Gully-ERFNet: a novel lightweight deep learning model for extracting erosion gullies in the black soil region of Northeast China

Abstract

Gully erosion is a critical environmental issue in the black soil region of Northeast China, causing severe soil degradation and posing significant threat to food security. Accurate gully extraction is essential for implementing effective control measures to mitigate environmental and agricultural impacts. Traditional machine learning methods rely on high-resolution digital elevation models (DEMs) but are hindered by data acquisition challenges and low automation levels. To resolve this issue, we introduce the Gully-ERFNet model, designed to extract gullies without high-resolution DEMs. Based on the deep learning ERFNet architecture, the Gully-ERFNet model integrates the Squeeze-and-Excitation attention mechanism to expand the global receptive field and the Pyramid Pooling Module to improve the model’s emphasis on gully features. We employed the Adaptive Correction Loss function to reduce the impact of incorrect labels on model performance. The model’s performance was evaluated using GF-2 satellite remote-sensing images of the Hailun Basin. The model achieved an F1-score of 81.50%, recall rate of 76.24%, and accuracy of 87.54% in gully extraction, outperforming existing gully extraction algorithms, such as Random Forest, U-Net, and DeeplabV3+. This model enables efficient gully extraction using only remote sensing imagery, addresses difficulties in data collection, and provides support for preventing and controlling gully erosion.