GF-UNet:A Cropland Extraction Method Based on Attention Gate and Adaptive Feature Fusion

Abstract

Cropland plays a critical role in maintaining national food security, but its extraction is often hindered by factors such as type of cropland, crop category, and surrounding vegetation, resulting in low extraction accuracy. This paper proposes a cropland extraction network, called GF-UNet, to address the challenges of accurately extracting cropland from very high-resolution (VHR) remote sensing images. GF-UNet builds on the Attention U-Net network and introduces Attention Gates (AGs) to improve the ability to discriminate between similar features of cropland and non-cropland in complex situations. This helps to improve the accuracy of cropland extraction. In addition, an Adaptive Feature Fusion Module (AFFM) is incorporated to integrate multi-scale cropland features, further enhancing the network's ability to identify cropland. The Spatial Feature Extraction Module (SFEM) is also introduced into the skip connection to improve the extraction of detailed features in the results. The research data used in the study consists of GF-2 satellite images of Xuan 'en County, Hubei Province, from June to September 2019. Comparative experiments were conducted with SOTA models, the results demonstrate that GF-UNet outperforms the other models in terms of accuracy, F1-score, and IoU. The accuracy, F1-score, and IoU of GF-UNet were reported as 91.25%, 92.41%, and 84.56%, respectively. The study also explores the impact of SFEM and AFFM on the experimental results. Compared to existing SOTA methods, GF-UNet proves to be more suitable for cropland extraction in complex scenes, providing a practical approach to addressing the challenges of cropland extraction in such scenarios. Povzetek: Članek predstavlja omrežje GF-UNet za natančno izločanje kmetijskih površin iz slik visoke ločljivosti s pomočjo daljinskega zaznavanja. GF-UNet uporablja modul za adaptivno združevanje značilnosti in prostorski modul za izboljšanje delovanja.