Automatic mapping of gully from satellite images using asymmetric non-local LinkNet: A case study in Northeast China

Abstract

Gully erosion can lead to the destruction of farmland and the reduction in crop yield. Gully mapping from remote sensing images is critical for quickly obtaining the distribution of gullies at regional scales and arranging corresponding prevention and control measures. The narrow and irregular shapes and similar colors to the surrounding farmland make mapping erosion gullies in sloping farmland from remote sensing images challenging. To implement gully erosion mapping, we developed a small training samples-oriented lightweight deep leaning model, called asymmetric non-local LinkNet (ASNL-LinkNet). The ASNL-LinkNet integrates global context information through an asymmetric non-local operation and conducts multilayer feature fusion to improve the robustness of the extracted features. Experiment results show that the proposed ASNL-LinkNet achieves the best performance when compared with other deep learning methods. The quantitative evaluation results in the three test areas show that the F1-score of erosion gully recognition varies from 0.62 to 0.72. This study provides theoretical reference and practical guidance for monitoring erosion gullies on slope farmland in the black soil region of Northeast China.

Keywords

• Gully erosion mapping
• High-resolution satellite images
• Deep learning
• Black soil