China-PMF-10: a 10-m national map of plastic-mulched farmlands in China of 2020 using deep semantic segmentation

Plastic mulching films are widely utilized in agriculture around the globe to increase soil temperature, retain soil moisture, and prevent pests. However, its large-scale and long-term application causes several environmental issues, including soil microplastic accumulation and “white pollution”. Therefore, obtaining reliable spatial distribution data on plastic-mulched farmlands (PMFs) is essential, particularly at a large scale. This study aimed to provide the first national-scale PMF map of China with a high resolution of 10 m. Specifically, this study proposed a hybrid framework for PMF mapping, which employed the VM-UNet semantic segmentation model for both updating weak labels and national-scale PMF mapping. First, a web crawler was used to gather news and academic literature related to China’s PMFs, extracting temporal and spatial distribution data on plastic mulch films. Afterwards, a national 1° grid system was established for data organization, where initial weak labels were generated via Google Earth Engine and Sentinel-2 imagery. These labels were subsequently refined through an active learning strategy within a transfer learning framework. Finally, VM-UNet was adopted for the national-scale prediction of China’s plastic-mulched farmland. Experimental results demonstrate that China-PMF-10 dataset achieved a user’s and producer’s accuracies of 86.30 ± 1.51% and 79.45 ± 1.90% for PMF, respectively. China’s total PMF area in 2020 was estimated at approximately 6.02 ± 0.58 Mha (95% CI). Furthermore, this study also found that the Hu Line emerged as a critical boundary, with more extensive PMFs and higher PMFs-to-cropland ratios to its west. This study provides a scientific basis for sustainable agricultural management and plastic pollution mitigation in China, and offers methodological insights for large-scale PMF monitoring in other regions.

塑料地膜（plastic mulching films）在全球农业领域应用广泛，可提升土壤温度、保持土壤墒情并防治虫害。然而，其大规模、长期的应用引发了诸多环境问题，包括土壤微塑料积累与“白色污染”。因此，获取可靠的地膜覆盖农田（plastic-mulched farmlands, PMFs）空间分布数据至关重要，在大尺度范围内尤为如此。本研究旨在绘制中国首幅10米高分辨率的全国尺度地膜覆盖农田分布图。具体而言，本研究提出了一种用于地膜覆盖农田制图的混合框架，该框架借助VM-UNet语义分割模型来更新弱标签并完成全国尺度的地膜覆盖农田制图。首先，本研究通过网络爬虫收集与中国地膜覆盖农田相关的新闻与学术文献，提取塑料地膜的时空分布数据。随后，构建了全国1°网格系统用于数据组织，通过谷歌地球引擎（Google Earth Engine）与哨兵二号（Sentinel-2）影像生成初始弱标签。随后，在迁移学习框架下，通过主动学习策略对这些标签进行精细化处理。最终，采用VM-UNet模型完成中国全国尺度的地膜覆盖农田预测制图。实验结果表明，China-PMF-10数据集在地膜覆盖农田类别上的用户精度与生产者精度分别为86.30±1.51%和79.45±1.90%。2020年中国地膜覆盖农田总面积估算为约6.02±0.58百万公顷（95%置信区间）。此外，本研究还发现胡焕庸线（Hu Line）是关键的分界线，该线以西的地膜覆盖农田分布更广，且地膜覆盖农田占耕地的比例更高。本研究为中国的可持续农业管理与塑料污染治理提供了科学依据，同时也为其他地区的大尺度地膜覆盖农田监测提供了方法学参考。