Integrating deep learning and agricultural remote sensing technology has made detecting changes in arable land a key technology for protecting arable land and food security. However, existing methods still face challenges such as crop phenological differences, intra-class confusion caused by fragmented arable land and low efficiency in multi-scale feature fusion. This paper proposes a dual-stream adaptive interaction and attention fusion network (DANet). First, a bidirectional cross-level edge fusion module (BCEDF) is designed to alleviate tiny target missed detection and edge blurring through multi-level residual and edge pyramid fusion strategies. Second, a difference-aware attention calibration module (DAAC) is con-structed to suppress the false detection of similar crops using dual-phase difference localization and cross-attention mechanisms. Finally, a dynamic hierarchical interaction fusion module (DHBIF) is designed to achieve cross-scale dynamic fusion of local and global features, enhancing the collaborative capability of multi-scale features. In addition, this study constructed a high-resolution farmland change detection dataset, GFSWCLCD. Experimental results show that DANet achieves F1 scores of 74.30% and 81.98% on CLCD and GFSWCLCD, significantly outperforming state-of-the-art methods.