR}} - {v_{level}}.$$

(3)

Where $\:f(x,y)$ denotes the correction value, $\:x,y$ denotes the horizontal and vertical coordinate values under the geodetic coordinate system of the InSAR monitoring point, $\:{v}_{InSAR}$ denotes the subsidence rate of InSAR, $\:{v}_{level}$ denotes the subsidence rate of the leveling point, and $\:{a}_{0}\sim{a}_{5}$ denotes the polynomial coefficients, respectively.

The polynomial coefficients obtained by solving Eq. (3) are used to correct the InSAR monitoring results:

$${V_{correct}}={V_{up}} - f(x,y).$$

(4)

where ${V_{up}}$ represents the vertical subsidence rate and ${V_{correct}}$ represents the corrected subsidence rate.

Subsequently, data splicing of neighboring strips and tracks is performed based on the corrected InSAR subsidence results.

Due to the different reference benchmarks for time-series InSAR processing between different strips of the same orbit and the uncertain phase stability of the reference target points, it is necessary to unify the reference benchmarks when splicing consecutive strips of the same orbit²⁷. Assuming there are three adjacent swaths within the same orbit, we selected the middle image as the reference master image, then the overall deviation of co-located coherent points in the overlapping regions of adjacent images is calculated. Subsequently, the auxiliary images are corrected based on the overall deviation. The specific formula is as follows.

$$\Delta {d_i}=\frac{1}{N}\sum\limits_{{i=1}}^{N} {({d_{{m_i}}} - {d_{{s_i}}})}$$

(5)

$${\bar {d}_s}={d_s}+\Delta {d_i}.$$

(6)

where $\:\varDelta\:{d}_{i}$ denotes the overall deviation from the main image, N represents the number of points in the overlapping region of the main and auxiliary images, $\:{d}_{{m}_{i}}$,$\:{d}_{{s}_{i}}$ represents the subsidence rate of the coherent target i in the overlapping region of the main and auxiliary images respectively, $\overline{d} _{s}$represents the subsidence rate of all the coherent targets in the corrected auxiliary image, and $\:{d}_{s}$ represents the subsidence rate of all the coherent points in the auxiliary image.

The essence of the multi-track integration approach is to accurately stitch together the subsidence parameters of coherent targets obtained from different orbits (subsidence rate, subsidence, and elevation correction parameters obtained by time series analysis). The stitching process mainly consists of two steps:

Firstly, unify the coordinate systems of adjacent orbits. The radar intensity images are geocoded to transform the coordinates of the subsidence parameters. Subsequently, polynomial regression is applied to precisely register the main and auxiliary orbit images. The registration function is represented as follows:

$$\left\{ {\begin{array}{*{20}{c}} {\Delta x(x,y)=\sum\limits_{{i=0}}^{l} {\sum\limits_{{j=0}}^{i} {{a_{i - j,j}}{x^{i - j}}{y^j}} } } \\ {\Delta y(x,y)=\sum\limits_{{i=0}}^{l} {\sum\limits_{{j=0}}^{i} {{b_{i - j,j}}{x^{i - j}}{y^j}} } } \end{array}} \right.$$

(7)

Where $\Delta x$, $\Delta y$ represents the image pair offsets in the x and y directions of the primary and secondary images, l is the polynomial order, and a, b are the polynomial coefficients.

Next, unify the reference baselines of adjacent orbits. Here, it is necessary to calculate and correct the offset based on the parameters estimated by time-series InSAR. The formula is as follows:

$${\bar {v}_i}=\frac{{{\sigma _{{m_i}}}{d_{{m_i}}}+{\sigma _{{m_i}}}{{\bar {d}}_{{s_i}}}}}{{{\sigma _{{m_i}}}+{\sigma _{{s_i}}}}}.$$

(8)

Where $\overline{d} _{si}$ is the subsidence rate of the secondary image coherent target i after correction, $\:{d}_{{m}_{i}}$ is the subsidence rate of the primary image coherent target i, $\:{\sigma\:}_{{m}_{i}}$ and $\:{\sigma\:}_{{s}_{i}}$ are the weight factors of the primary and secondary orbits, respectively.

