Pre- and within-season crop type classification trained with archival land cover information

REMOTE SENSING OF ENVIRONMENT

卷:264

Article

Johnson, David M.¹  Mueller, Richard¹ 

[1] Natl Agr Stat Serv, USDA, 1400 Independence Ave SW, Washington, DC 20250 USA

关键词: Crop;    Land cover;    Classification;    Predictive;    Real-time;    Landsat;    Sentinel-2;    Random forest;    Without training data;    Cloud-based;   

DOI  :  10.1016/j.rse.2021.112576

来源: Elsevier

【 摘 要 】

Crop type maps were created without the traditional need for in-season training data across the Corn Belt and Great Plains regions of the United States. This was accomplished through machine learning of historical land cover information, paired with a time-series of multi-spectral satellite imagery composites spanning the growing season, to develop rulesets, which were used for real-time prediction in the current year. Specifically, a decade's worth of annual 30 m resolution crop specific maps, known as the Cropland Data Layer (CDL), provided the foundation, and prior and current year's satellite imagery from Landsat 7 and 8 and Sentinel-2a and-2b built upon it. Four modeling scenarios, all using random forests, were performed to understand the crop mapping abilities of the datasets independently and combined. They were 1) use of CDLs only (i.e. prediction based solely on crop rotation history) 2) use of Landsat 7 and 8 bottom-of-atmosphere surface reflectance imagery only, 3) use of Sentinel-2a and-2b top-of-atmosphere imagery only, and 4) integration of the CDL, Landsat, and Sentinel-2 information together in a unified effort. Furthermore, the model runs were generated monthly, beginning in April, through the growing season to provide understanding of classification performance as a function of time. The 2020 crop year, relatively normal in terms of planting and weather, was used for the test. Accuracy statistics were generated by randomly sampling 50 counties and comparing those classification outputs to the actual 2020 CDL. Pixel-level results showed that prediction by midsummer using only the CDL information provided a crop type map with corn and soybean consumer and producer agreement above 70% and winter wheat just below 50%. The early season imagery-based classifications were markedly worse. However, as Landsat or Sentinel-2 imagery accumulated through July, those classifications became significantly better than those reliant on the use of the CDL information only. Ultimately, the very best crop maps resulted from integrating the CDLs with a full season's worth of Landsat and Sentinel-2 imagery. At that point in late September, the corn and soybean agreements were around 85% and winter wheat near 70%. All analysis was performed within Google Earth Engine cloud-based public imagery repository and high-performance computing system. The classification outputs provide practitioners with US crop type maps in near real-time.

【 授权许可】

Free