Image classification of sugarcane aphid densities using machine learning

Since 2013, the sugarcane aphid Melanaphis sacchari (Zehntner) has caused significant yield loss across the sorghum growing region in the U.S. Adequate management of sugarcane aphid depends on monitoring the pest and spraying insecticides once the infestation reaches an economic threshold. However, scouting this pest is time-consuming and inefficient in sorghum. We propose using computer vision models to assist pest monitoring to automatically classify sugarcane aphid infestation on leaves according to images of different density levels. We used 5,048 images collected during field scouting, and we evaluated the performance of four models: Inception v3, DenseNet 121, Resnet 50, and Xception. We trained the models to classify aphid densities into 6 classes: no aphids (0 aphids/leaf), no threat (1-10, 11-39 aphids/leaf), and consider using insecticide (40-125, 126-500, and >501 aphids/leaf) to manage sugarcane aphid in field conditions. Among these models, Inception v3 and Xception showed an overall accuracy score of 86%. The models correctly classified aphids as above or below threshold density over 97% of the time. In the future, the methodology developed and the models tested in this study can be used in sampling protocols or mobile applications to assist sorghum growers and researchers to automate scouting sugarcane aphids in sorghum fields.

Keywords: sorghum, integrated pest management, sugarcane aphid, automation, and classification.