Methods for analyzing the contribution of sedimentation inducing factors

The population density data used in this study comes from the annual world population density data provided by the Oak Ridge National Laboratory in the United States (https://landscan.ornl.gov/). Rainfall data is sourced from the 1 km monthly rainfall dataset provided by the Tsinghua University Environmental Big Data Platform. NDVI data is sourced from the MODIS dataset under MOD13A²⁸. Due to the limited spatial density of measurement points for groundwater level data, it is difficult to support the analysis and evaluation of the relationship between groundwater level changes and surface subsidence after the opening of the South-to-North Water Diversion Middle Route project. Therefore, this study uses GRACE gravity satellite data and GLDAS data to invert groundwater storage changes based on the water balance equation, with the specific formula as follows²⁹:

$${\vartriangle _{GWS}}={\vartriangle _{TWS}} - ({\vartriangle _{SM}}+{\vartriangle _{CAN}}+{\vartriangle _{SWE}}+{\vartriangle _{SW}}).$$

(9)

In the equation, $\:{\vartriangle}_{GWS}$ represents the change in groundwater storage, $\:{\vartriangle}_{TWS}$represents the change in land water storage, $\:{\vartriangle}_{SM}$represents the change in soil moisture, $\:{\vartriangle}_{CAN}$represents the change in total canopy water storage, $\:{\vartriangle}_{SWE}$ represents the change in snow water equivalent, and $\:{\vartriangle}_{SW}$ represents the change in surface runoff. All variables are expressed in terms of equivalent water height (mm).

This study uses a random forest (RF) regression algorithm to quantify the contribution of NDVI, population density, precipitation, and groundwater storage to ground subsidence³⁰. RF is a nonparametric supervised learning method applied to classification and regression, which was first proposed by Breiman in 2001 and is one of the practical algorithms for the Bagging integration strategy. Random forest utilizes the bootstrap self-help method for resampling, which randomly draws n samples from the entire sample set and puts them back to generate a new training set, constructs mutually independent decision trees by training the new sample set, and combines the trained n decision trees into a forest³¹.

Feature importance is a measure used to assess the usefulness of a feature in the classification prediction process of a model; the higher the importance of a feature, the more useful the feature is to the model. Machine learning algorithms based on tree models provide a “feature importance” toolkit based on the structural score gain to calculate which feature to choose as a split point. The importance of a feature is the sum of the number of times it appears in all the trees, i.e., the more attributes of a feature are used to build a decision tree in the model, the higher its importance is. The higher its importance is. This paper uses the importance ranking of features to reveal the magnitude of the effect of different induced factors on sedimentation.

Results

MTLS-InSAR deformation monitoring result

To ensure the integrity and continuity of the data, this study selected Sentinel-1 satellite images covering the MRP, including 14 cities along the route such as Beijing, Baoding, Shijiazhuang, and Handan. Figure 4 illustrates the average deformation rate from 2017 to 2021 for the South-to-North Water Diversion Middle Route Project, calculated using a joint leveling correction method with long-strip, multi-orbit time-series InSAR. This method used 1209 SAR images for experiments along the SNWD-MRP corridor. By integrating actual leveling data into the conventional time-series InSAR monitoring method, a quadratic polynomial fitting function was constructed using the leveling data and InSAR results. The coefficients from this function were used to calibrate the InSAR results. Table 3 shows the positions and annual average subsidence rates of the measured leveling points.

The selected leveling data are shown in Table 3, these points are located in the southwestern part of the study area on orbit number 40 and in the northeastern part on orbit number 142. Points 1 to 6 are located in Beijing, points 7 and 8 are in Tianjin, and points 9 to 13 are within Henan Province. Subsequently, the calibrated result image was used as a reference image to estimate and correct the overall bias of adjacent strips and to stitch them together. Finally, the images from adjacent orbits were unified to a common reference coordinate system and reference datum, achieving the stitching of adjacent orbit data. Due to the limited availability of measured point data, the validation data for this monitoring result is the same as the calibration data. After calibration, the RMSE of the annual subsidence data is ± 5.1 mm.

