Edge-aware and spectral–spatial information aggregation network for multispectral image semantic segmentation

Semantic segmentation is a fundamental task in the field of remote sensing image intelligent interpretation and computer vision. Multispectral images have attracted more researchers’ attention because they can accurately describe different types reflection spectra. However, inaccurate multispectral feature description leads to edge semantic ambiguity misclassification small objects. In this article, we propose novel network named edge-aware spectral–spatial information aggregation net (ESSANet) capture both high-level features low-level details for images. Specifically, on one hand, order improve representation ability discriminant features, design two-stream extraction via 3D hybrid convolution multi-level network. On other eliminate effect ambiguity, develop siamese structure multi-stage loss function. Experimental results show that our method achieved 3.5% 4.09% mean intersection over union (mIoU) score improvements 2.59% 3.32% Kappa compared with competitive baseline algorithm SEN12MS US3D datasets, respectively. addition, proposed paper also achieves better trade-off between speed accuracy.