Estimating Gully Erosion Induced by Heavy Rainfall Events Using Stereoscopic Imagery and UAV LiDAR

Highlights

• We developed a large-scale erosion estimation method that fuses data from Chinese Gaofen-series high-resolution satellites and UAVs.
• Analyze and discuss the far-reaching impacts of anthropogenic and natural factors on soil conservation across the Loess Plateau.

What are the main findings?

• Using multi-source high-resolution remote sensing data, we quantitatively assessed the impact of heavy precipitation on erosion and established a large-scale, generalizable methodology.
• Using multi-source high-resolution datasets, we quantitatively assessed how natural processes and human activities influence erosion trends.

What is the implication of the main finding?

• We developed a systematic, high-precision erosion-rate analysis method that integrates China’s Gaofen-series satellites with UAV observations, is scalable to large areas, and enables a systematic assessment of how precipitation events affect erosion rates on the Loess Plateau.
• We elucidated the causes of soil erosion under the combined effects of human and natural factors and proposed methodological guidelines.

Abstract

Gully erosion, driven by the interplay of natural processes and human activities, results in severe soil degradation and landscape alteration, yet approaches for accurately quantifying erosion triggered by extreme precipitation using multi-source high-resolution remote sensing remain limited. This study first extracted digital surface models (DSM) for the years 2014 and 2024 using Ziyuan-3 and GaoFen-7 satellite stereo imagery, respectively. Subsequently, the DSM was calibrated using high-resolution unmanned aerial vehicle photogrammetry data to enhance elevation accuracy. Based on the corrected DSMs, gully erosion depths from 2014 to 2024 were quantified. Erosion patches were identified through a deep learning framework applied to GaoFen-1 and GaoFen-2 imagery. The analysis further explored the influences of natural processes and anthropogenic activities on elevation changes within the gully erosion watershed. Topographic monitoring in the Sandu River watershed revealed a net elevation loss of 2.6 m over 2014–2024, with erosion depths up to 8 m in some sub-watersheds. Elevation changes are primarily driven by extreme precipitation-induced erosion alongside human activities, resulting in substantial spatial variability in surface lowering across the watershed. This approach provides a refined assessment of the spatial and temporal evolution of gully erosion, offering valuable insights for soil conservation and sustainable land management strategies in the Loess Plateau region.

1. Introduction

Soil erosion is one of the most severe environmental challenges across the globe [1]. External forces such as anthropogenic disturbances, hydraulic and wind actions cause the detachment and transport of surface soil, directly leading to land resource degradation [2]. Moreover, soil erosion reduces carbon storage, thereby contributing to climate warming. Its socio-economic consequences are profound, particularly for rural communities dependent on agriculture, where it exacerbates vulnerability and poverty. Gully erosion represents one of the most conspicuous forms of soil erosion [3], characterized by its sudden onset, destructive impact, and irreversibility. Beyond causing the loss of arable land and ecosystem degradation, gully erosion can also trigger secondary hazards such as landslides and debris flows [4]. Therefore, accurate and systematic identification and quantification of gully erosion are critical for land resource conservation and the sustainable development of ecological environments.

Changes in rainfall characteristics, including rainfall amount, rainfall intensity, and rainfall spatial-temporal distribution, directly affect the process of runoff and soil erosion. Ongoing warming has increased water-holding capacity of the atmosphere, which contributes to changes in precipitation patterns and intensity [5]. The frequency and severity of extreme rainfall events have intensified significantly in recent years [6]. Saturation excess runoff, induced by long-duration and low-intensity rainfall, contribute significantly to soil erosion, particularly under conditions of bare soil surfaces or shallow soil layers [7]. Excess surface runoff, induced by extreme rainfall events, can remove nutrient-rich topsoil, depleting organic matter and essential minerals, thereby accelerating soil erosion [8]. In most climatic zones (especially in warm-humid areas), high-intensity and short-duration rainfall is more frequent than long-duration rainfall. High-intensity rainfall may induce a shift in runoff generation from saturation-excess dominated to infiltration-excess dominated [9]. Infiltration excess runoff accumulates and removes the soil from narrow channels to considerable depths, resulting in gully erosion [10]. Increasingly frequent rainfall events can accelerate gully erosion rates, with mega-gulling showing drastically increase under high-magnitude rainfall events [11]. Under extreme rainfall conditions, severe rill erosion and dense rill networks provide abundant concentrated flows and sediments for ephemeral and permanent gully formation, resulting in serious gully erosion and threatening watershed geomorphology [12]. Although the severity of runoff and erosion driven by rainfall extremes is mediated by vegetation, high-intensity and short-duration rainfall events can induce severe runoff and soil loss under all vegetation types [13,14]. Therefore, quantifying extreme rainfall-induced erosion risks is vital for understanding climate threats to ecological security and guiding targeted soil conservation strategies.

