PFMANet: A Progressive Multi-Level Feature Aggregation Network for Cropland Change Detection

Cultivated land is a critical resource and foundation for food production, and the automatic and efficient extraction of cultivated land change detection information is of significant research importance for land conservation. Existing deep learning-based methods for cultivated land change detection have not adequately addressed the issues of feature misalignment and information redundancy in multilevel feature fusion, which hinders further improvements in model performance. To alleviate these issues, this paper designs a Progressive Multilevel Feature Aggregation Network (PFMANet) for cultivated land change detection. The network narrows the semantic gap by fusing adjacent features and employs a multilevel residual dual-attention feature enhancement module to strengthen local features, adaptively emphasizing key information across both spatial and channel dimensions, thus generating more detailed change detection results. Experimental results on the self-constructed Leshan dataset demonstrate that our method outperforms several existing change detection approaches, including BIT, BaseTransformer, MCTNet, and RDPNet. Key quantitative metrics, such as Precision and IoU, show improvements ranging from $\mathrm{3. 2 6 \%}$ to $\mathrm{1 3. 2 \%}$ and $\mathrm{1. 4 3 \%}$ to $\mathrm{1 0. 0 \%}$, respectively.