A comparison of machine learning algorithms for regional wheat yield prediction using NDVI time series of SPOT-VGT

Abstract

This study compares two machine learning algorithms to predict regional winter wheat yields. The models, based on Boosted Regression Trees (BRT) and Support Vector Machines (SVM), are constructed of Normalized Difference Vegetation Indices (NDVI) derived from low resolution SPOT VEGETATION imagery. Three types of NDVI-related predictors were used: Single NDVI, Incremental NDVI and Targeted NDVI. BRT and SVM were first used to select features with high relevance for predicting the yield. Periods of high influence spanning from March to June were detected by both machine learning methods. After feature selection, BRT and SVM models were applied to the subset of selected features for yield forecasting. BRT seems to consistently outperform SVM.