Automated in-season crop-type data layer mapping without ground truth for the contiguous US

Abstract

The use of modern technology, such as machine learning, has ushered in a modern agricultural revolution, significantly enhancing crop yields, efficiency, and nutritional quality. To fully capitalize on these advancements, real-time crop mapping becomes crucial, enabling data-driven decisions on a large scale for sustainable agriculture. Although the US government's agro-geoinformatics data, they are only available after the growing season. Therefore, this study aimed to develop an in-season Cropland Data Layer (ICDL) for the contiguous United States, making use of trusted pixels. The process involved combining Landsat 8-9 and Sentinel 2 data to form a mosaic, followed by training a supervised classification model using random forest. This approach was applied to each state, and the resulting in-season maps were merged to create the final map. The accuracy of the ICDL was validated against CDL and ground-truth data, as well as NASS statistics report data. These products were also made freely accessible to the public through iCrop. In assessing corn and soybean fields in Nebraska and Iowa, F1-scores of (0.911, 0.845) and (0.959, 0.969) were achieved by the end of July 2022, respectively. The ICDL facilitates simple estimation of crop acreage without extensive surveying and can also provide real-time information to farmers so that they can make informed decisions on logistics, harvest, irrigation, and fertilizer use.