MF-Dfnet: a deep learning method for pixel-wise classification of very high-resolution remote sensing images

Semantic segmentation of high-resolution remote sensing images is very important. However, the targets in the high-resolution optical satellite images are always various in size, which lead to multiscale problems resulting in difficulty of locating and identifying the target. High-resolution remote sensing is more complex than natural phenomena; this leads to false alarms due to a greater intraclass inconsistency. Thus, the pixel-wise classification of high-resolution remote sensing images becomes challenging. Aiming at the above problems, we propose a multiscale feature and discriminative feature network (MF-DFNet). We introduce the hierarchical-split block (HSB) and the residual receptive field block module (RRFBM) to extract multiscale information to address multiscale problems. We also introduce a foreground-scene relation module to enhance the discrimination of features and deal with the false alarm phenomenon. In addition, the channel attention block (CAB) is introduced to select more discriminative features. We use two publicly available remote sensing image datasets (Vaihingen and Massachusetts building) for the experiments in this paper. Compared to current advanced models, our results show that MF-DFNet achieves state-of-the-art performance and can effectively improve the integrity and correctness of semantic segmentation in high-resolution remote sensing images.