Gully erosion estimation has transitioned from conventional field-based surveys to geoinformatics-driven remote sensing approaches, with the advancement of remote sensing technologies [15]. Remote sensing-based estimation methods have gained widespread acceptance due to their robustness, cost-efficiency, and high accuracy, significantly facilitating soil erosion assessments over extensive spatial scales, particularly in remote or inaccessible regions [16]. Technological advancements have driven a shift in medium-resolution imagery applications (e.g., Landsat, Sentinel) from initial reliance on visual interpretation and terrain indices to the widespread adoption of object-based image analysis and machine learning algorithms [17,18,19,20,21]. Nevertheless, medium-resolution imagery remains insufficient for accurately capturing the morphological information of gully erosion [22]. Consequently, recent research has increasingly focused on high-resolution imagery (sub-meter to meter scales), which enables more precise detection and classification of various gully erosion [23]. Although high-resolution imagery has enhanced gully erosion detection, current technologies continue to face significant challenges, such as missing three-dimensional structural parameters and inadequate multi-source data integration. Meanwhile, the resolution limitations of satellite-derived elevation data introduce substantial errors, making the detection of local erosion insufficient to finely quantify all topographic changes. Addressing these limitations necessitates the development of a synergistic estimation framework that leverages high-resolution imagery. In this regard, domestically developed satellites demonstrate distinct advantages. The GaoFen7 satellites, China’s first equipped with laser altimetry, offers 1:10,000-scale stereoscopic mapping capabilities that provide critical support for estimating gully erosion depth. The GaoFen1 and GaoFen2 satellites supply 2 m (0.8 m) panchromatic and 8 m (3.2 m) multispectral imagery, which effectively facilitate the extraction of gully erosion patches. Additionally, the Ziyuan-3 satellite (2.1 m panchromatic) enables dynamic monitoring of gully erosion length and area. Therefore, integrating high-resolution data from GaoFen1, GaoFen2, GaoFen7, and Ziyuan-3 satellites into a multi-source collaborative estimation system establishes a more robust technical foundation for the accurate identification and continuous monitoring of gully erosion. For multi-source satellite–derived elevations, higher-resolution data (e.g., UAV) are still required to calibrate ground control points; the calibrated DSM then enables more accurate erosion estimation and evaluation. In addition, to characterize fine-scale local erosion, large-area satellite datasets benefit from high-precision, intelligent delineation of erosion patches, for example via convolutional neural networks for detection and segmentation [24]. Meanwhile, there are many other competitive approaches for remote sensing data classification using deep learning. For example, improving the FPN and fully convolutional head enables pixel-level segmentation of flood-affected buildings, reducing localization error by 30% [25]; integrating multi-scale features with a Swin Transformer achieves 97.78% accuracy on the AID dataset while reducing the number of parameters by 40% [26]; incorporating attention mechanisms with GANs improves the segmentation accuracy of small objects in complex backgrounds, outperforming the original Mask R-CNN baseline [27]. Using a multimodal feature fusion framework that combines the strengths of Vision Transformers and CNNs can significantly improve classification accuracy on remote sensing datasets from North America and Europe [28]. A systematic comparison of Transformer and CNN model performance on the EuroSAT (Europe) and UC Merced (USA) datasets shows that Transformers surpass traditional methods when sufficient data are available [29]. Moreover, by leveraging vision-language models, zero-shot land-cover classification can be achieved on the EuroSAT dataset, enabling generalization to unlabeled regions without any training [30]. While such deep-learning methods alleviate the limitations of traditional models in identifying small-scale, high-precision soil erosion, post-processing is still needed to improve the boundary accuracy of subtle topographic changes. Accordingly, applying a unified, training-free (zero-shot) deep-learning segmentation model [31] as a post-processing step can further enhance the accuracy of small-scale soil-erosion mapping from satellite imagery.

The Loess Plateau, spanning the middle and upper reaches of the Yellow River in North China, covers 6.6% of the land yet sustains 8.5% of the population in China [32]. As one of the planet’s most severely eroded regions, gully erosion alone generates more than 70% of hillslope erosion and about half of total sediment loss [33]. Ecological restoration programs, most notably the 1999 Grain-for-Green Project, have raised vegetation cover from 32% in 1999 to 63% in 2018, effectively mitigating gully incision [34]. Nevertheless, heavy rainfall events caused by climate change increasingly challenge these gains. Gully erosion is readily activated when rainfall intensity, duration, accumulated volume, or previous rainfall reach critical thresholds [35,36]. Frequent short-duration rainstorms between July and September in the Loess Plateau often trigger extreme erosion [37,38], seriously undermining the stability of the basin’s geomorphology [39,40]. The Loess Plateau in central Gansu Province serves as a typical region characterized by a high density of gullies and significant cutting depth. Among these regions, gully erosion is particularly severe in the Sandu River watershed, located in the hilly and dissected terrain of the Loess Plateau, where rainstorms are concentrated and intense during the rainy season.

Soil erosion poses a significant threat to ecosystem sustainability and agricultural productivity, particularly in sensitive regions like the Sandu River watershed, where complex terrain and intensified human activities accelerate land degradation. This study presents an integrated framework for quantifying gully erosion dynamics by synergizing multi-source remote sensing data and deep learning techniques. First, we generate an initial digital surface model (DSM) using Ziyuan-3 and GaoFen7 stereoscopic imagery, then enhance its accuracy through systematic correction with unmanned aerial vehicle-derived data for reliable gully depth calculation (2014–2024). Second, high-resolution GaoFen1 and GaoFen2 imagery are processed using deep learning models to automatically identify and delineate erosion patches. Third, we systematically evaluate the influence of both natural factors (i.e., gully erosion, vegetation restoration) and anthropogenic drivers (i.e., mining activities, terracing) on elevation changes during 2014–2024 using the corrected DSM time series and erosion patch maps. This study targets China’s Gaofen-series satellite data and, for the Loess Plateau, implements a two-stage intelligent erosion-extraction workflow—U-Net pre-screening followed by SAM-based post-processing. By incorporating both decadal erosion evolution and the impact of a single extreme precipitation event into a unified measurement framework, we provide a reproducible template for quantifying long-term soil loss as well as event-scale scour intensity. The approach is applicable to monitoring geomorphic change in ecologically fragile regions and has direct implications for soil and water conservation planning.

2. Materials and Methods

2.1. Materials

2.1.1. Study Area

The Sandu River watershed is located in the upper reaches of the Wei River Basin, which is a major hydrological subsystem of the Yellow River Basin, covering an area of 2484 km² (Figure 1). The watershed exhibits a continental monsoon climate, featuring a mean annual temperature of approximately 10.2 °C and precipitation of around 473.4 mm [41]. According to the spatial distribution of the average annual precipitation in the watershed over the last 11 years (Figure 2), the low precipitation areas of the watershed are mainly located in the north-central and western parts of the region. The upper part of the watershed shows the highest annual precipitation, with a significant decrease in the middle and upper reaches. The basin’s six-year mean annual total precipitation is 395.63 mm (

Table A2. Annual Cumulative Precipitation (2019–2024).

Category	Cumulative Precipitation (mm)
Year	2019	2020	2021	2022	2023	2024
Qin’an County	488.3	589	432.4	313.2	253.5	297.4
Tongwei County	475.7	544.5	423.3	270.7	197.8	348.5

), with a substantial share of rainfall concentrated in July–September (

Table A1. Heavy Rainfall Statistics (2014–2024).

