State-of-the-art computer vision techniques for automated sugarcane lodging classification

Sugarcane crop lodging is an agronomic condition that critically affects the cane yield and sugar quality. Lodging also impedes intercultural management and harvest operations. Currently, no approaches known exist to non-invasively assess sugarcane lodging. This study is therefore focused on expedited and autonomous sugarcane lodging detection using state-of-the-art computer vision techniques. A total 1600 digital red-green-blue (RGB) images of lodged and non-lodged sugarcane were acquired for two cultivars and two years (2020, and 2021) of growing seasons. These images were augmented to obtain a total of 6400 images. Based on tested proportions for minimum model overfittings, 80 % of the 6400 images were used for training, 10 % for validation, and remaining 10 % for testing of seven state-of-the-art deep learning (DL) models; ResNet50, GoogLeNet, DarkNet53, Inception V3, Xception, AlexNet, and MobileNetV2. When validated, amongst all, ResNet50 demonstrated the highest lodging prediction accuracy of 98.5 %, followed by GoogLeNet (98.0 %), DarkNet53 (97.6 %), InceptionV3 (97.6 %), Xception (97.1 %), AlexNet (96.5 %), and MobileNetV2 (93.6 %) and respective model precisions of 98.6 %, 98.6 %, 97.5 %, 97.2 %, 96.9 %, 96.1 %, and 93.1 %. Maximum accuracies and minimum model overfitting were observed for batch size of 16 and 30 epochs for all the models. The overall error rate for MobileNetV2, AlexNet, Xception, DarkNet53, InceptionV3, GoogLeNet, and ResNet50 models were 6.4 %, 3.5 %, 2.9 %, 2.4 %, 2.4 %, 2.0 % and 1.5 %, respectively. Best performing ResNet50 model was again tested on 50 independent images from real field conditions from both years and a net accuracy of 94 % was obtained. The residual blocks and skip connection features of ResNet50 help optimizing training parameters and therefore achieved better performance compared to other DL models. Autonomous lodging assessments with AI models could help guide supervised harvest operations without compromising the lodged crop, yield predictions, and site-specific management of other intercultural operations.