Classification of soybean response to off-target dicamba exposure using UAV-based imagery and machine learning methods

Though dicamba tolerant soybean genotypes are widely adopted by the world, but non-resistant soybean genotypes are still in fields and are subjected to exposure to dicamba as off-target and are damage reported by farmers from the past couple of years. Non-resistant soybean genotypes are vulnerable to dicamba, and severity of damage depends on growth stage, dosage, frequency, duration of exposure, and potential genetic background. From the past few years, development of crop high-throughput phenotyping has shown potential in assessing crop traits and response in different stages of life cycle. The goal of this study was to develop a UAV imagery-based classification system to help breeders to precisely and quickly select resistive genotypes that are exposed to off-target dicamba damage. An RGB (red-green-blue) camera was used to collect imagery data of 240 soybean breeding lines at the reproductive stages from five different fields. Dicamba damage scores were visually assessed by the breeders and are obtained on the same day or next day of image acquisition. Dicamba scores were grouped into three classes namely class 1, class 2, and class 3. Seven image features such as canopy coverage, color hue, color saturation, triangular greenness index (TGI), green leaf index (GLI), entropy, and contrast were extracted to classify the dicamba scores that are associated with the image features. ANOVA result of each image feature shows that damage of dicamba varies significantly among fields than the variation among three classes of dicamba score. Extreme gradient boosting (XGBoost) model were developed to classify three classes of dicamba scores using image features. Classification model were developed using standardized image feature data and an overall classification accuracy of 0.75 were obtained from the model. Whereas, class 1 showed higher accuracy of 0.91, and 0.83 were obtained for both class 2 and 3. Furthermore, class 2 resulted higher kappa, precision, recall, sensitivity, specifity, and F1 score than other two classes. Along with these, image features entropy, hue, and GLI indicate that these image features have higher effect in classifying dicamba scores and are obtained from XGBoost feature importance. This study reflected and demonstrated that the use of UAV-RGB imagery techniques has the potential to select dicamba resistive soybean genotypes.