Figure 4 shows multiple instances of uneven ground subsidence along the entire route, mainly in the BTH region, including Beijing, Tianjin, Langfang, Hengshui, Cangzhou, Handan, and Zhengzhou. The largest subsidence is in the southern part of the Jiaozuo to Xinxiang canal section, with a maximum rate of 187 mm/yr. A large subsidence funnel in southeastern Hebei Province has a maximum rate of about 130 mm/yr. The study also focuses on a 10 km buffer zone on both sides of the canal’s central line, as shown in Fig. 4 (b). Ground subsidence along the central canal has transitioned from stable to accelerated sinking from south to north, with five significant subsidence areas within the buffer zone. These include two areas along the Tianjin branch line, two small subsidence funnels near the Handan to Xingtai canal, and one along the Jiaozuo to Xinxiang section in the south.

Fig. 4

(a) Average subsidence rate of 2017 to 2021 after multi-track and long-strip mosaicking. (b) Average subsidence rate of 2017 to 2021 within 10-km Buffer Zone on both sides of the canal centerline. (GMT, Version 6.5; https://docs.gmt-china.org/).

Starting from June 2017, cumulative subsidence was extracted for June 2017 to December 2018, 2019, 2020, and 2021, as shown in Fig. 5. The spatial distribution pattern of cumulative ground subsidence from 2017 to 2021 is similar to that of the subsidence rate, showing continuous development over time. The largest cumulative subsidence in Hebei Province is in southeastern Handan City, with a maximum of 133 mm (June 2017 to December 2018), 215 mm (June 2017 to December 2019), 468 mm (June 2017 to December 2020), and 585 mm (June 2017 to December 2021).

Fig. 5

(a) Cumulative subsidence for 2017–2018; (b) Cumulative subsidence for 2017–2019; (c) Cumulative subsidence for 2017–2020; (d) Cumulative subsidence for 2017–2021. (GMT, Version 6.5; https://docs.gmt-china.org/).

Table 3 Measured levels and subsidence Rates.

Contribution analysis of subsidence inducing factors

As discussed above, surface subsidence aligns closely with the range of groundwater level fluctuations and the extent of groundwater over-extraction, further confirming that excessive groundwater extraction is the main controlling factor causing surface subsidence in the North China Plain. However, the relationship between surface subsidence and groundwater depletion is a complex systemic issue influenced by multiple factors. Therefore, in this section, we take the 10 km buffer zone on both sides of the Tianjin branch corridor of the South-to-North Water Diversion Middle Route as an example. We comprehensively select groundwater storage, population density, rainfall, and NDVI as factors inducing subsidence. Using machine learning methods, we quantify the impact of each inducing factor on surface subsidence.

Due to the different temporal and spatial resolutions of the data, to facilitate the subsequent input of data into machine learning training, all data have been resampled to a spatial resolution of 1 km and converted to an annual average rate, as shown in Fig. 6.

Fig. 6

Characteristic vectors for subsidence attribution analysis, 2017–2021 (a) Mean annual maximum synthetic NDVI; (b) Mean annual rate of change in groundwater storage, mm/yr; (c) Mean annual rainfall, mm/yr; and (d) Mean annual population density (people/km²). (ArcGIS Pro, Version 3.1; https://www.esri.com/zh-cn/arcgis/products/arcgis-pro/overview).

To further investigate the impact of various inducing factors on surface deformation, this study selected a random forest regression model based on the grid search method. This model uses the grid search method to train and find the optimal combination within the parameter threshold range. The optimized hyperparameters based on the grid search method are as follows: max_depth = 80, max_samples_leaf = 1, min_samples_split = 6, and n_estimators = 700. The RMSE for the test set is 0.98, and the MSE is 0.96. The training results are shown in Fig. 7. As can be seen from the figure, among the various inducing factors of subsidence, the change in groundwater storage has the highest proportion, followed by rainfall, NDVI, and population density. The change in groundwater storage accounts for 75.1% of the inducing factors for ground subsidence. These results once again confirm that groundwater depletion is the main controlling factor causing ground subsidence.

Fig. 7

Feature Importance.

Discussion

Analysis of time-series evolution trend of key subsidence areas

Following the implementation of the MRP, a detailed examination of ground subsidence in Beijing was conducted. As shown in Fig. 8, the annual average subsidence rate from June 2017 to December 2021 showed significant uneven patterns, particularly in the eastern and southeastern regions. Major subsidence funnels were identified in Chaoyang and Tongzhou Districts, with rates of 71.9 mm/yr and 81.2 mm/yr, respectively. Other areas, such as parts of Haidian, Shunyi, and Daxing Districts, also exhibited subsidence trends, albeit at lower rates.

