EAS-CNN: automatic design of convolutional neural network for remote sensing images semantic segmentation

Abstract

Accurate and effective semantic segmentation methods for remote sensing are important for applications such as precision agriculture, urban planning, and disaster monitoring. Convolutional neural networks (CNN) have achieved remarkable performance in the field of remote sensing semantic segmentation. However, the design of CNNs is both time-consuming and necessitates a substantial amount of domain expertise and experience. To address the aforementioned issues, we propose a neural architecture search method called EAS-CNN. The method constructs a search space based on a U-shaped structure and utilizes a fixed-length encoding solution based on gene expression suppression to preserve potential useful information during the evolution process. Furthermore, an improved genetic strategy is proposed to enhance search efficiency and save computational resources. In this paper, we evaluate the proposed EAS-CNN method against state-of-the-art semantic segmentation methods and verify its effectiveness. Experimental results show that EAS-CNN achieves high OA values of 91.2% and 91.6% on the Vaihingen and Postman datasets, respectively. Furthermore, we conduct a thorough analysis of the experimental results and summarize effective design patterns for model architecture to enhance remote sensing semantic segmentation tasks.