10 m crop type mapping using Sentinel-2 reflectance and 30 m cropland data layer product

• DOI: 10.1016/j.jag.2022.102692
• Publication Date: 2022-01-29T01:01:30Z
• Authors: Khuong H. Tran; Hankui K. Zhang; John T. McMaine; Xiaoyang Zhang; Dong Luo
• Keywords: Physical geography; Time series; 04 agricultural and veterinary sciences; GB3-5030; Environmental sciences; Cropland data layer; Machine learning; 0401 agriculture, forestry, and fisheries; High resolution; GE1-350; Sentinel-2; Land cover mapping

The 30 m resolution U.S. Department of Agriculture (USDA) crop data layer (CDL) is a widely used type map for agricultural management and assessment, environmental impact food security. A finer can potentially reduce errors related to area estimation, field size characterization, precision agriculture activities that requires growth information at scales than field. This study develop method mapping using Sentinel-2 10 bands (i.e., red, green, blue, near-infrared) examine the benefit derived map. was conducted two areas with significantly different sizes types in South Dakota California, respectively. surface reflectance normalized difference vegetation index (NDVI) acquired 2019 growing season were generate monthly median composites as classification input. training evaluation samples from CDL by (i) finding good quality pixels (ii) identifying single representative pixel time series each pixel. random forest algorithm trained 80% evaluated 20% remaining samples, results showed high overall accuracies 94% 83% California areas, major crops both obtained user's producer's (>87%). There agreement between class proportions R2 ≥ 0.94 root mean square error (RMSE) ≤ 3%. More importantly, compared CDL, has much less salt-pepper boundary-aliasing effects defines better small features (e.g., fields, roads, rivers). potential large discussed.