Fine Extraction of Cultivated Land Parcels in Chengdu Plain Area Based on CDUSU-Net Network Model of Gaofen-2 Imagery

Accurate extraction of cropland information is an important prerequisite for agricultural remote sensing applications. This paper addresses the challenge of low accuracy in extracting cultivated land information in plain areas from high-resolution remote sensing images, with a particular focus on the inadequate extraction of fragmented cultivated land. The primary contribution of this research is the development of a novel model named Convolutional Dual up Sample U-Net (CDUSU-Net),which is based on U-Net (an end-to-end segmentation framework). This model incorporates a Atrous Spatial Pyramid Pooling module that integrates attention mechanisms and multi-scale features, significantly enhancing the accuracy of land information extraction. Additionally, the traditional upsampling method in the decoder is innovatively replaced with a Dual up Sample module, further refining the extraction process. By leveraging Gaofen-2 satellite imagery, this study offers a substantial improvement in the precision of cultivated land extraction, particularly for fragmented landscapes. In order to test the effectiveness of the method, the model is applied to the extraction of cultivated land in a region of Chengdu Plain, and six classical semantic segmentation network models are selected for comparison. The results show that the intersection ratio, recall, precision, and F1 Score of the CDUSU-Net network model for cultivated land information extraction are 92.85%, 97.96%, 94.57%, and 95.75%, respectively, which are better than the classical full convolutional neural network method. The research results will provide important technical and methodological support for the fine recognition of cultivated land information from high-resolution remote sensing images.