Rainstorm Statistics—Cumulative Precipitation Value (Unit: mm)
Site Name	Site Code	Site Type	Precipitation	Precipitation Level	Statistical Time Period	Statistical Basis	Occurrence Time
Huajialing	52996	Basic Station	39.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	9 July 2015
Huajialing	52996	Basic Station	31.8	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	14 July 2015
Huajialing	52996	Basic Station	44.6	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	23 June 2016
Huajialing	52996	Basic Station	40	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	27 July 2017
Huajialing	52996	Basic Station	46.7	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	6 August 2017
Huajialing	52996	Basic Station	46.3	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	2 July 2018
Huajialing	52996	Basic Station	47.4	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	2 August 2019
Huajialing	52996	Basic Station	61	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	26 June 2020
Huajialing	52996	Basic Station	38.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	18 July 2020
Huajialing	52996	Basic Station	31.5	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	18 August 2022
Tongwei	53908	General Station	37.8	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	15 July 2022
Tongwei	53908	General Station	113.3	Severe Rain	Accumulated value from 20:00 to 20:00	Rainfall in the last 24 h: 100.0 to 249.9 mm	23 July 2024
Zhuanglang	53917	General Station	32.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	22 June 2022
Zhuanglang	53917	General Station	32.1	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	15 July 2022
Gan’gu	57001	General Station	71.9	Heavy Rain	Accumulated value from 20:00 to 20:00	Rainfall in the last 24 h: 50.0 to 99.9 mm	15 July 2021
Qin’an	57002	General Station	74	Heavy Rain	Accumulated value from 20:00 to 20:00	Rainfall in the last 24 h: 50.0 to 99.9 mm	15 July 2021
Qin’an	57002	General Station	33.9	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	24 April 2022
Qin’an	57002	General Station	40.6	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	11 July 2022
Wushan	57004	General Station	39.4	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	24 April 2022
Tianshui	57006	General Station	71.5	Heavy Rain	Accumulated value from 20:00 to 20:00	Rainfall in the last 24 h: 50.0 to 99.9 mm	15 July 2021
Tianshui	57006	General Station	43.4	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	24 April 2022
Tianshui	57006	General Station	41.3	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	22 June 2022
Maiji	57014	Basic Station	57.1	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	18 May 2017
Maiji	57014	Basic Station	57	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	7 August 2017
Maiji	57014	Basic Station	59.8	Heavy Rain	Accumulated value from 08:00 to 20:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	2 August 2019
Maiji	57014	Basic Station	57.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	20 August 2019
Maiji	57014	Basic Station	38.7	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	7 May 2020
Maiji	57014	Basic Station	53.7	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	25 July 2020
Maiji	57014	Basic Station	31.4	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	16 August 2020
Maiji	57014	Basic Station	38.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	17 August 2020
Maiji	57014	Basic Station	43.8	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	24 April 2022
Maiji	57014	Basic Station	32.6	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	14 July 2022
Maiji	57014	Basic Station	47.4	Heavy Rain	Accumulated value from 20:00 to 08:00	Rainfall in the last 12 h: 30.0 to 69.9 mm	27 July 2022
According to the Precipitation Classification Standard (GB/T 28592–2012) [78], issued and implemented by the China Meteorological Administration on 1 August 2012: Precipitation of less than 0.1 mm within 24 h is classified as trace precipitation (scattered drizzle). Precipitation of 0.1–4.9 mm within 12 h or 0.1–9.9 mm within 24 h is classified as light rain. Precipitation of 5.0–14.9 mm within 12 h or 10.0–24.9 mm within 24 h is classified as moderate rain. Precipitation of 15.0–29.9 mm within 12 h or 25.0–49.9 mm within 24 h is classified as heavy rain. Precipitation of 30.0–69.9 mm within 12 h or 50.0–99.9 mm within 24 h is classified as rainstorm. Precipitation of 70.0–139.9 mm within 12 h or 100.0–249.9 mm within 24 h is classified as severe rainstorm. Precipitation of ≥140.0 mm within 12 h or ≥250.0 mm within 24 h is classified as extraordinary rainstorm. In this study, items 5, 6, and 7 are used to count the number of rainstorm events and record their specific occurrence times.

) [42].

In the Sandu River watershed, land use is dominated by agricultural cultivation, with sparse vegetation and severe soil erosion. Its average annual erosion modulus reaches 8560 t/km²—the highest among all tributaries in the upper Wei River. The middle and lower reaches feature highly developed gully networks with elevated bifurcation ratios. The riverbed consists primarily of clayey sand, with gravel and cobble deposits being exceptionally rare throughout the channel [43]. Gully erosion exhibits significant spatial variability across the watershed, attributable to heterogeneous topographic gradients, pedologic characteristics, and vegetative cover [44]. This variability may be further amplified by extreme precipitation events. To investigate these spatial variations in erosion dynamics, we conduct a comparative analysis of five characteristic sub-watersheds within the watershed: Huanglong River, Jimaligou, Cunpinggou, Zhangjiagou, and Zaotangou.

2.1.2. Datasets

To characterize erosion patterns in the Sandu River watershed, we employed multi-source remote sensing imagery (

Table 1. Datasets used in this study.

Name	Resolution	Time	Bands	Sensor	Scenes	Data Preprocessing
CB04A	2 m	2024	Blue, Green, Red, Nir	WPM	1	Geometric correction, orthorectification, and image fusion (the raw Level-1A data had already undergone radiometric calibration and atmospheric correction).
GF-1/GF1D	2 m	2014, 2015, 2019, 2024	Blue, Green, Red, Nir	PMS1/PMS2/PMS	22
GF-2	0.8 m	2024	Blue, Green, Red, Nir	PMS1/PMS2	2
ZY302A	2 m	2014, 2023, 2025	Blue, Green, Red, Nir	PMS	5
GF-6	2 m	2019, 2024	Blue, Green, Red, Nir	PMS	3
ZY302A	10 m	2013–2017, 2019, 2020, 2022–2025	Three-line-array data	TMS	23	Tie-/control-point matching, block (bundle) adjustment, spatial correlation analysis, and DSM-based correction
GF-7	0.8 m	2020, 2023, 2024, 2025	Blue, Green, Red, Nir	DLC	12
UAV	0.2m	2024	Blue, Green, Red, Nir; LiDAR		5	Import of flight-image metadata (overlap ratio and camera parameters), aerial triangulation/AT, ground control point selection, and data fusion
Meteorological data	Daily	2014–2024	Station		9 stations	Data correspondence, Kriging interpolation

), including Ziyuan-3 (ZY-3), GaoFen7 (GF-7), GaoFen1 (GF-1), GaoFen2 (GF-2), and GaoFen6 (GF-6). High-resolution stereo-pair images from ZY-3 (10 m resolution for three-line-array camera data, 2 m resolution for digital orthophoto map) and GF-7 (0.8 m spatial resolution) were used to generate digital surface model (DSM), while GF-2 (0.8 m spatial resolution) and GF-1 (2 m spatial resolution) data enabled precise identification of erosion patches. Unmanned aerial vehicle (UAV) digital orthophoto map (0.2 m spatial resolution) and LiDAR data acquired in 2024 was applied to correct ZY-3-derived DSMs. Satellite data spanned 2014–2024, with additional years to ensure temporal coverage. Daily meteorological data (2014–2024) from 9 stations (e.g., precipitation) were sourced from China Meteorological Administration stations within the watershed (two stations) and in the vicinity of the watershed (Figure 2).

