Field Patch Extraction Based on High-Resolution Imaging and U2-Net++ Convolutional Neural Networks

• Published in: Remote sensing (Basel, Switzerland) Vol. 15; no. 20; p. 4900
• Main Authors: Long, Chen, Wenlong, Song, Tao, Sun, Yizhu, Lu, Wei, Jiang, Jun, Liu, Hongjie, Liu, Tianshi, Feng, Rongjie, Gui, Abbas, Haider, Lingwei, Meng, Shengjie, Lin, Qian, He
• Format: Journal Article
• Language: English
• Published: Basel MDPI AG 01.10.2023
• Subjects: Accuracy; Agricultural land; Agricultural management; Agriculture; Algorithms; Artificial neural networks; Boundary maps; channel-spatial attention mechanism; Deep learning; depth-wise separable convolution; Economic development; Efficiency; farmland parcel extraction; High resolution; high-resolution remote sensing imagery; Image resolution; Land surveys; Modules; Neural networks; Precipitation; Remote sensing; Satellite imagery; Satellites; Water conservation; China; Yunnan China
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs15204900

Accurate extraction of farmland boundaries is crucial for improving the efficiency of farmland surveys, achieving precise agricultural management, enhancing farmers’ production conditions, protecting the ecological environment, and promoting local economic development. Remote sensing and deep learning are feasible methods for creating large-scale farmland boundary maps. However, existing neural network models have limitations that restrict the accuracy and reliability of agricultural parcel extraction using remote sensing technology. In this study, we used high-resolution satellite images (2 m, 1 m, and 0.8 m) and the U2-Net++ model based on the RSU module, deep separable convolution, and the channel-spatial attention mechanism module to extract different types of fields. Our model exhibited significant improvements in farmland parcel extraction compared with the other models. It achieved an F1-score of 97.13%, which is a 7.36% to 17.63% improvement over older models such as U-Net and FCN and a more than 2% improvement over advanced models such as DeepLabv3+ and U2-Net. These results indicate that U2-Net++ holds the potential for widespread application in the production of large-scale farmland boundary maps.