Data driven modeling of corn yield: a machine learning approach

With the increase in global population and growing demand for food, there has been considerable research in leveraging data in the agricultural domain to improve yields. Tremendous amounts of data are generated on farms, ranging from amount of water used for irrigation to the quantities of fertilizers applied. To our knowledge, this study is the first that uses high-resolution crop data (280,000 points at 10-meter scale) to improve understanding and prediction of the impact of hydrology-related variables, namely topography, soil, and weather, on yield. Supervised machine learning techniques, namely decision trees and random forests, are used to develop data-driven predictive models. A case study of corn fields in Iowa demonstrates how an ensemble technique like random forest can improve upon simpler models like decision trees. In addition, the random forest model is used to develop partial dependence plots of corn yield versus different feature variables (derived from topography, soil and weather). These plots help in understanding how yield varies with changes in different feature variables. For example, there is an optimum topographic range in which yield is high. Corn yield is higher in gentle depressions as compared to steep slopes and very deep depressions. Further, very high and very low precipitation during the emergence stage (VE) is most likely to lead to lower yield. The model described in this study can also be used to develop intra-field importance maps that delineate importance of a particular type of a variable (like precipitation) on corn yield at fine scales. These maps can be used to visually inspect the interaction of topography and precipitation and the resultant impact on corn yield, and could be used to support more fine-scale farm management strategies (e.g., irrigation only where needed)