Effect of image resolution on automatic detection of whitefly (Hemiptera: Aleyrodidae) species on tomato leaflets using deep learning

Abstract

Deep learning approaches facilitate the rapid and accurate identification of insect pest species in agriculture. However, challenges arise when detecting tiny pests, such as whiteflies (Hemiptera: Aleyrodidae). Although high-resolution images are preferred for accurate detection, they reduce the image acquisition throughput. Hence, this study investigates the impact of image resolution on the accuracy of detecting the adults of two whitefly species, Bemisia tabaci Gennadius and Trialeurodes vaporariorum Westwood, on tomato leaflets. The following state-of-the-art object detectors were used: Faster R-CNN, SSD, YOLOv3, detectoRS, YOLOX, and deformable DETR. The results showed that the models with high-resolution datasets for training and testing performed well in insect detection, although performance declined with decreased image resolution. While, for many models, most errors occurred in species classification, those trained on low-resolution images had a higher proportion of localization errors. Poor performance was also observed when image resolution differed between training and inference. Moreover, the models in which the training dataset was assembled from high- and low-resolution images exhibited similar performance to those trained over each image resolution. Ultimately, detectoRS demonstrated the best performance for lower-resolution images or a mixture of multiple-resolution images, while YOLOX exhibited the highest accuracy for high-resolution images. This study highlights the importance of image resolution for optimal performance in insect pest-detection tasks.