Potential of ensemble tree methods for early-season prediction of winter wheat yield from short time series of remotely sensed normalized difference vegetation index and<i>in situ</i>meteorological data

Abstract

We aimed at analyzing the potential of two ensemble tree machine learning methods—boosted regression trees and random forests—for (early) prediction of winter wheat yield from short time series of remotely sensed vegetation indices at low spatial resolution and of in situ meteorological data in combination with annual fertilization levels. The study area was the Huaibei Plain in eastern China, and all models were calibrated and validated for five separate prefectures. To this end, a cross-validation process was developed that integrates model meta-parameterization and simple forward feature selection. We found that the resulting models deliver early estimates that are accurate enough to support decision making in the agricultural sector and to allow their operational use for yield forecasting. To attain maximum prediction accuracy, incorporating predictors from the end of the growing season is, however, recommended.