KDANet: a farmland extraction network using band selection and dual attention fusion – a case study of paddy fields and irrigated land in Qingtongxia, China

Food security is a critical global issue, and accurate monitoring of farmland is essential to address this challenge. However, farmland extraction accuracy is often hindered by variations in land types and scales. We propose the Kolmogorov–Arnold Dual Attention Multi-Feature Fusion Network (KDANet), which combines band selection with dual attention mechanisms to overcome these challenges. Using the SEaTH-Span algorithm, phenological image bands specific to farmland types are selected to optimize data effectiveness. KDANet extracts local and global features through dual attention, with linear Kolmogorov-Arnold Networks (KAN) fusing them to improve accuracy and stability. Based on Sentinel-2 imagery of Qingtongxia, KDANet significantly outperformed mainstream methods such as DeepLabV3, U-Net, and SegFormer, achieving Precision gains of 11.5% and 8.5% and Intersection over Union (IoU) improvements of 3.5% and 4.3% for paddy fields and irrigated land. This method provides robust support for intelligent farmland monitoring and food security. Code is released at https://github.com/Chen-Qi-005/KDANet.