As shown in Fig. 9, six typical points in the Chaoyang and Tongzhou Districts were analyzed using Sentinel-1 A subsidence data. Polynomial fitting revealed that points P1 and P2 in Chaoyang District showed severe subsidence, with cumulative amounts of 30.16 cm and 30.43 cm. Points P3 and P4 in Tongzhou District had cumulative subsidence of 18.28 cm and 18.55 cm. Points P5 and P6 in Sanjianfang Township showed lesser subsidence of 12.02 cm and 18.54 cm. The subsidence trends indicated an initial acceleration followed by a deceleration post-2020, suggesting the effectiveness of the SNWD Project and subsidence control measures.

Fig. 8

Annual average subsidence rate in Beijing from 2017.6 to 2021.12. (GMT, Version 6.5; https://docs.gmt-china.org/).

Fig. 9

Continuous subsidence sequence of 6 typical points. (Origin, Version2019; https://www.originlab.com/).

According to Fig. 4, subsidence funnels are observed along the Tianjin branch line. Therefore, a 10 km buffer zone on both sides of the main canal of the Tianjin branch line was selected as the study area. We used average subsidence rate data from time-series InSAR technology for the Tianjin branch line from June 2017 to December 2021, as shown in Fig. 10. Two main subsidence funnels, S1 and S2, are identified. S1, at the junction of Baoding City and Langfang City, has a maximum subsidence rate of 67.2 mm/yr in Zhongying Village in the southern buffer zone. S2, at the junction of Langfang City and Tianjin City, has a maximum subsidence rate of 98.8 mm/yr in the northeast buffer zone, far from the canal. The subsidence areas within the buffer zone encompass over a dozen villages. To further analyze subsidence near the branch line, the cumulative subsidence amounts of eight severely affected villages from June 2017 to December 2021 were extracted, as shown in Fig. 10.

Fig. 10

Average subsidence rate of Tianjin branch line 2017.6 ~ 2021.12. (GMT, Version 6.5; https://docs.gmt-china.org/).

Combining the analyses from Figs. 10 and 11, the S1 funnel covers six villages. Villages closer to the main canal of Tianjin, such as Dazhuang, Mazhuang, and Quantou, have subsidence rates ranging from 24.8 to 25.8 mm/yr. In contrast, villages farther from the canal, like Senan, Xuma, and Zhongying, experience more severe subsidence, with rates between 42.0 and 67.2 mm/yr. Specifically, subsidence rates increase with distance from the canal. At the end of the Tianjin branch line, S2, a broader subsidence area, has a more significant overall subsidence rate than S1. Villages in S2, such as Yangjia, Youzhangbu, and Yangjiabu, exhibit subsidence rates from 39.2 to 98.8 mm/yr. Thus, villages near the S1 funnel receive effective water supply, while those farther from the canal and at the end of the Tianjin branch line experience significant subsidence.

Fig. 11

Cumulative subsidence of villages in Tianjin branch line from 2017.6 to 2021.12 (T1 is Dazhuang village, T2 is Mazhuang village, T3 is Circle head village, T4 is Sennan village, T5 is Xuma village, T6 is Yangjiabao village, T7 is Youzhangbao village, T8 is Zhongying village). (Origin, Version2019; https://www.originlab.com/).

Figure 12 shows average subsidence rates from June 2017 to December 2021 between Handan City and Xingtai District in Hebei Province. Significant and uneven subsidence is observed on the right side of the canal section, with a maximum rate of 130 mm/yr. Notably, subsidence funnels S3 and S4 within the 10 km buffer zone along the main canal encompass villages such as Nanjiang and Beicheng. To understand subsidence evolution in this area, cumulative subsidence values need to be extracted and analyzed.

Figure 13 illustrates cumulative subsidence and trends in Beicheng and Nanjiang Villages from June 2017 to December 2021. Beicheng Village has a gentle subsidence rate of 8.2 mm/yr, with a cumulative subsidence of 5.49 cm. Conversely, Nanjiang Village exhibits a more severe subsidence rate of 40 mm/yr, with a cumulative subsidence of 22.42 cm. However, due to the distance from the main canal, subsidence in Nanjiang Village does not threaten the canal’s safe operation.

