EAGNet: A method for automatic extraction of agricultural greenhouses from high spatial resolution remote sensing images based on hybrid multi-attention
Hongzhou Li, Yuhang Gan, Yujie Wu, Li Guo
Abstract
The timely and accurate acquisition of greenhouse information is crucial for strategically planning modern agriculture. However, existing methods are affected by the close spacing between agricultural greenhouses, intra-class diversity, and inter-class similarity, resulting in missed and incorrect extraction phenomena. Here, we propose a model for agricultural greenhouse extraction (i.e., EAGNet), which includes a residual block improvement module (RBIM) and boundary segmentation module (BSM) that solve the problem of densely distributed agricultural greenhouse-boundary adhesion. We constructed a class attention module (CAM) to address the leakage extraction phenomenon in agricultural greenhouses caused by intra-class diversity and introduced an object contextual representation module (OCRM) to address the incorrect extraction of agricultural greenhouses caused by the similarity between classes. Experiments on a self-made agricultural greenhouse dataset showed that EAGNet achieved the best extraction results among all compared methods.