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

Cropland is the strategic support for maintaining national food security, and accurately obtaining farm information is crucial for food production and security. Due to the complexity of agricultural environments in very high-resolution (VHR) remote sensing images, cropland is often interfered with by the type of cropland, the category of crops, and surrounding vegetation, leading to difficulties in extracting cropland and low extraction accuracy. To solve this problem, this article proposed a cropland extraction network called GF-UNet. First, based on the Attention U-Net network, GF-UNet used attention gates (AGs) to enhance the discriminative ability between partially similar features of cropland and non-cropland in complex situation. Next, the adaptive feature fusion module (AFFM) was introduced to integrate multi-scale cropland features, which could enhance the ability of GF-UNet to identify cropland. Finally, the Spatial Feature Extraction Module (SFEM) was introduced into the skip connection to improve the extraction of detailed features in the results. In this paper, GF-2 satellite images of Xuan 'en County, Hubei Province from June to September 2019 were used as research data, and U-Net, PSPNet, DeepLabv3+ and Attention U-Net networks were selected to conduct comparative experiments with GF-UNet. The results shown that the accuracy, F1-score and IoU of GF-UNet were better than other models, which were 91.25%, 92.41% and 84.56%, respectively. At the same time, the influence of SFEM and AFFM on the experimental results were also explored. Compared with the existing methods, GF-UNet was more suitable for cropland extraction in complex scenes and provided a practical method for cropland extraction in complex scenes.