关键词: 冬闲田, 皖江平原, Sentinel-1/2, FR-Net, 深度学习, 闲田利用

Abstract:

The efficient utilization of cropland resources serves as the cornerstone for safeguarding national food security and promoting sustainable agricultural development. As a pivotal grain production base within the Yangtze River Economic Belt, the Wanjiang Plain is characterized by traditional double-cropping systems, specifically rice-wheat and rice-rapeseed rotations. However, influenced by factors such as rural labor migration, fluctuating agricultural profitability, and climate change, the phenomenon of winter fallow fields (WFF) has become increasingly prevalent in this region. Due to frequent cloud cover and rain during winter, landscape fragmentation, and complex planting structures in the Wanjiang Plain, traditional monitoring methods relying on single-source optical remote sensing or coarse-resolution imagery struggle to accurately identify fragmented fallow parcels. This has resulted in a scarcity of high-precision, long-time-series thematic datasets, thereby constraining the scientific assessment of regional cropland utilization efficiency. Leveraging the Google Earth Engine (GEE) cloud platform, this study constructed a multi-source remote sensing collaborative observation dataset spanning from 2019 to 2024. First, Sentinel-1 Synthetic Aperture Radar (SAR) and Sentinel-2 optical imagery corresponding to key winter phenological stages were selected. A multi-dimensional feature cube was constructed by extracting SAR backscatter coefficients (VV/VH), spectral bands, and the Normalized Difference Vegetation Index (NDVI) to effectively mitigate cloud interference and capture distinct phenological characteristics. Second, based on 352 winter fallow field samples and 325 non-winter fallow field samples, a cascaded mapping strategy integrating "Random Forest (RF) pre-classification + Fine Resolution Network (FR-Net)" was employed. The RF model was utilized to generate initial probability maps, followed by the application of the FR-Net deep learning model—incorporating residual structures—for semantic segmentation and edge refinement. This approach effectively resolved boundary ambiguity issues common in fragmented parcels. This dataset comprises annual raster data of the spatial distribution of winter fallow fields in the Wanjiang Plain from 2019 to 2024, with a spatial resolution of 10 m and a coordinate system of WGS 1984 UTM Zone 50N. Results indicate that the winter fallow phenomenon in the study area is both extensive and persistent. Validated against independent samples, the dataset achieves a six-year average F1-score of 87.21% and an Overall Accuracy (OA) of 85.64%, demonstrating high mapping accuracy and spatial consistency. This dataset can directly support agricultural departments in planning the development and utilization of winter fallow fields, estimating grain production potential, and researching the cropland ecosystem carbon cycle. It provides reliable data support for regional agricultural planting structure adjustment and policy formulation.

Data summary:

Key words: winter fallow fields, Wanjiang plain, sentinel-1/2, FR-Net, deep learning, utilization of fallow land