|
Abstract:
|
Preseason crop-type forecasting has emerged as a novel application in machine learning and agriculture. A reliable algorithm for early crop-type prediction has many uses, including crop mapping, planted acreage forecasting, and area survey imputation. The primary method of preseason crop forecasting in the United States uses the NASS Cropland Data Layer (CDL), which is an annual crop specific land cover data set produced using satellite imagery and administrative data. Historical crop rotations derived from the CDL can be used to predict the future crop type in any given land area. The dataset obtained from the CDL is large, containing hundreds of millions of rows per state. Current approaches rely on sampling to make their machine learning algorithms feasible. In this work, the authors propose an alternative method that uses all the data in a fast and memory-efficient manner. The proposed method leverages the fact groups of crop pixels tend to exist in homogeneous fields. By summarizing pixels to the field level we can significantly reduce model size while still using all the data (no sampling). This new approach is more scalable and accurate than existing methods.
|