2.2. Methods

2.2.1. Overall Technical Framework

The technical flowchart begins with the preparation of input datasets, including stereo imagery, a baseline digital orthophoto map (DOM), a baseline digital elevation model (DEM), topographic maps, and ground control point (GCP). Control point matching is first conducted to achieve geometric consistency across datasets. This is followed by block adjustment through a regional network approach to further refine spatial alignment and reduce systematic errors. Elevation error analysis is then performed to quantify discrepancies between generated digital surface models and reference elevations. Spatial correlation analysis of DSM errors is undertaken to assess the pattern and extent of error distribution. Subsequently, a continuous spatial surface model is constructed to represent elevation differences and model systematic deviations (Figure 3).

2.2.2. DSM Calibration and Validation Based on UAV

Digital surface models (DSMs) for two distinct periods were generated using stereo imagery acquired from the ZY-3 satellite in 2014 and the GF-7 satellite in 2024. We employed the Radial Basis Function (RBF) interpolation method to correct multi-temporal digital surface model data within the study area, enabling the extraction of detailed channel change information. RBF interpolation is a technique used for interpolating scattered data in high-dimensional spaces. It establishes a continuous functional relationship between elevation differences ($∆Z$) at control points and their spatial coordinates $(x_i,y_i)$, thereby constructing a continuous spatial surface model of elevation differences to perform interpolation at any location within the region. First, stable ground features with minimal change from 2014 to 2024 were selected as control points, prioritizing roads, buildings, and engineered structures, while excluding cultivated land, channels, water bodies, and vegetation-covered infrastructure. We selected 170 control points distributed essentially uniformly over the study area for DSM elevation correction (Figure 4). During acquisition, these points were rigorously cross-checked against a decade of Google Earth imagery to ensure no positional changes or site reconstruction had occurred; additionally, control point locations were verified in situ during UAV operations. Field-measured terrain elevations obtained via unmanned aerial vehicle (UAV) surveys were assigned to these control points (Figure 3). The UAV survey covered five sample areas (Figure 4): (A) Jimali Village (flight extent: 104.8641–104.8860°E; 35.0395–35.0623°N), (B) Zhangjia Village (105.6138–105.6478°E; 34.8348–34.8570°N), (C) Huanglong Village (105.2426–105.2594°E; 35.2717–35.2949°N), (D) Zaotan Village (105.6742–105.6947°E; 34.7949–34.8139°N), and (E) Cunping Village (105.6670–105.6914°E; 34.8267–34.8471°N).Second, DSM spatial data were generated using tri-stereo satellite imagery (ZY-3 satellite), facilitating the construction of the continuous spatial surface model of elevation differences. Finally, the multi-temporal DSM datasets across the study area were corrected using either a global shift combined with local interpolation or purely local interpolation methods. Specifically, an elevation difference model is constructed using UAV LiDAR DSM as the high-precision reference and UAV image DSM as the target for correction. The process begins by estimating the global offset using the sample mean on a stratified hold-out training set. Subsequently, a decision on whether to perform local residual field modeling is made based on Moran’s I (p < 0.05 and I > 0.2). The preferred method is block-based RBF interpolation (block size 512, overlap 64, selecting ‘thin_plate’ if control points > 1000, otherwise ‘linear’), with robustness at the boundaries enhanced by adding 8 virtual boundary points (ratio = 0.05, estimated via IDW). If the RBF matrix is ill-conditioned or computational resources are limited, the method falls back to ordinary Kriging (with an exponential variogram model).For quality control, coordinate/coverage checks are performed, followed by outlier removal (|z-score| < 3; relaxed to |$∆Z$| ≤ 15 m for small samples) applied to the elevation differences and stratified hold-out partitioning based on Y-direction bins. For pixel-level application, a safety threshold is enabled (falling back to the original value if |correction amount| > 20 m) before generating the final output.

Regarding error propagation, statistics are computed for the distribution of errors within 1 m and 2 m tolerance limits, and the Mean Squared Error (MSE) is calculated to represent the overall uncertainty.

We used the MSE of elevation difference (σ) to assess the accuracy and reliability of elevation points. This evaluation involved calculating residuals between measured and true elevation values, typically expressed as standard deviation.

$$\sigma ={\displaystyle \frac{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2}{n}}$$

(1)

where $x_i$ denotes the elevation difference, $\overline{x}$ denotes the average of elevation difference.

2.2.3. Analysis of Changes in Gully Erosion Depth

The analysis of gully depth changes primarily relies on terrain data that have been corrected using a continuous spatial surface model of elevation differences. Specifically, the centerline data of gullies within the Sandu River watershed were first extracted by integrating hydrological analysis models in ArcGIS software with high-resolution optical remote sensing imagery. Subsequently, spatial analysis tools in ArcGIS software were used to assign elevation values from the corrected terrain data of two different periods to the gully centerline data. Finally, elevation differences were calculated to assess changes in gully depth in the Sandu River watershed from 2014 to 2024.

2.2.4. Identification of Gully Erosion Using Optical Remote Sensing Imagery

We used five sub-watersheds within the Sandu River watershed including the Huanglong River, Jimaligou, Cunpinggou, Zhangjiagou, and Zaotangou as sample areas for gully erosion. First, a 200 m buffer was generated by acquiring stream data based on the hydrologic analysis tool in the ArcGIS Pro 2.8 software. Erosion patches were then manually interpreted in these buffer zones by integrating erosion gully shape, texture, vegetation cover, feature changes, and other characteristics from the high-resolution optical images for the years 2014, 2019, and 2024. The minimum identification unit requirement for erosion patches is 4 × 4 pixels (based on GF-2 imagery), and its validation is based on UAV aerial survey data conducted on-site in October 2024. A total of 244 effective erosion gully patches were identified within the sample area, of which 8 erosion patches were attributed to change from 2014 to 2019, 150 to 2019–2024, and 86 to continuous change from 2015 to 2024. The total area of the patches was 92,191.49 m², the minimum patch was 16.44 m², the maximum patch was 9434.47 m², and the average area was 358.72 m². Meanwhile, each patch was turned into a point with geometric center, and the distance between its center point and the centerline of the river channel had an average of 19.42 m, a minimum of 0.24 m, and a maximum of 99.24 m.

