DEANet: Dual Encoder with Attention Network for Semantic Segmentation of Remote Sensing Imagery

Remote sensing has now been widely used in various fields, and the research on automatic land-cover segmentation methods of remote imagery is significant to development technology. Deep learning methods, which are developing rapidly field semantic segmentation, have applied segmentation. In this work, a novel deep network—Dual Encoder with Attention Network (DEANet) proposed. network, dual-branch encoder structure, whose first branch generate rough guidance feature map as area attention help re-encode maps next branch, proposed improve encoding ability an improved pyramid partial decoder (PPD) based parallel put forward make fuller use features form along receptive filed block (RFB). addition, edge module using transfer method introduced explicitly advance performance areas. Except for loss function composed weighted Cross Entropy (CE) Union subtract Intersection (UsI) designed training, where UsI represents new region-based aware replaces IoU adapt multi-classification tasks. Furthermore, detailed training strategy network well. Extensive experiments three public datasets verify effectiveness each our framework demonstrate that achieves more excellent over some state-of-the-art methods.