Convolutional Neural Network and Transfer Learning to Automate Plant Pathology

Abstract

Plant diseases have still been a significant challenge to food security worldwide, and manual diagnosis by trained plant pathologists is time consuming, subjective and challenging to scale across both smallholder and commercial farms. This paper brings together the recent advances in the application of convolutional neural networks (CNNs) and transfer learning for automated plant disease recognition, across the plant disease benchmark, field-captured images, and crop-specific pipelines for apple, tomato, maize, rice, cassava and wheat. Existing deep architectures pre-trained on the ImageNet corpus such as AlexNet, GoogLeNet, VGGNet, ResNet and DenseNet are considered as feature extractors and then fine-tuned on comparatively small, domain-specific leaf-image collections, for which the classification accuracy is shown to increase from below ninety-two percent to above ninety-nine percent for equal sizes of samples when transfer learning is applied. The synthesis also examines the scale and variety of the datasets, real-time detection systems, attention-augmented residual networks, interpretability through saliency, and case studies of mobile deployments for cassava and wheat, which illustrate a frequently cited disparity between laboratory-trained accuracy (100%+), and field-trained accuracy (78%–91%). Across many quantitative measures, it is clear that deeper networks do not perform as well as residual and densely connected ones, that data augmentation and background variability have a significant impact on generalization, and that lightweight networks can achieve reasonable accuracy while providing on-device inference. Ethical, regulatory and adoption issues for agricultural extension services in resource-constrained environments are discussed. The results suggest that transfer learning is the most data and compute efficient route to becoming deployable plant pathology diagnostic tools and that further work needs to reduce the gap observed between lab and field performance.