The U-Net model was utilized for erosion identification (Figure 5), and the erosion areas were accurately identified through its powerful image segmentation capability. All deep learning models were implemented with the PyTorch framework. The proposed method was executed under Python 3.10.12, Torch 2.2.0, and CUDA 11.7, and all experiments were conducted on an NVIDIA Tesla A40 GPU. Sample selection was based on five typical sub-watersheds covering a variety of shapes, textures, and vegetation cover levels to ensure a diverse and representative sample. The data were derived from three periods of pre-processed high-resolution optical imagery in 2014, 2019, and 2024, resulting in 244 high-confidence gully erosion patches that were used to train, validate, and test the model. The sample allocation follows a ratio of 8:1:1 divided into training, validation and testing sets to ensure that the model gets full learning as well as generalization capabilities. The training, validation, and testing sets comprise 196, 24, and 24 samples, respectively. To ensure the model’s robustness in identifying erosion patches under varying rainfall regimes and vegetation conditions, we used precipitation data and Gaofen high-resolution imagery to assign the selected high-confidence patches to subregions representing distinct rainfall and vegetation environments within the study area. Specifically, during model training and testing, we employed stratified sampling so that high-confidence patches from each rainfall/vegetation environment were evenly distributed between the training and test sets. For the training of the model, we optimize it with a cross-entropy loss function, use the Adam optimizer (with the learning rate set to 0.001), and incorporate data augmentation techniques to improve the model’s robustness and resistance to overfitting. To prevent overfitting, we inserted Dropout layers in the U-Net decoder to reduce feature co-adaptation and enhance generalization, and we applied data augmentation (geometric and radiometric transforms) together with early stopping based on validation loss/IoU as standard regularization. The accuracy of the final model (i.e., the IoU metric on the test set) reaches 87%, showing excellent classification results. In addition, for the initial results output from the U-Net model, we applied Segment Anything Model (SAM) for post-processing, Parameters used during post-processing are summarized in

Table 2. Key parameters of the SAM.

Parameters Name	Final Parameters	Explanation
points_per_side	128	Defines sampling points along one image side.
pred_iou_thresh	0.86	Filters masks based on predicted quality, within the range [0, 1].
stability_score_thresh	0.92	Adjusts cutoff for stability score calculation.
crop_n_layers	1	Determines the number of image crop layers, where each layer includes 2i image subdivisions.
crop_n_points_downscale_factor	2	Downscales sampled points per side for layer n by a factor of 2n.
min_mask_region_area	80	Removes small regions and holes in masks smaller than the specified area.

. During SAM-based post-processing of erosion patches, we evaluated extraction performance using a confusion matrix, reporting the false-alarm rate (the proportion of tiles incorrectly identified as erosion patches) and the model’s false-positive rate. The SAM stage achieved an erosion-patch detection accuracy of 0.882, a false-alarm rate of 0.094, and an overall accuracy of 0.928. This process effectively eliminates the noise and misclassification of model prediction, optimizes the boundary fineness and accuracy of segmented patches, and thus further improves the accuracy and practicality of soil erosion monitoring.

2.2.5. Accuracy Evaluation Metrics

A confusion matrix is used to assess whether the predicted class labels are consistent with the reference land-cover types; it is a commonly used evaluation tool. In this study, we adopt mapping accuracy, overall accuracy (OA), and Cohen’s Kappa as the evaluation Mapping accuracy refers to the per-class accuracy, i.e., the ratio of correctly classified samples for a given class (the diagonal entry of the confusion matrix) to the reference total of that class (often termed the producer’s accuracy). It represents the probability that a reference sample of that class is correctly mapped.

Overall accuracy (OA) is the ratio of the number of correctly classified samples to the total number of samples:

$$OA={\displaystyle \frac{\sum _k N_{kk}}{N_{\mathrm{total} }}}$$

(2)

where N_kk is the number of correctly classified samples for class k, and N_total is the total number of samples.

Kappa measures the agreement between classification and reference data after accounting for chance agreement.

Given that our task involves only two classes (erosion vs. non-erosion), OA provides a clear and informative summary of the classification performance. The confusion matrix of the classification accuracy of the deep learning model in this study is shown in

Table 3. Confusion Matrix of Classification Accuracy of the Deep Learning Model.

Classification Results	Erosion (%)	Non-Erosion (%)	User’s Accuracy (%)
Erosion (%)	90.62	9.38	90.6
Non-erosion (%)	12.5	87.5	87.5
Producer’s Accuracy	0.8	0.91
	Overall Accuracy: 0.89		Kappa: 0.78

.

3. Results

3.1. Accuracy Assessment and Correction of Parameters

We first conducted aerial surveys using a LiDAR sensor-equipped UAV on the Huanglong River, Jimaligou, Cunpinggou, Zhangjiagou, and Zaotangou subwatersheds. We compared the elevation data obtained from GF7 and ZY3 imagery using the 2024 UAV aerial survey elevation data as a baseline. The results show that the median error in elevation between GF7-DSM and UAV data is 0.8433 m, while it is 4.2738 m between ZY3-DSM and UAV data, and 4.2401 m between GF7-DSM and ZY3-DSM data, indicating that the GF7-DSM data meets the requirement of accuracy, but there is an obvious systematic offset error in ZY3-DSM data. The ZY3-DSM data for 2014 and 2024 were then corrected based on a continuous model of elevation difference spatial surfaces. We further compare the corrected ZY3-DSM data with the UAV elevation data and GF7-DSM elevation data. We find that the median error in elevation between the corrected ZY3-DSM data and the UAV data is 1.7702 m, and the median error in elevation between the GF7-DSM and the ZY3-DSM data is 1.7395 m. This demonstrates that the systematic offset error of the ZY3-DSM data has been effectively corrected after the correction process.

