Large-scale agricultural greenhouse extraction for remote sensing imagery based on layout attention network: A case study of China

Abstract

Rapid and accurate agricultural greenhouse extraction with remote sensing imagery is essential for providing spatial information for precision agriculture. Benefiting from local spatial perception, deep learning based object extraction methods have achieved satisfactory performances in extracting geo-objects. However, they fail to work for large-scale greenhouse extraction since the objects are sparsely distributed in the background and densely distributed in the foreground, where the local spatial perception causes redundant computation and false detection problems. In this paper, we propose a layout attention network (LANet) framework for large-scale greenhouse extraction using remote sensing imagery, which replaces the local spatial perception with spatial layout perception, i.e., a sparse global layout to identify the sparse background and a dense local layout to identify the dense foreground. To address the shortcoming of the sparse background, which leads to redundant computation, a sparse global layout awareness module is formulated as a scene classifier. This accommodates the global layout attention map of the global scene features by adopting a layout-shared convolutional neural network (CNN) backbone for generating class-agnostic layout priors and global channel attention for aggregating discriminative global layout features, ensuring robust sparse background identification. Then, to alleviate the problem of the dense foreground, which causes false detection, a dense local layout awareness module is proposed to incorporate the local layout attention map and rotated region of interest (RRoI) features. The RRoI features are then further embedded to guide the initial RRoIs for object location refinement by aligning the initial RRoI locations in a layout-sensitive attention mechanism and achieving semantic enhancement by taking the local layout density as a semantic prior to assign a reliable class score map. The experimental results obtained on an agricultural greenhouse benchmark dataset and a large-scale agricultural greenhouse extraction dataset illustrate that the proposed framework can outperform the state-of-the-art object extraction methods in both speed and accuracy, and has a high generalization ability for large-scale dense object extraction.