Extraction of Terraces Across Representative Regions Using an Improved DeepLabv3+ Network

Accurate mapping of terraces is important for farmland protection, food security, and the management of terrace abandonment. However, extracting terraces from high-resolution remote sensing imagery remains challenging because of complex backgrounds, fragmented ridge structures, weak boundary contrast, and high similarity between terraces and surrounding land-cover types. This study proposes AMF-Deep, an Attention Multi-scale Fusion DeepLabv3+ model, for binary terrace extraction across representative terrace regions and morphologies. AMF-Deep was evaluated using 0.61 m high-resolution remote-sensing imagery from five representative terrace regions in China, encompassing diverse terrace morphologies and background conditions. In the proposed model, MobileNetV2 is adopted as the backbone to reduce computational cost, a modified multi-scale ASPP module is used to capture terrace morphology at different spatial scales, CBAM enhances low-level boundary and texture features, ECA refines high-level channel responses before ASPP, and NAM recalibrates semantic features after ASPP. A residual decoder is further introduced to support low-level feature transfer and boundary recovery. Experimental results show that AMF-Deep achieved the highest regional performance in Huangling, with mIoU, mPA, Accuracy, Recall, and Precision of 89.47%, 95.17%, 96.70%, 94.99%, and 93.68%, respectively. Stepwise ablation experiments further indicated that the modified ASPP module, the differentiated attention arrangement, and the residual decoder contributed to the final performance. Qualitative seasonal visualization suggests that winter images and dryland terraces in Shexian remain more challenging because of weak texture contrast and high background similarity. These results indicate that AMF-Deep has potential for terrace extraction across representative regions, although further quantitative seasonal evaluation and cross-region validation are still needed.