A Novel Residual-Concatenate Feature Fusion Strategy for Building Segmentation From Remote Sensing Images

This paper presents a novel encoder-decoder architecture based on U-Net integrated with a residual-concatenate multi-scale feature fusion (RCMFF) strategy to perform efficient building segmentation from remote sensing images. The proposed method considers an end-to-end deep learning framework based on UNet that consists of mainly three modules: i) an encoder with convolutional block attention module (CBAM) employed in their lateral connections; ii) the proposed residual-concatenate multi-scale feature fusion (RCMFF) module; and finally, iii) a decoder with refinement layers. The proposed RCMFF module integrates semantically relevant discriminative features inherent within the last encoder layer with reduced number of parameters thereby simultaneously improving the segmentation accuracy as well as the computational speed. The effectiveness of the proposed module is assessed using two benchmark archives: Massachusetts (MAS) and WHU building datasets. The proposed model demonstrates an overall improvement of 2.30% and 7.01% compared to the baseline models DeepLabV3+ and UNet++, respectively.