Specifically, in the Huanglong River sub-watershed the mean square error of elevation difference (σ) from 2014 to 2024 was 1.83 m, and after correction the value was 0.99 m (Figure 6b), indicating a 46.13% reduction in elevation error. In the Jimaligou sub-watershed, the σ values dropped from 1.71 m to 1.28 m following adjustment (Figure 6c), corresponding to a 25.12% decrease. For the Cunpinggou, the mean square error of elevation difference initially measured 2.53 m and was reduced to 1.84 m post-correction (Figure 6d), indicating a 27.36% improvement in accuracy. Zhangjiagou exhibited a substantial reduction, with σ values declining from 3.10 m to 2.44 m (Figure 6e), amounting to a 21.35% decrease in error. Lastly, in Zaotangou sub-watershed, the σ decreased from 3.09 m to 2.34 m after correction (Figure 6f), reflecting a 24.11% reduction in elevation discrepancy. Overall, 805 validation points were selected in the Sandu River watershed, which in turn corrected the 2014 and 2024 DSM data. Elevation errors were reduced by 18.97% (σ = 2.95 for before; σ = 2.39 for after) in the Sandu River watershed (Figure 6a).

The corrected DSM results for the Sandu River watershed in 2014 and 2024 are shown in Figure 7. The corrected DSM results indicate that the elevation range in the Sandu River watershed was 1069–2474 m in 2014 and 1091–2481 m in 2024. The minimum and maximum elevation of the DSM show slight increases in the whole watershed, which may be associated with channel sedimentation and vegetation recovery in the watershed. The DSM elevations in the northern region are higher than those in the southern region, while the western part shows greater elevations compared to the eastern part. This is closely associated with the topographic distribution of the Sandu River watershed, where higher mountains dominate the northern and western parts of the watershed, while the downstream southern and eastern areas feature a denser drainage network (Figure 1).

3.2. Dynamic Variation of Gully Erosion Depth

The difference in erosion depth at each pixel along the central line of gullies in the Sandu River watershed between 2014 and 2024 is illustrated in Figure 8. Over the past decade, the average gully erosion depth in the Sandu River watershed was 2.6m, with an annual rate of 0.26m yr⁻¹ and a maximum observed depth of 19.3 m (Figure 8a,b). Gully erosion has reduced the average elevation in the Sandu River watershed from 1615.5 m in 2014 to 1612.9 m in 2024 (Figure 8c). Topographic analysis of the Sandu River watershed between 2014 and 2024 documented measurable elevation declines across all monitored sub-watersheds, demonstrating distinct spatial patterns in erosional response (Figure 8d–h). The Huanglong River in the northern part of the watershed decreased from 1819.1 m to 1815.9 m (−3.2 m, Figure 8d), while the Jimaligou in the western part declined from 1879.4 m to 1876.5 m (−2.9 m, Figure 8e). In the eastern part of the watershed, three adjacent sub-watersheds exhibited distinct elevation changes: Cunpinggou declined from 1314.0 m to 1311.6 m (−2.4 m, Figure 8f), Zhangjiagou showed a dramatic drop from 1283.0 m to 1275.0 m (−8.0 m, Figure 8g), and Zaotangou experienced the smallest reduction from 1211.8 m to 1209.8 m (−2.0 m, Figure 8h). Examining the watershed-scale DEM of Difference (DoD) (Figure 9b) shows that the Sandu River watershed experienced pronounced soil erosion over the past decade. To further evaluate the impact of the most recent extreme rainfall, we rectified the 2023 ZY-3 satellite DSM using UAV-derived control points and generated DoD maps for the pre- and post-storm periods (Figure 9a,b). Statistical analysis indicates that the mean erosion depth associated with the 2023 event was 0.36 m, with a maximum observed depth of 3.34 m (Figure 9c).These spatially heterogeneous changes demonstrate varying erosion susceptibilities among the watershed systems, which may be closely associated with differences in natural environmental conditions and anthropogenic disturbance intensities across sub-watersheds [13].

The spatial distribution of erosion patches over the past decade is shown in Figure 10. Widespread elevation decline occurred across the Sandu River watershed from 2014 to 2024, exhibiting spatially heterogeneous surface lowering magnitudes. The eastern watershed exhibits greater elevation reduction compared to the western watershed, which may be related to the regional topographic gradient descending from west to east. The larger depth of gully erosion is primarily in the central part of the watershed, with localized maximum elevation losses exceeding 100 m. Compared to vegetated areas, loess-covered regions exhibit greater elevation loss and higher susceptibility to gully erosion, consistent with existing research [38]. Erosion depth is greater along terrace bases, likely results from concentrated runoff imposing enhanced shear stress on bottom soils [45].

3.3. Drivers of Topographic Variations

The Sandu River watershed experienced an intricate change in elevation from 2014 to 2024, with the spatial heterogeneity pattern of surface deformation influenced by a variety of factors, including gully erosion, vegetation restoration, mining activities and terracing (Figure 11). Gully erosion remains the main natural driver of elevation change (Figure 11a), especially in the steep gullies of Huanglong River and Zhangjiagou sub-watersheds, where annual rates of lowering reach 0.18–0.34 m/year. DSM variations reveal that these erosion hotspots account for about 20–38%% of the entire area. The erosion process exhibits strong seasonal characteristics, with particularly heavy rainfall events triggering extensive slope damage and gully retreat. However, large-scale vegetation restoration efforts (Figure 11b), such as Grain for Green Project and slope protection, may offset soil erosion, thereby resulting in a net elevation increase of 0.19–0.55 m per year. Our analyses of vegetation cover and elevation changes indicate significant differences in elevation trajectories between areas with increased vegetation cover and those without restoration. Compared to cropland, forest land represents the most effective soil retention and can significantly reduce gully erosion. Human activities exhibit contrasting effects on elevation changes, particularly mining activities and terracing. Open-pit mining operations have typically resulted in localized depressions that have lowered elevations by 1.52–8.92 m over the past decade, while disposal of waste soil and waste rock stockpiles in adjacent areas have resulted in artificial elevations of 3.44–8.56 m (Figure 11c). Traditional terrace leveling and maintenance activities have produced significant elevation changes (Figure 11d), resulting in a net elevation loss of 0.19–0.77 m per year.