Fig. 12

Annual average subsidence rate of Handan to Xingtai canal section from 2017.06 to 2021.12. (GMT, Version 6.5; https://docs.gmt-china.org/).

Fig. 13

Cumulative subsidence of S3 and S4 funnels from 2017.06 to 2021.12 (BC represents Beicheng Village, NJ represents Nanjiang Village). (Origin, Version2019; https://www.originlab.com/).

Figure 14 illustrates the average subsidence rate from June 2017 to December 2021 in the section from Jiaozuo to Xinxiang in Henan Province. From the figure, it can be observed that there is a “strip-shaped” subsidence belt, S5, on the south side of the Jiaozuo to Xinxiang canal section. This subsidence belt covers over ten villages including Zhangtianhe Village, Lu Village, Guanzhuang Village, Wu Village, Zhaotun Township, Pan Village, Wenzhuang Village, Xiwanghe Village, Dongwanghe Village, and Zhaogu Village from west to east. The subsidence rate in this area ranges from 25.6 to 187 mm/yr.

Figure 15 is illustrated the cumulative subsidence map, it can be observed that Zhangtianhe Village, which is closest to the canal, experiences relatively gentle subsidence, which has a subsidence rate of 25 mm/yr. The most severely affected area is located in Dongwanghe Village, where the cumulative subsidence exceeds 80 cm. The subsidence rates in other villages range from 51 ~ 111.9 mm/yr. In terms of the overall subsidence trend, except for Dongwanghe Village, which shows a continuous sinking trend, the subsidence trends in the remaining villages are gradually easing. Overall, the foundation of the Jiaozuo to Xinxiang main canal is basically stable, and there is no large-scale subsidence occurring.

Fig. 14

Annual average subsidence rate of Jiaozuo to Xinxiang drainage section from 2017.06 to 2021.12. (GMT, Version 6.5; https://docs.gmt-china.org/).

Fig. 15

Cumulative subsidence of Jiaozuo to Xinxiang canal section from 2017.06 to 2021.12 (P1 represents Wenzhuang Village, P2 represents Zhaogu Village, P3 represents Dongwanghe Village, P4 represents Dabeicheng Village, P5 represents Pan Village, P6 represents Guanzhang Village, P7 represents Lu Village, and P8 represents Zhangtianhe Village). (Origin, Version2019; https://www.originlab.com/).

Correlation analysis between subsidence and groundwater level

To further investigate the controlling effect of groundwater level changes on subsidence along the route of the MRP, this study selected a total of 14 monitored groundwater wells for correlation analysis. Combining groundwater level changes with InSAR subsidence results reveals that subsidence in Beijing Municipality and the Tianjin branch line has transitioned from slow sinking to stability, while most cities in Hebei Province continue to exhibit a trend of sustained subsidence.

Table 4 Presents the correlation coefficients between each groundwater well and InSAR results. It can be observed from the table that the subsidence results from InSAR lag behind the groundwater level changes to varying degrees, with a lag of 1 to 2 months observed in most areas compared to groundwater level changes.

Table 4 Groundwater wells and InSAR correlation analysis.

Figure 16 shows the correlation between groundwater wells in Beijing Municipality and InSAR monitoring results. The correlation coefficients range from 0.6 to 0.76, indicating consistent interannual variations in groundwater level changes and InSAR subsidence. Six wells displayed a slow sinking trend from 2018 to 2019, followed by a gradual rise from 2020 to 2021. InSAR data indicate subsidence rates plateaued after 2020, with minor fluctuations but a stable overall trend.

Additional deep groundwater wells near Beijing’s subsidence funnels were analyzed (Fig. 17), revealing a yearly rise in deep groundwater levels. InSAR results show subsidence gradually leveling off over time.

It is important to highlight that Well No. 6, located in the Xiaotangshan area of Changping, lies within a major geothermal resource enrichment zone in Beijing. The extraction of geothermal fluids may induce formation compaction and associated surface deformation. The ground displacement in this region is influenced by multiple factors: (1) non-poroelastic deformation due to geothermal exploitation; (2) additional deformation caused by municipal engineering projects (e.g., load effects from subway construction); and (3) differential settlement at the Quaternary/bedrock interface. These deformation mechanisms, which are not primarily governed by groundwater dynamics, significantly weaken the linear correlation between water level variations and InSAR-derived surface displacements.

