Herbicide Drift Injury Mapping and Yield Loss Prediction in Cotton Using Remote Sensing Tools

Off-target movement of herbicides onto sensitive crop cultivars is a serious concern in agricultural landscapes. An ability to map herbicide injury and predict their impact on crop yield using drone images can allow for rapid and informed management decision making by growers. In this study, both multispectral and thermal imagery-based approaches were implemented for evaluating injury caused by tembotrione (Laudis^®) and dicamba (Xtendimax^®) in cotton. The spectral responses of cotton plants to these herbicides were recorded with DJI Matrice 600 drones mounted with a multispectral camera (Micasense RedEdge) and a thermal camera (Infrared Inc.) at 7, 14, and 21 days after herbicide application (DAA) at two different growth stages (match-head square and early-bloom). Deep neural networks were used for detection of herbicide injury and both univariate and multivariate regression techniques were used for predicting cotton lint yield reduction. Several sets of remote sensing variables, including vegetation indices, thermal values, and canopy height model were used in the mapping and the prediction process. The analysis of imagery at 14 DAA for match-head square stage showed that herbicide injury can be mapped with fair accuracy (73% and 68% respectively for tembotrione and dicamba) using aerial imagery. Results also showed that the multivariate regression technique is more effective in cotton yield loss prediction (R² values of 0.79 and 0.88 for tembotrione and dicamba, respectively) compared to the univariate regression technique (R² values of 0.73 and 0.84). The analysis for other DAA imagery is ongoing. Overall, this study demonstrates that aerial imagery can be used reliably to map herbicide injury and predict yield loss in cotton, though additional research is required to improve prediction accuracies.