4. Discussion

4.1. Potential and Uncertainty for Estimating Gully Erosion from Multi-Source Remote Sensing Imagery

Combining multi-source high-resolution remote sensing images to estimate and monitor gully erosion has significant advantages, especially in the Loess Plateau where the environment is more complex. Such intricate landscapes present limitations to traditional field-based observation methods, such as scale and time constraints. We used DSM from GF7 and ZY3 stereo satellite imagery to estimate gully erosion depths and corrected them using UAV data. We also used high spatial resolution GF2 imagery to delineate erosion regions. It should be noted that all correction models for the ZY-3 satellite DSMs in this study used 2024 UAV ground-truth data. Although the control points selected from the UAV data ensure that no planimetric (geometric) changes occurred during the 2014–2024 observation period, they cannot preclude vertical subsidence. Subsidence at control points may therefore have a potential impact on the accuracy of the ZY-3 DSM correction [46]. This error source has limited influence on our basin-scale assessment of erosion trends and on the detection of extreme-rainfall-induced erosion events using domestic high-resolution satellites combined with UAV data, but it introduces uncertainty into high-precision estimates of erosion rates. Consequently, we recommend acquiring pre-event UAV data to calibrate large-area DSMs for the pre-event epoch, and reacquiring post-event high-resolution DSMs to calibrate the large-area DSMs for the post-event epoch. Moreover, the corrected ZY-3 DSM shows a median vertical error of 1.77 m relative to the UAV data, which is comparable to the mean annual erosion depth of 2.6 m. This level of uncertainty warrants caution in interpreting local erosion rates, especially when the observation window is short. Nevertheless, the proposed framework—integrating high-resolution satellite imagery with UAV data—remains robust for basin-scale estimation of erosion trends and is particularly effective at capturing erosion events triggered by extreme rainfall. Despite being implemented only in the Sandu River watershed, this method demonstrates the synergistic potential of combining multi-source remote sensing sensors for gully erosion monitoring. The DSM generated by GF7 and ZY3 provides detailed topographic information that accurately quantifies surface elevation changes associated with gully formation and development, which is more subtle when quickly identifying gully erosion due to heavy precipitation events. Following ground control and calibration, the ZY3 imagery can provide high vertical accuracy. Combining ZY3 and GF7 on the temporal scale can effectively capture the elevation distribution and fine-scale surface features for the study area over the past 10 years. Such accuracy mapping is critical to estimate gully erosion depths accurately, especially in areas where elevation changes are driven by a combination of natural and human activities. The acquisition of UAV data is essential to bring the different data sources into a harmonized framework. GF2 imagery with sub-meter spatial resolution enhances the identification and mapping of erosional features, including the depiction of gully boundaries, and especially the monitoring of gully erosion that has been overlooked in coarse datasets. Moreover, using multi-source remote sensing imagery addresses several challenges inherent to single-sensor data. For example, the limited temporal coverage of single-sensor data, and the fact that optical imagery is affected by vegetation cover or shading effects. We also employed advanced image analysis techniques, such as deep learning techniques and Segment Anything Model, to further improve the accuracy of gully detection with multi-source inputs. Compared with approaches that use MODIS and Landsat to assess soil-erosion intensity [47], the systematic framework developed here—integrating domestic high-resolution satellite imagery with UAV data—likewise demonstrates the central role of remote sensing in quantifying erosion dynamics. Moreover, unlike the conventional erosion-rate unit (t·km⁻²·a⁻¹), our DSM-based estimates are expressed in m·yr⁻¹, providing an intuitive, elevation-change–centric perspective that captures periods of intense erosion and reveals the impacts of soil- and water-conservation measures (e.g., Grain for Green) on soil loss across the Loess Plateau [48].

4.2. Erosion Drivers and Human Impacts

The observed topographic changes in the Sandu River watershed are driven by complex interactions between natural processes and anthropogenic disturbances. Soil erosion, as a common phenomenon in nature, is the process of destruction, denudation, transport and deposition of soil and its soil-forming parent material by external forces such as water, wind, gravity and human activities. Dynamic erosion pattern continuously drives the micro-variation in surface elevation [49]. From the statistics of ten years of heavy rainfall data and the elevation differences in the watershed before and after the most recent storm, it can be determined that rainfall plays an important role in erosion within the Sandu River basin, with extreme rainfall events directly causing rapid soil loss [50]. At the same time, due to the relatively low resolution of available rainfall data, the complete quantitative relationship between rainfall and erosion cannot be derived directly from the data. However, by combining information such as the frequency of heavy rainfall events and annual cumulative precipitation with soil erosion rates and severity, it can be roughly inferred that water erosion is one of the dominant factors driving soil erosion processes in the Sandu River basin. During high-intensity rainfall events, concentrated surface runoff develops into linear flow paths that exert shear stress on the soil, initiating gully erosion features ranging from rills to ephemeral gullies and ultimately permanent gullies [51,52]. As erosion progresses, the channels expand both vertically and laterally, creating a rough terrain with alternating highs and lows. In addition, gravitational erosion induces instability and downslope movement of weathered debris or unstable rock-soil masses on slopes, triggering geological hazards such as collapses and landslides [53]. This process creates localized depressions and protrusions on mountain surfaces, thereby altering topographic relief. As slopes steepen during erosion, the associated increase in overland-flow shear stress further promotes the co-mobilization of coarse and fine particles; once a critical slope threshold is exceeded, erosion rates escalate rapidly [54]. Slope form and length also exert strong controls: convex slopes exhibit significantly higher erosion rates than concave slopes [55]—a contrast that is readily captured and evaluated using UAV-calibrated, high-precision DSMs. From a soil-type perspective, erosion in sandy soils rises sharply with slope and then declines due to surface armoring by coarse particles, whereas in silt/clay soils the erosion rate continues to increase with slope [56]. These topographic and edaphic controls provide important guidance for targeted soil- and water-conservation measures across different sub-basins.

Vegetation cover significantly mitigates gully erosion in the Sandu River watershed, thereby stabilizing surface topography and reducing terrain dissection. Under precipitation conditions, vegetation roots and stems play different roles in the process of runoff erosion, altering the runoff process while increasing resistance to soil erosion [57,58]. Plant roots can effectively increase soil cohesion and consolidate soil, reducing erodibility while improving soil stability, thereby mitigating landslide and collapse risks and contributing to terrain smoothing [59,60]. Vegetation canopy interception promotes gradual rainwater infiltration while suppressing rapid surface runoff, thereby reducing flow erosivity and soil erosion potential [61]. Furthermore, vegetation growth could curb gully evolution during erosion by hindering soil transport, reducing soil loss rates, and increasing critical shear stress [62].

Compared to natural processes, anthropogenic activities may lead to faster rates in reshaping surface topography [63,64]. Open-pit mining, as one of the prominent anthropogenic activities in the Sandu River watershed, can rapidly and significantly alter surface topography within a relatively short period. The open-pit mining process excavates and destroys the original landform, while waste dumps overlay and compress it, thereby forming a new special landform type characterized by depressions and elevated mounds. Moreover, open-pit mining exerts severe ecological impacts on land by altering vegetation, soil composition, and subsurface geological structures [65,66]. These modifications consequently induce changes in surface hydrology, groundwater levels, and flow pathways, ultimately exacerbating the extent of land subsidence [67,68].

