Real-Time Semantic Segmentation of Remote Sensing Images for Land Management

Abstract

Remote sensing image segmentation is a crucial technique in the field of land management. However, existing semantic segmentation networks require a large number of floating-point operations (FLOPs) and have long run times. In this paper, we propose a dual-path feature aggregation network (DPFANet) specifically designed for the low-latency operations required in land management applications. Firstly, we use four sets of spatially separable convolutions with varying dilation rates to extract spatial features. Additionally, we use an improved version of MobileNetV2 to extract semantic features. Furthermore, we use an asymmetric multi-scale fusion module and dual-path feature aggregation module to enhance feature extraction and fusion. Finally, a decoder is constructed to enable progressive up-sampling. Experimental results on the Potsdam data set and the Gaofen image data set (GID) demonstrate that DPFANet achieves overall accuracy of 92.2% and 89.3%, respectively. The FLOPs are 6.72 giga and the number of parameters is 2.067 million.