Assessing normalized difference vegetation index (NDVI) data to estimate winter wheat yields and analyze winter wheat by homogeneous subregions at field scale in Kansas

Abstract

Wheat (Triticum aestivum L.) is the 4th largest staple crop produced worldwide. While global demand has increased over the last 15 years, the rate of increase of global cereal production has slowed or stagnated. Accurate information about crop production is key for local-scale research, farmers, and decision-making evaluation due to the typically high spatial variability in agricultural production, especially in environmentally heterogeneous high-producing regions. The main goal of this dissertation was to investigate the potential of satellite imagery in predicting winter wheat yields and analyze winter wheat yields by homogeneous subregions at field scale in Kansas, the largest producer of winter wheat in the U.S. The first chapter examined the performance of different satellite sensors (from coarse to moderate resolution - MODIS, Landsat, and Sentinel) in predicting winter wheat yields. The following chapters analyze the winter wheat yield prediction using environmentally distinct subregions regarding weather and management practices and multisource data (NDVI, weather, and climate). Linear Regression and a robust machine learning model, (i.e., Random Forest) were applied to predict winter wheat yields. The results, using NDVI predictor variables, were not enough to explain field-scale winter wheat yield variability across much of Kansas, where Landsat USGS achieved the lowest prediction error among all sensors (RMSE = 0.95 Mg ha⁻¹). The results proved to be more accurate when using Landsat NDVI variables to predict winter wheat yields in more homogeneous subregions (NC, SC, and West), with the best prediction in NC (RMSE = 0.76 Mg ha⁻¹). NC, SC, and West Kansas achieved the best results when including weather and management variables along with NDVI (RMSE of 0.59 Mg ha⁻¹ , 0.66 Mg ha⁻¹, and 0.69 Mg ha⁻¹ in NC, SC, and West), and outperformed the prediction when using all fields-yields across Kansas ( RMSE=0.78 Mg ha⁻¹). The prediction model showed that it is possible to predict yield in early crop developmental stages; however, after adding weather and management variables, NDVI predictor variables in the late stages of the growing season were the most important for winter wheat yield prediction. NDVI was more significant in predicting winter wheat yields in NC and West than in SC Kansas. NC showed management of fertilizers ( N, P, Cl) as good yield predictors and could be used along with NDVI to estimate yields. SC and West predictor variables relied more on variables related to environmental conditions or management practices related to environmental conditions, such as fungicide application, soil water storage, and sowing date. Overall, this research demonstrates that the applicability of empirical winter wheat yield modeling using NDVI predictor variables in Kansas is environmentally dependent. Lastly, winter wheat yield prediction using satellite imagery at the field scale could be benefited using this subregional scheme in Kansas.