Objective To improve the efficiency and accuracy of cultivated land reserve resource information extraction, meet the demands of modern agricultural development for land resource dynamic monitoring.

Method This paper took Yuncheng District in Yunfu City of Guangdong Province as the study area, and proposed a method for extracting cultivated land reserve resources by integrating object-oriented rule construction and deep learning. Using GF-6 high-resolution satellite imagery, multi-scale image segmentation was performed, and a stepwise elimination method was applied to construct land classification identification rules, extracting samples of typical land types. Subsequently, based on the rule-based samples, a training label dataset for the U-Net deep learning model was constructed to accomplish the extraction and classification of cultivated land reserve resources.

Result For Yuncheng District, the optimal segmentation scale was determined to be 300. At this scale, features of the same category were effectively segmented, with clear boundaries between grassland and bare land. The overall precision of the proposed method in the study area reached 87.3%, while the mean intersection over union and F1-score achieved 75.4% and 86.7%, respectively, enabling precise extraction of complex feature boundaries. The deep learning approach based on the improved U-Net effectively reduced misclassification, particularly in areas with blurred boundaries and mixed pixels, and improved precision by approximately 5 percentage points compared to traditional object-oriented method.

Conclusion The remote sensing intelligent extraction method developed in this study demonstrates both high precision and time efficiency. It can provide robust support for local land use planning, cultivated land resource management, and ecological conservation, showing promising potential for broader application.