Improvement Of In-season Crop Mapping For Illinois Cropland Using Multiple Machine Learning Classifiers
Hui Li, Liping Di, Chen Zhang, Li Lin, Liying Guo
Abstract
Large-area crop type identification and mapping for cropland are intensively crucial for agriculture research, yield forecast, and disaster management. The United States Department of Agriculture (USDA) produces the Cropland Data Layer (CDL) for Contiguous United States cropland that involves crop type spatial distribution with 30m resolution. However, CDL is always published in early next year, which cannot meet the needs for in-season agricultural applications. In this paper, we embark on solving the questions above and introduce a large-area efficient crop mapping approach. We utilized historical CDL data to extract crop trusted pixels in Illinois. The trusted pixels were as training data of the machine learning model for crop type classification. We combined random forest and minimum distance model to time-series classification. We used Google Earth Engine to produce four in-season crop maps of the Illinois cropland in May-August. The validation result shows the overall accuracy of Illinois in-season crop mapping 2021 up to 91% and the major crops classification accuracy is around 92%.