Gully Extraction in Northeast China’s Black Soil Region: A Multi-CNN Comparison with Texture-Enhanced Remote Sensing

Gully erosion poses a serious threat to soil fertility and agricultural sustainability in Northeast China’s black soil region. Accurate and efficient mapping of erosion gullies is critical for enabling targeted soil conservation and precision land management. In this study, we developed a texture-enhanced deep learning framework for automated gully extraction using high-resolution GF-1 and GF-2 satellite imagery. Key texture parameters—specifically mean and contrast features derived from the gray-level co-occurrence matrix (GLCM) under a 5 × 5 window and 32 gray levels—were systematically optimized and fused with multispectral bands. We trained and evaluated three convolutional neural network architectures—U-Net, U-Net++, and DeepLabv3+—under consistent data and evaluation protocols. Results demonstrate that the integration of texture features significantly enhanced extraction performance, with U-Net achieving the highest overall accuracy (90.27%) and average precision (90.87%), surpassing DeepLabv3+ and U-Net++ by margins of 6.06% and 9.33%, respectively. Visualization via Class Activation Mapping (CAM) further confirmed improved boundary discrimination and reduced misclassification of spectrally similar non-gully features, such as field roads and farmland edges. The proposed GLCM–CNN integrated approach offers an interpretable and transferable solution for gully identification and provides a technical foundation for large-scale monitoring of soil and water conservation in black soil landscapes.