TL;DR
This paper introduces a novel framework for precisely mapping gully erosion using a fusion of high-resolution satellite imagery and UAV LiDAR data. By calibrating satellite-derived elevation models with UAV data and employing deep learning for erosion patch identification, the study quantifies erosion dynamics over a decade, revealing the significant impact of extreme rainfall and human activities on landscape alteration in the Loess Plateau.
What Problem Does This Paper Address?
Gully erosion is a major driver of soil degradation and landscape change, particularly in regions like the Loess Plateau, which are susceptible to extreme rainfall events. Accurately quantifying this erosion, especially at a fine scale and over long periods, is challenging. Traditional methods are often labor-intensive and limited in spatial coverage. While remote sensing has advanced, existing approaches struggle to capture the detailed 3D structural parameters of gullies and integrate multi-source data effectively. This paper aims to bridge this gap by developing a robust, scalable methodology for precisely estimating gully erosion, considering both natural and anthropogenic influences.
Methods
The study employs a multi-stage approach:
- Digital Surface Model (DSM) Generation and Calibration:
- Initial DSMs were generated from stereoscopic imagery of China's Ziyuan-3 (ZY-3) and GaoFen-7 (GF-7) satellites.
- These DSMs were then calibrated and enhanced for elevation accuracy using high-resolution Unmanned Aerial Vehicle (UAV) photogrammetry and LiDAR data. This step is crucial for improving the precision of elevation difference measurements.
- A Radial Basis Function (RBF) interpolation method was used for DSM correction, leveraging UAV data as a high-precision reference.
- Gully Erosion Patch Identification:
- A deep learning framework, specifically a U-Net model, was trained on high-resolution GaoFen-1 (GF-1) and GaoFen-2 (GF-2) imagery to automatically identify and delineate erosion patches.
- The Segment Anything Model (SAM) was used as a post-processing step to refine the accuracy of the identified erosion patches, improving boundary precision.
- Analysis of Erosion Dynamics and Drivers:
- Corrected DSMs from different time points (2014 and 2024) were used to quantify gully erosion depths.
- The study analyzed the influence of natural factors (gully erosion, vegetation restoration) and anthropogenic activities (mining, terracing) on elevation changes within the watershed.
Datasets
- Satellite Imagery:
- Ziyuan-3 (ZY-3): Used for initial DSM generation (2m resolution).
- GaoFen-7 (GF-7): Used for initial DSM generation (0.8m resolution).
- GaoFen-1 (GF-1) / GF1D: Used for erosion patch identification (2m resolution).
- GaoFen-2 (GF-2): Used for erosion patch identification (0.8m resolution).
- GaoFen-6 (GF-6): Used for general imagery (2m resolution).
- UAV Data:
- Digital orthophoto map (0.2m resolution) and LiDAR data: Used for calibrating satellite-derived DSMs.
- Meteorological Data:
- Daily precipitation data from 9 stations (2014-2024) sourced from the China Meteorological Administration.
Key Insights and Ideas
- Synergistic Power of Multi-Source Data: The study highlights that combining high-resolution satellite data with UAV-derived ground truth is essential for accurate, large-scale gully erosion monitoring. Satellite data provides broad coverage, while UAV data offers the necessary fine-scale precision.
- UAV Data as a Critical Calibrator: The research demonstrates that UAV data is indispensable for correcting systematic errors in satellite-derived DSMs, significantly improving the reliability of elevation change measurements.
- Deep Learning for Precise Erosion Mapping: The application of U-Net and SAM models proves highly effective in accurately identifying and delineating subtle erosion patches, overcoming limitations of traditional image analysis methods.
- Bifurcated Impacts of Human Activities: The study reveals that human activities like mining and terracing have complex and often contrasting effects on topography, leading to both depressions and artificial fills, and altering natural erosion patterns.
- Vegetation Restoration as a Mitigation Strategy: Evidence suggests that large-scale vegetation restoration efforts can effectively offset soil erosion, leading to net elevation gains and improved surface stability.
Key Results
- Net Elevation Loss: The Sandu River watershed experienced a net elevation loss of 2.6 meters between 2014 and 2024, with erosion depths reaching up to 8 meters in some sub-watersheds.
- DSM Correction Accuracy: The median error in elevation between corrected ZY3-DSM and UAV data was 1.7702 m, indicating effective correction of systematic errors. Elevation errors in the Sandu River watershed were reduced by an average of 18.97% after correction.
- Gully Erosion Depth: The average gully erosion depth over the past decade was 2.6m, with an annual rate of 0.26m/year. The maximum observed depth was 19.3m.
- Erosion Patch Identification Accuracy: The U-Net model achieved an IoU metric of 87% on the test set. Post-processing with SAM resulted in an erosion-patch detection accuracy of 0.882 and an overall accuracy of 0.928.
- Drivers of Erosion: Gully erosion, particularly triggered by extreme precipitation, was identified as the most severe natural driver of elevation reduction (0.18–0.34 m/year). Vegetation restoration led to net elevation gains (0.19–0.55 m/year). Mining operations caused localized depressions (1.52–8.92 m) and artificial fills (3.44–8.56 m), while terracing resulted in net losses (0.19–0.77 m/year).
Limitations and Future Work
- Limitations:
- The inherent spatial resolution and vertical accuracy of satellite DSMs (GF-7, ZY-3) can lead to underestimation or overestimation of subtle topographic changes, especially in vegetated areas.
- The limited spatial coverage of UAV surveys might introduce localized biases in calibration.
- GF-2 imagery's spectral and spatial features may not capture all microscale erosion features or distinguish them from other surface disturbances.
- Matching between different datasets can affect change detection accuracy.
- The precipitation data used had limitations, including missing monthly records and lower spatial resolution, hindering detailed basin-wide rainfall-erosion relationship analysis.
- Future Work:
- Acquire pre-event UAV data to calibrate large-area DSMs and reacquire post-event data for post-event calibration.
- Integrate more temporal datasets and advanced data fusion techniques for improved erosion monitoring robustness.
- Incorporate higher-fidelity precipitation datasets and analyze rainfall-erosion attribution at shorter temporal scales.
- Refine elevation models and water-erosion process models by integrating higher spatiotemporal-resolution 3D elevation data.
- Conduct quantitative analysis linking soil types with high-precision erosion-rate estimates to examine how erosion rates vary across different soil types and under various influences.
- Develop low-altitude, cooperative unmanned intelligent systems for more efficient and automated DSM calibration.
Summary updated