Edge Guided Context Aggregation Network for Semantic Segmentation of Remote Sensing Imagery

Semantic segmentation of remote sensing imagery (RSI) has obtained great success with the development deep convolutional neural networks (DCNNs). However, most existing algorithms focus on designing end-to-end DCNNs, but neglecting to consider difficulty in imbalance categories, especially for minority categories RSI, which limits performance RSI semantic segmentation. In this paper, a novel edge guided context aggregation network (EGCAN) is proposed RSI. The Unet employed as backbone. Meanwhile, an branch and extraction are designed comprehensive enhancement modeling. Specifically, promote entire comprehension further emphasize representation information, consists three modules: module (EEM), dual expectation maximization attention (DEMA), (EGM). EEM created primarily accurate tracking. According that, DEMA aggregates global contextual features different scales along spatial channel dimensions. Subsequently, EGM cascades aggregated into decoder process capture long-range dependencies error-prone pixels region acquire better labels. Besides this, exploited presented rich multi-scale information through elaborate hybrid pyramid pooling (HSPP) distinguish taking small percentage background. On Tianzhi Cup dataset, algorithm EGCAN achieved overall accuracy 84.1% average cross-merge ratio 68.1%, improvement 0.4% 1.3% respectively compared classical Deeplabv3+ model. Extensive experimental results dataset released ISPRS Vaihingen Potsdam benchmarks also demonstrate effectiveness over other state-of-the-art approaches.