The implementation of soil and water conservation measures is one of the main anthropogenic factors contributing to topographic changes in the watershed. The Sandu River watershed is almost entirely covered by loess, with sparse vegetation and high sediment load in the river. To address the harsh natural conditions and severe soil erosion in the watershed, comprehensive soil and water conservation measures, primarily involving the conversion of sloping land to terraces, have been implemented since the 1970s. Terracing, as an effective measure to enhance the soil water holding capacity, can thoroughly alter landforms by transforming natural slopes in to stair-steps of flat surfaces. Compared to natural slopes or terrace risers, terracing significantly reshapes the original slope morphology, including gradient, length, curvature, and surface roughness [69]. This transformation modifies hydrologic connectivity, rebalances runoff and infiltration patterns in watersheds, and generates micro-watersheds to improve rainwater harvesting efficiency [70]. The implementation of diverse terracing measures (e.g., level ditch, slope, counter-slope, and half-moon terraces) improved watershed conditions, thereby increasing the survival rate of planted vegetation such as trees, shrubs or grasses [71]. In this way, even very little precipitation can be infiltrated efficiently, avoiding runoff, achieving water storage and reducing the quantity of sand [72].

4.3. Limitation and Future Works

There are several limitations in this study. First, the estimation of gully erosion depths depends mainly on DSM from GF-7 and ZY-3 imagery, which are limited by the inherent spatial resolution and vertical accuracy of these data. These limitations can lead to underestimation or overestimation of subtle topographic changes, especially in regions with complicated topography or dense vegetation cover. Second, although UAV-derived data were used to calibrate the ZY3 DSM, the limited spatial coverage of the UAV survey may introduce localized biases in the calibration process. Third, the identification of erosion areas based on GF-2 images is affected by the spectral and spatial features of the sensors, which may not capture all the microscale erosion features or distinguish between erosion and other surface disturbances. Finally, matching between different datasets can affect the accuracy of change detection and erosion assessment. Future efforts should consider the integration of more temporal datasets, as well as advanced data fusion techniques to improve the robustness of erosion monitoring in different environments.

Moreover, this study treats rainfall as the primary driver of erosion. However, the precipitation products obtained from the China Meteorological Administration are based on all basic and ordinary stations in and around the Sandu River watershed, and some monthly records are missing. As a result, we were unable to conduct a more detailed basin-wide, fine-scale spatiotemporal analysis of rainfall–erosion relationships, even though the interaction between local topography and rainfall is crucial for understanding rainfall-induced erosion [73]. To further improve the accuracy of rainfall-erosion attribution, our framework can incorporate additional, higher-fidelity precipitation datasets from other sources, enabling analyses at shorter temporal scales. In parallel, because the proposed erosion-estimation scheme relies on domestic satellites and UAVs, the elevation model used here can be further refined—along with calibration of the water-erosion process model—by integrating higher spatiotemporal-resolution 3D elevation data [74]. In addition, different soil physical properties have important impacts on erosion. The Sandu River Basin is dominated by Huangmiantu (loessal) and Heilutu soils (Figure 12a); Huangmiantu is generally more prone to erosion, whereas Heilutu—with relatively higher organic matter and clay contents and a more stable structure—shows greater resistance [75]. Spatial maps show that individual soil properties correlate with erosion rates (Figure 12), but higher-precision analysis requires integrating multiple soil properties. The watershed exhibits a con together with other factors. As a next step, we can perform a quantitative analysis linking soil types with high-precision erosion-rate estimates to examine how erosion rates vary across soil types under heavy-precipitation events and different human influences.

4.4. Methodological Extensions to Other Fields

This study develops a systematic approach that uses UAV data to calibrate domestic high-resolution satellite imagery and then evaluates soil and water erosion based on DSM products. Our results demonstrate that UAV-based correction of satellite-derived DSMs is effective. Accordingly, the framework can be generalized beyond erosion applications. For example, in the Heihe River Basin, DSMs can support vegetation-structure surveys of Populus euphratica, where elevation information helps distinguish degradation states of trees versus shrubs [76]; on the Loess Plateau, DSMs can facilitate large-area cropland mapping, as crop canopy heights exhibit pronounced phenological (seasonal) variations [77]. The approach can be further enhanced by developing low-altitude, cooperative unmanned intelligent systems to achieve more efficient, automated, high-precision DSM calibration—for instance, synchronizing UAV acquisitions with ground-based unmanned platforms that conduct concurrent intelligent 3D surveys. These possibilities highlight the method’s broad applicability and substantial potential for advancement.

5. Conclusions

This study employed multi-source high resolution remote sensing imagery and deep learning methods to identify gully erosion in the Sandu River watershed and further investigated the characteristics of topography changes over the past decade under the combined effects of erosion and other environmental and anthropogenic factors. The main findings of our study can be summarized as follows:

Topographic monitoring in the Sandu River watershed revealed a net elevation decrease of 2.6 m (from 1615.5 m in 2014 to 1612.9 m in 2024) attributable to gully erosion processes during the study period. More than 50% of the watershed area experienced measurable elevation reduction, with varying magnitudes of surface lowering. Across the monitored sub-watersheds, the measured gully erosion depths across sub-watersheds varied from 2.0 m to 8.0 m of elevation loss, with the maximum reduction (−8.0 m) observed in Zhangjiagou, indicating higher erosional dynamics.

The complex elevation changes in the Sandu River watershed result from coupled natural and anthropogenic drivers. Gully erosion drives the most severe natural elevation reduction, with localized surface lowering reaching 0.18–0.34 m yr⁻¹ in steep gullies, particularly triggered by extreme precipitation events. Conversely, large-scale vegetation cover can enhance surface stability with measurable net elevation gains of 0.19–0.55 m yr⁻¹, demonstrating that increased vegetation cover serves as an effective measure for erosion control. In addition, human activities exhibit bifurcated impacts on elevation changes. Mining operations concurrently produce localized depressions (1.52–8.92 m depth) and adjacent artificial fills (3.44–8.56 m depth), while terrace maintenance results in systematic net losses of 0.19–0.77 m yr⁻¹. This study provides further insights into the risk of soil erosion in ecologically fragile regions and highlights the urgent need for coordinated landscape management.