We study how to incorporate the Normalized Difference Vegetation Index (NDVI) derived from remote sensing satellites to improve soybean yield predictions in ten major producing states in the United States. Unlike traditional methods which assume that a global OLS model applies to all observations, we account for geographical heterogeneity by using the Flexible Fourier Transform (FFT) model. Results show that there is considerable heterogeneity in how responsive soybean yield is to NDVI over the growing season. Out-of-sample cross-validation indicates that accounting for geographical heterogeneity improves the forecasts in terms of smaller prediction error compared to models assuming away geographical heterogeneity.