MADNet: cropland change detection network for the complex terrain and dense vegetation hilly region in the Southwestern China

Detecting cropland changes in complex hilly terrains presents significant challenges due to the irregular shapes, small sizes, and fragmented distributions of the land. To address these challenges, we introduce MADNet, a Multi-Scale Attention and Dynamic Sampling Network specifically designed for such environments. MADNet leverages RFAConv to dynamically adjust receptive fields, thereby enhancing the capture of edge and texture features. The LSKA-enhanced ASPP module extracts multi-scale features, while the EMA attention mechanism reinforces feature representation. Additionally, dynamic upsampling using DySample efficiently reconstructs high-resolution feature maps. Experimental results on the SHCD, luojiaSET-CLCD, and CLCD datasets demonstrate MADNet's superior performance, achieving F1 scores of 82.06%, 70.98%, and 68.76%, respectively. We also introduce SHCD (https://github.com/LeeJiEunx/SHCD.git), a novel dataset specifically designed for hilly cropland change detection, which aims to advance research in this domain. Our findings underscore MADNet’s potential in supporting decision-making for agricultural policy formulation and land resource management.