SE-CapsNet: Automated evaluation of plant disease severity based on feature extraction through Squeeze and Excitation (SE) networks and Capsule networks

Abstract

Diseases in plants harm the quantity of the overall food production as well as the quality of the yield. Early detection, diagnosis and treatment can greatly reduce losses, both economic and ecological. Intuitively, reduction in the use of agrochemicals due to timely detection of the disease, would greatly help in mitigating the environmental impact. In this paper, the authors have proposed an improved feature computation approach based on Squeeze and Excitation (SE) Networks, before processing by the original Capsule networks (CapsNet) for classification, for estimating the disease severity in plants. Two SE networks, one based on AlexNet and another on ResNet have been combined with Capsule networks. Leaf images for the devastating Late Blight disease occurring in the Tomato crop have been utilized from the PlantVillage dataset. The images, divided into four severity stages i.e. healthy, early, middle and end, are downscaled, enhanced and given as input to the SE networks. The feature maps generated from the two networks are separately given as input to the Capsule Network for classification and their performances are compared with the original CapsNet, on two image sizes 32X32 and 64X64. SE-Alex-CapsNet achieves the highest accuracy of 92.1% and SE-Res CapsNet achieves the highest accuracy of 93.75% with 64X64 image size, as compared to CapsNet that results in 85.53% accuracy. The classification accuracies of six state-of-the-art CNN models namely AlexNet, SqueezeNet, ResNet50, VGG16, VGG19 and Inception V3 are also presented for comparison purposes. Accuracy as well as precision, recall, F1-score, validation loss etc. measures have been recorded and compared. The findings have been validated by implementing the proposed approaches with another dataset, achieving similar resultant accuracy measures. The implementation was also accomplished with datasets after noise addition in six different variations, to verify the robustness of the proposed model. Based on the performances, the proposed techniques can be exploited for disease severity assessment in other crops as well and can be extended to other areas of applications such as plant species classification, weed identification etc. In addition to improved performance, with reduced image size, the proposed methodology can be utilized to create a mobile application requiring low processing capabilities, to be installed on reasonably priced smartphones for practical usage by farmers.