*2.1. Wide-Area Deformation Monitoring Using Stacking*

The stacking technology can calculate the deformation rate based on weight and average the unwrapped phases of the multitemporal DInSAR dataset. The stacking technology assumes linear ground-deformation changes, and temporal randomly distributed phase noise, such as atmospheric delay phase. Assuming *N* + 1 SAR images of one frame covering the WSA constitute *M* InSAR pairs, the displacement phase can be separated as

$$\overline{\phi} = \sum\_{i=1}^{M} \phi\_i \cdot \Delta t\_i / \sum\_{i=1}^{M} \Delta t\_i^2 \tag{1}$$

in which *φ* is the rate of deformation phase change. *φ<sup>i</sup>* and Δ*ti* are the interference phase and the time interval of the *i*-th InSAR pair, respectively.

The rate of the deformation phase change would be converted to the deformation rate (*Vdef*) as,

$$V\_{\rm def} = \lambda \cdot \overline{\Phi} / 4\pi \tag{2}$$

where *λ* is the wavelength of the SAR sensor. The multitemporal DInSAR data in each frame is processed using stacking technology to obtain the ground deformation of the WSA.

#### *2.2. ROI Detection Based on Wide-Area Deformation Rate*

Luo et al. [39] proposed an improved method to automatically identify and evaluate geological hazards using TS–InSAR results. By judging and analyzing the deformation rate and time series in the monitoring area, the method can automatically identify the deformation region and evaluate its hazard grade. In this study, we improve this method to accurately delineate the ROI.

To improve the accuracy of ROI detection, we first apply spatial domain filtering to the wide-area monitoring results to obtain deformation results with good spatial consistency. Then, we set the thresholds for deformation rate, extension radius, and minimum clustering area. When the absolute value of the deformation rate is greater than the deformation rate threshold, it is considered to be an active point. Otherwise, it is a stable point. Buffer zones are established around the active points according to the extension radius. The active points are clustered following the principle of spatial proximity relationship [43]. The clustering regions are smoothed to refine the boundary. The robust deformation regions and their area are obtained by removing regions smaller than the minimum clustering area. The ROI can be finally located based on spatial clustering and the area of deformation.

A detailed description of the intelligent recognition part of the method can be found in [39]. It should be noted that InSAR can only obtain one-dimensional (1D) deformation along the line-of-sight (LOS) direction of the SAR sensor, so the InSAR data of one geometry is insensitive to the deformation of some regions, especially landslides [44]. To obtain more reliable deformation detection results, we need to use the above method and InSAR data from different observational geometry. The detection results of multitrack InSAR data are taken together as the final deformation regions. Then, we can adaptively determine the ROI and perform fine monitoring.

#### *2.3. ROI Deformation Refinement Using Advanced TS–InSAR*

When calculating the wide-area deformation rate, we select the InSAR data with the same acquisition time from different frames to facilitate the splicing of the results from different frames and to maintain the consistency of the wide-area deformation rate. To accurately monitor the deformation in the ROI, we first crop the registered InSAR datasets. The cropped datasets are used to obtain the time-series and multidimensional ground deformation of the ROI. Detailed steps are as follows.


If multi-sensor and multitemporal InSAR data covering the ROI are available, we can collect all data to analyze the long-term deformation and understand the deformation spatio-temporal evolution features based on the data-overlapping and deformation model [47,48].

#### *2.4. Variable-Scale Deformation Product Generation*

The low-spatial-resolution deformation rate can be used to detect a stable surface in the WSA, which greatly reduces the task and data volume of wide-area InSAR deformation monitoring. In addition, we obtain the fine results of the deformation time series with a high spatial resolution of ROI using advanced TS–InSAR technology. A variable-scale deformation product organization structure includes low-spatial-resolution deformation rates in stable areas of the WSA and the high-spatio-temporal-resolution deformation in the ROI. Hence, we superimpose the high-spatio-temporal-resolution deformation at the corresponding regions of the ROI on the wide-area deformation rate results to improve the spatial and temporal dimensions of the deformation in the ROI. At this stage, we can obtain variable-scale deformation products in the WSA, which only contain low-spatial-resolution deformation rates in stable regions, and fine monitoring results in the ROI.

