Corn Grain Yield Prediction Using UAV-based High Spatiotemporal Resolution Multispectral Imagery

Abstract

Food demand is expected to rise significantly by 2050 due to the increase in population; additionally, receding water levels, climate change, and a decrease in the amount of available arable land will threaten food production. Yield precision maps, an application of precision agriculture, can be used by farmers to address these challenges by reducing input costs and optimizing yield. These maps can be created using machine learning models trained on field data (e.g., imagery data). Although performing satellite-based remote sensing to gather imagery has some advantages, using unmanned aerial vehicles (UAV)s is favorable over satellite-based approaches due to the higher spatial and temporal resolutions of the imagery. Vegetation indices (VI)s can be computed from the imagery and can represent the state or condition of vegetation. The present work performed yield prediction regression experiments that analyzed the effects of image spatial resolution (satellite vs. UAV) on prediction results, compared and ranked the prediction power of 33 VIs (and 5 raw-bands) over the growing season, and explored the optimal image acquisition date that produced the best prediction results. We gathered yield data and UAV-based multispectral imagery from a Canadian smart farm and trained random forest (RF) and linear regression (LR) machine learning models. High spatial resolution data generally led to better prediction results than lower spatial resolution data, especially for the RF model, where regardless of the VI choice or image acquisition date, good results were obtained. VIs that included the near-infrared and/or red-edge band generally performed better than the red-green-blue (RGB) VIs. The best performing VIs were: simple ratio index (near-infrared (NIR) & red-edge), normalized difference vegetation index with red-edge instead of red, normalized green index, green chlorophyll index, and simple ratio index (NIR & green). When higher spatial resolution imagery was available, optimized soil-adjusted vegetation index and renormalized difference vegetation index also performed well. We found that imagery from the middle of the growing season produced the best prediction results, even for some RGB VIs and raw-bands.