FMCNet: A Fuzzy Multiscale Convolution Network for Remote Sensing Image Segmentation

Abstract

Due to being affected by factors such as imaging distance, lighting, ground features, and environment, objects in the same class may have certain differences, and different classes of objects often produce similar visual features in remote sensing images. This phenomenon leads to an uncertainty problem in segmentation of remote sensing images, i.e., intra-class heterogeneity and inter-class blurring. To alleviate this problem, a fuzzy multiscale convolution neural network (FMCNet) is proposed in this paper. By extracting receptive fields of different scales, sizes and aspect ratios, the detailed information in remote sensing objects is fully represented. The relationship between their adjacent pixels is effectively expressed by fuzzy logic learning to alleviate the uncertain segmentation. The proposed method achieves overall accuracies of 85.33%, 86.31%, and 85.39% on the Vaihingen, Potsdam, and Gaofen Image datasets respectively. It demonstrates superior performance compared to existing popular methods.