Quarterly cropland mapping in sample shortage regions using Sentinel-2 images based on the knowledge-driven OCSVM-RF

• Knowledge was used to overcome the label shortage problem in practice. • Knowledge-driven OCSVM-RF was used to achieve quarterly cropland mapping. • Knowledge-driven OCSVM-RF performs well across different terrain regions. Croplands are a critical resource for human sustenance, and their quantity and quality have direct implications for national food security. The development of remote sensing technology has provided robust data support for timely regional cropland mapping. However, traditional cropland mapping methods depend on localized ground samples, limiting model transferability and product timeliness, particularly in regions with complex terrain. Consequently, this study developed a cropland mapping framework that enables quarterly mapping without the need for manual annotation, based on quarter-averaged composite Sentinel-2 images. This was achieved by integrating cropland prior knowledge with single-category (One-Class Support Vector Machine, OCSVM) and multi-category (Random Forest, RF) machine learning models (i.e., knowledge-driven OCSVM-RF). Experiments in regions with different terrains (hilly and plain) were performed, and the results showed that the knowledge-driven OCSVM-RF could effectively extract quarterly cropland distribution information for different terrain regions within a given year. Meanwhile, it significantly outperformed other automatic sample generation methods and achieved a performance comparable to that of the traditional supervised RF. In addition, we evaluated the feasibility of using knowledge-driven OCSVM-RF-generated labels to drive U-Net, a model commonly used in agricultural remote sensing communities. The results showed that the knowledge-driven OCSVM-RF-generated labels effectively facilitated U-Net model training, enabling the model to learn deep abstract features from quarter-averaged composite Sentinel-2 images for accurate cropland and non-cropland discrimination across various terrains and quarters within a given year. This further demonstrates the robustness and reliability of the proposed knowledge-driven OCSVM-RF and its generated dataset. This study could provide new options for high-frequency cropland mapping in sample shortage regions.