Figure 18 depicts the correlation between groundwater wells along the Tianjin branch and InSAR data. Correlation coefficients range from 0.6 to 0.9, indicating a high agreement between groundwater changes and InSAR subsidence. Four wells show accelerated groundwater decline from January 2018 to June 2020, followed by a gradual rise. Tianjin and Langfang experienced accelerated subsidence initially, transitioning to a slower subsidence trend, necessitating enhanced monitoring of the canal section to mitigate potential impacts.

Fig. 16

Correlation diagram between groundwater wells and InSAR monitoring results in Beijing (a: Goubei Village, Shunyi District; b: Wangjiachang, Shunyi District; c: Xiaonandi, Tongzhou; d: DaNiuFang, Haidian District; e: Xiaotangshan, Changping; f: Xicheng Library). (Origin, Version2019; https://www.originlab.com/).

Fig. 17

Correlation between deep groundwater wells and InSAR results in Beijing (a: Louzhizhuang Middle School, Chaoyang District; b: Daludian, Chaoyang District; c: Dongxiaoying, Tongzhou; d: Beijing University of Technology). (Origin, Version2019; https://www.originlab.com/).

Fig. 18

Correlation between groundwater wells and InSAR monitoring results of Tianjin branch line (a: Xiaoshenzhuang, Langfang City; b: Fuxingzhuang Village, Wangqingtuo Town, Wuqing District, Tianjin City; c: Donggu Harbor, Langfang City; d: Houyi Village, Yongqing County, Langfang City). (Origin, Version2019; https://www.originlab.com/).

To further analyze changes in deep groundwater along the Tianjin branch, this study selected four wells for analysis (Fig. 19). Figure 19 (a) and (b) are within the Tianjin urban area, while (c) and (d) are in the S2 subsidence funnel. Water levels and cumulative subsidence generally rise annually, with noticeable seasonal fluctuations in the S2 funnel’s deep groundwater. Overall, Tianjin’s urban area no longer subsides due to sufficient groundwater replenishment, while subsidence in the funnel slowed after 2020, indicating some deep groundwater replenishment.

Fig. 19

Correlation between deep groundwater wells and InSAR monitoring results in Tianjin branch (a: Tianjin Vocational University, Beichen District; b: Beibaitansi Village, Yangliuqing Town, Xiqing District; c: Xiaoshenzhuang, Langfang; d: Houyi Village, Yongqing County, Langfang City). (Origin, Version2019; https://www.originlab.com/).

Figure 20 depicts the correlation between Hebei Province’s central/eastern groundwater wells and InSAR results. Groundwater level changes correlate with InSAR subsidence (correlation coefficient: 0.65 to 0.87). From 2018 to 2021, groundwater levels in Baiheji Township, Linzhang County, Handan City, and Mucun Township, Shenzhou City, Hengshui City, showed continuous decline and subsidence. Gaoliang Village, Weixian County, Xingtai City, and Beizong Village, Shanglin Town, Xian County, Cangzhou City, initially sank slowly, then rose slightly, yet continued to subside. Feng Wei²⁹ et al. attributed ongoing subsidence in the North China Plain’s central/eastern regions to over-extraction of deep, primarily saline groundwater, without timely replenishment.

Fig. 20

Correlation between groundwater wells and InSAR monitoring results in the east-central part of Hebei Province (a: Gaoliang Village, Weixian County, Xingtai City; b: Beizong Village, Shanglin Township, Xian County, Cangzhou, China; c: Baiheji Township, Linzhang County, Handan City; d: Mucun Township, Shenzhou City). (Origin, Version2019; https://www.originlab.com/).

Well No. 13, located in Linzhang County, lies within the agricultural zone of the North China Plain and is subject to intensive seasonal irrigation activities. These activities lead to rapid recharge and discharge cycles in the shallow aquifer system. The deformation signal derived from InSAR (Interferometric Synthetic Aperture Radar) primarily reflects elastic deformation within the middle-to-deep aquifer system. However, the fluctuations in shallow groundwater levels may exhibit temporal decoupling from the InSAR observations. Specifically, irrigation-induced abrupt changes in shallow water levels typically occur on a weekly timescale, whereas InSAR monitoring operates on a monthly timescale. This mismatch in temporal resolution results in poor correlation between the two datasets. Additionally, hydraulic hysteresis effects within the multi-layered aquifer system further diminish the statistical correlation between shallow water level changes and InSAR-detected deformation.