#### **3. Study Area and Data Processing**

#### *3.1. The Turpan–Hami Basin*

The Turpan–Hami basin, consisting of the Turpan and the Hami depressions, is an intermountain basin located in northwest China (Figure 2). Since the end of the Early Permian period, the Turpan–Hami basin has developed following the model of "faultdepression foreland". It is a typical faulted basin, with limited sedimentary range, great lateral variation of sedimentary thickness, and multiple depositions and subsidence centers. The geological conditions and active tectonic motion contribute to oil and gas accumulation and make the Turpan–Hami basin the largest coal-derived petroleum-producing basin in China [49]. Moreover, there are many mineral resources in this basin, e.g., coal, iron, and potassium (sodium) saltpeter. It is the world's largest potassium (sodium) saltpeter resource. Aydingkol Lake, located in the middle of the Turpan depression, is the lowest depression in China, 154.31 m below sea level [50]. Centering on Aydingkol Lake, the Turpan depression presents a roughly three-ring shape. The outermost ring has high snow-capped mountains. The middle ring is the Gobi gravel belt. The inner ring is an oasis plain belt, most of which belongs to a piedmont sloping plain, and accumulates a large area of fine soil alluvium. The water in the basin mainly comes from rainfall and meltwater from the surrounding mountains. The Tianshan mountains, e.g., Bogurda Mountain and Harlick Mountain, are in the north of the Turpan depression. The Flaming Mountains fault zone lies nearly east–west in the Turpan depression, between Turpan city and Shanshan county (Figure 2). Weathered material is transported from the Tianshan mountains to the center of the basin by water flow, but is blocked by the Flaming Mountains fault line and accumulates in the northern part of the mountains. The surface water and groundwater from the Tianshan mountains are also blocked by the Flaming Mountains fault line. The head height of the shallow aquifers is raised on both sides of the Flaming Mountains, creating overflow zones and an oasis in these areas.

The Turpan–Hami basin has a typical continental warm temperate desert climate, with abundant heat and extremely little precipitation. It has 3200 h of sunshine in a year. The hydrogeology, climate, and lighting conditions make it an ideal place for growing cantaloupe, grapes, cotton, and off-season vegetables. Groundwater is the main source of agricultural water in the arid area. Previously, karezes were the predominant underground water conservancy project in this region. A karez uses the principle of water potential artware to divert water from shallow aquifers to the surface for irrigation. There are more than 2000 karezes in the Turpan–Hami basin, accounting for more than 70% of the total number of karezes in Xinjiang [51,52]. However, many electromechanical wells have been built in the Turpan–Hami basin since the 1960s. Groundwater exploitation has increased yearly, with the annual overexploitation reaching 2.48 × 1010 m3, leading to the continuous decline of groundwater level. Advanced water conservancy facilities have reduced people's dependence on karezes. Meanwhile, the water supply source of karezes is shallow aquifers. The continuous reduction of groundwater level directly leads to the decrease or even drying-up of karezes [53]. The number of water-filled karezes in the Turpan depression decreased from 1237 in 1957 to 214 in 2014 [51]. In addition, the increased demand and excess consumption of water resources in upstream areas have seriously threatened the water supply of Aydingkol Lake, resulting in water area shrinkage. The exploitation

of groundwater and mineral resources will make the surface of the Turpan–Hami basin unstable and threatened by potential geohazards.

**Figure 2.** The Turpan–Hami basin and SAR data coverage.

### *3.2. InSAR Datasets*

To monitor wide-area deformation in the Turpan–Hami basin, we collected eight frames of InSAR data covering the whole Turpan–Hami basin from the Sentinel-1 satellite. The Sentinel-1 satellite began operation in April 2014, and has different observation periods in different regions, resulting in inconsistent periods of SAR data in different regions. To ensure the consistency of deformation rates from multiple frames, we selected the images (628 images in total) from the eight frames acquired from October 2017 to May 2020 (Table 1). The spatial coverage of each dataset is shown in Figure 2.

**Table 1.** Acquisition periods of the datasets.


Wide-area InSAR deformation shows that many subsidence funnels are concentrated in the south part of the Flaming Mountains fault zone in the Turpan depression (hereafter referred to as the SFM–def region). The SFM–def region (the yellow box in Figure 2) was selected as an application demonstration area of ROI to carry out the fine monitoring of the deformation time series. Four frames from the ALOS-1/PALSAR dataset spanning from 2007 to 2010 (green rectangles in Figure 2) and a descending track from the Sentinel-1 dataset (red rectangle in Figure 2) covering the SFM–def region were collected. The common monitoring time of the Sentinel-1 ascending (AT41F135) and descending (DT121F449) tracks data is from 2015 to 2020 (Table 1). These data were used to precisely monitor the long-term and fine deformation in the SFM–def region.

