Mapping 10-m Plastic Greenhouses of China Based on Multisource Remote Sensing Products

Plastic greenhouses (PGs), as an important modern agricultural infrastructure, can significantly enhance crop yield and quality through efficient and intensive production practices. Timely and accurate mapping of PGs is essential for food security and sustainable agricultural development. However, insufficient training samples and the constraints of existing mapping methods have confined most studies to small-scale regions, and there is a notable absence of high-resolution PGs datasets at the national level. Moreover, traditional PGs extraction relies heavily on labor-intensive and time-consuming manual sample collection, which poses great challenge to meet the demands of national-scale mapping. To address these limitations, we propose a large-scale PGs mapping framework based on automated training sample generation for mainland China. Initially, a high-quality pool of positive and negative samples was established by integrating existing multi-source land use/land cover (LULC) products and thematic PGs data through spatial overlay analysis and spectral filtering operations. Uniform and sufficient training samples were then selected to construct multi-source features based on Sentinel images and other auxiliary data, and finally, a tile-based local random forest approach was implemented to produce a 10m PGs dataset for China in 2020 (ChinaPGs10) on the Google Earth Engine (GEE) platform. Evaluation results show that ChinaPGs10 performs well on two nationwide test sets, with product accuracies of PGs reaching 84.64% and 86.15%. And cross-comparison with PGs-30 and PGs-3 reveals high spatial consistency of different products, which further proves the effectiveness and robustness of the proposed framework.