Moreover, Well No. 13 is situated at the margin of an alluvial-proluvial fan, where the hydrogeological structure consists of interbedded sand-gravel and clay layers. This heterogeneous stratigraphy can cause attenuation of deformation signals as they propagate through the subsurface. Consequently, this attenuation affects the accuracy of InSAR measurements and reduces the reliability of correlation analyses between groundwater level changes and surface deformation.

Water resources management recommendations

To address the challenges of sustainable groundwater management along the South-to-North Water Diversion Project route, this study proposes an integrated governance framework combining “stratified regulation-ecological recharge-intelligent monitoring”. Regarding extraction control, a differentiated management approach should be implemented for confined and phreatic aquifers, including total withdrawal control and water level thresholds for deep confined aquifers, and dynamic quota systems synchronized with crop growth cycles for shallow phreatic aquifers. For ecological restoration, a multi-source artificial recharge system should be established utilizing return flow channels from the diversion project, with priority given to constructing recharge basins at alluvial fan apexes maintaining annual recharge volumes no less than 20% of extraction amounts. In terms of monitoring, an integrated space-ground observation network incorporating InSAR, distributed fiber optic sensing and machine learning should be developed, along with coupled groundwater-subsidence early warning models. The study further recommends piloting “precision irrigation-water rights trading” initiatives in the Beijing-Tianjin-Hebei coordinated development zone to enhance agricultural water efficiency through smart irrigation infrastructure and water market mechanisms, supported by improved implementation rules for the Groundwater Management Regulations and cross-regional ecological compensation schemes. This comprehensive approach specifically accounts for the hysteresis effects in North China Plain aquifer systems, aiming to establish a sustainable groundwater utilization paradigm featuring “total quantity control, stratified utilization, and dynamic equilibrium” through 3–5 years of demonstration and replication, thereby providing transferable experience for water resource management in similar regions.

Conclusion

To address the issue of severe accuracy loss in traditional wide-area long-term InSAR surface deformation monitoring, this paper proposes a long-strip, multi-track time-series InSAR method integrated with leveling correction. This approach enables the extraction of full-domain surface deformation for the South-to-North Water Diversion Project from June 2017 to December 2021. It conducts in-depth long-term evolutionary analysis of significant subsidence areas and examines the impacts of groundwater data on land subsidence across spatio-temporal scales. Finally, we used machine learning to quantify the contributions of population density, rainfall, NDVI, and groundwater changes to surface subsidence, reaching the following conclusions:

• (1) Using 1,209 SAR image datasets, we conducted experiments along the middle route of the South-to-North Water Diversion Project, validating the feasibility of this method for wide-area surface deformation extraction. The algorithm accuracy was verified as ± 5.1 mm through integration with in-situ annual subsidence measurements.

• (2) Ground subsidence monitoring reveals that the SNWD corridor is near large subsidence funnels in the BTH region. Persistent subsidence funnels are observed in eastern Beijing, particularly Chaoyang and Tongzhou districts, with a maximum subsidence rate of 81.2 mm/yr and a cumulative subsidence of 34.8 cm over five years. Significant subsidence also occurs along a line from Baoding to Handan in southeastern Hebei Province, with a maximum rate of 130 mm/yr and a cumulative subsidence of nearly 60 cm over five years. Within a 10 km buffer zone along the MRP, there are five subsidence funnels, with two significant ones in the Tianjin branch route, showing maximum rates of 67.2 mm/yr and 98.8 mm/yr.

• (3) Groundwater measurement data indicated that deep groundwater levels in the subsidence funnels of Beijing and Tianjin urban areas showed a year-by-year rising trend, while shallow groundwater exhibited seasonal fluctuations. Further statistical comparison revealed a high correlation between groundwater level changes and InSAR-derived subsidence changes, with correlation coefficients ranging from 0.6 to 0.9.

• (4) Machine learning analysis of subsidence-inducing factors, including population density, precipitation, NDVI, and groundwater storage changes, showed that groundwater storage changes accounted for 75.1% of land subsidence, followed by precipitation, NDVI, and population density. This indicates that groundwater depletion remains the primary controlling factor for ground subsidence.