#### *3.3. Data Processing*

We preprocessed all InSAR datasets covering the WSA. In each frame, one image was selected as the master image to register and resample the rest images. Multitemporal InSAR pairs were generated from SAR data in the same frame, based on the appropriate spatio-temporal baseline thresholds. All multitemporal DInSAR pairs were processed using GAMMA software [54] and two-pass DInSAR technology [55] to obtain multitemporal deformation signals. The shuttle radar topography mission (SRTM) digital elevation model (DEM) with a resolution of 30 m [56] was employed to remove the topographic phases. The point targets with a coherence lower than 0.3 were eliminated [57]. Least-squares-based filtering and the minimum cost flow method [58] were then applied to further suppress phase noise [59] and unwrap the differential interferogram, respectively.

The eight frames of the Sentinel-1 data were preprocessed with a spatial baseline (perpendicular) and temporal baseline of 100 m and 48 days, respectively, and a multi-looking operation of 20:4. Then, stacking was used to process all the multitemporal deformation signals in each frame, to obtain a wide-area deformation rate of the Turpan–Hami basin. The adaptive deformation detection method proposed in Section 2.2 was used to delineate the deformation regions. The thresholds of the deformation rate, extension radius, and minimum clustering area were set as ±2 cm/year, 250 m, and 1 km2, respectively.

For the SFM–def region, we set the multi-looking parameters of ALOS-1/PALSAR and Sentinel-1 data as 3:8 and 8:2, respectively. The improved IPTA method was used to compute the four frames of the ALOS-1/PALSAR data and the ascending/descending tracks from the Sentinel-1 datasets to obtain long-term and high-resolution displacements. Moreover, MSBAS technology was used to obtain multidimensional deformation from the ascending/descending tracks of the Sentinel-1 datasets. Then, we obtained the variable-scale deformation product of the Turpan–Hami basin, which consists of low-spatial-resolution deformation rates in the stable areas and high-spatio-temporal-resolution deformation in the SFM–def region.

#### **4. Results**

#### *4.1. Monitoring and Detecting the Wide-Area Deformation in the Turpan–Hami Basin*

The wide-area ground subsidence in the Turpan–Hami basin (Figure 3) shows that the surface of the Turpan–Hami basin is generally stable. The regions with deformation account for a small proportion of the whole. The main deformation type is subsidence. Based on the deformation detection threshold set in Section 3.3, we identified 32 deformation areas (the funnel) in the Turpan–Hami basin (the blue lines in Figure 3). The area of each funnel is shown in Table 2. The detected deformation area accounts for about 2.4‰ of the total monitoring area.

Analyzing the hydrogeology and land cover of the deformation areas, we divided the ground deformation in the Turpan–Hami basin into three types:


**Figure 3.** Wide-area subsidence rate map and the detected deformation regions. The numbers identify the location of the top 10 deformation regions. (**a**) The SFM–def region in Figure 4. (**b**) One of the major mining areas. Background image: Google Maps satellite image.

**Figure 4.** Deformation rate along LOS directions from (**a**) ALOS-1/PALSAR, (**b**) ascending track Sentinel-1, and (**c**) descending track Sentinel-1 data. Negative values indicate the direction away from the SAR satellite, while positive values indicate the opposite. (**d**) The hydrogeology of this area. (**e**,**f**) The deformation rate along the up–down and east–west directions calculated from ascending/descending tracks Sentinel-1 data. The red dotted line delineates the central area of the subsidence funnels from 2007 to 2010. The magenta dotted line delineates the central area of the subsidence funnels from 2015 to 2020.


**Table 2.** The area of the deformation funnels.

The largest deformation funnel is distributed in the SFM–def region, with an area of 437.6 km2, surrounded by small funnels (Figure 3a). The optical images show that the subsidence funnels in the SFM–def region are highly correlated with the location of agricultural areas. Aydingkol Lake is in the south of the SFM–def region (Figure 3a). In recent years, the area of the lake has continuously shrunk, and a large area of saline–alkali land has appeared. There is obvious ground uplift in these saline–alkali regions. In addition, multiple subsidence funnels are observed close to some mines, e.g., the funnel cluster in Figure 3b. The wide-area deformation results are discussed in detail in Section 5.
