Improving crop yield estimation by applying higher resolution satellite NDVI imagery and high-resolution cropland masks

This research contributes to the literature by investigating if and by how much higher resolution satellite imagery improves crop yield estimation accuracy at the county level when paired with a high-resolution cropland mask. Satellite imagery is an interesting big data source that has potential applications in agriculture. When applying satellite imagery for crop yield estimation, practitioners choose which resolution (i.e., grid size) of images to use. Processing higher resolution images requires greater computing resources compared to lower resolution images. Practitioners may choose to use lower resolution images, but there may be a loss in crop yield model estimation accuracy. The cost of computation has decreased significantly with the advent of cloud computing and open access computing portals such as Google Earth Engine. These technologies have made satellite image processing more economical. The objective of this research is to quantify the crop yield estimation accuracy improvement that could be achieved by using higher resolution normalized difference vegetation index (NDVI) with a cropland mask. NDVI (a measure of crop greenness) data was collected for 48 U.S. states for four crops over 11 years. The crops investigated were corn, soybeans, spring wheat, and winter wheat. Each crop yield regression model estimation showed improved accuracy (R2) as the satellite NDVI resolution increased. Results suggest that using higher resolution satellite NDVI provides more accurate crop yield estimation compared to lower resolution satellite NDVI. This study is believed to be the most comprehensive study to date using NDVI to estimate crop yield, analyzing 48 states in the U.S. and four crops over 11 years using three resolution levels.