Cotton Leaf Disease Identification Using Transfer Learning
Asaduzzaman Herok, Sabbir Ahmed
Abstract
Accurate identification and classification of plant diseases play a vital role in safeguarding global food production and ensuring the economic well-being of the stakeholders. Various solutions have been developed using Deep Learning (DL) based systems for different essential crops. However, the research landscape regarding intelligent solutions for automatically classifying cotton leaf diseases remains largely unexplored despite cotton being a significant commercial crop in many regions of the world. This article describes a method for detecting diseases in cotton leaves based on transfer learning. Unlike the previous works focusing on a limited number of classes and/or samples, we combined samples from different resources and curated a dataset of seven cotton leaf diseases and one healthy class. The dataset reflects the challenges of heterogeneous sources, varying backgrounds, illumination, inter-class similarity, etc., making it suitable for estimating performance in real-life scenarios. After that, we assessed the effectiveness of various pretrained Deep Convolution Neural Networks, including InceptionV3, ResNet152V2, VGG16, InceptionResNetV2, Xception, MobileNetV2, and DenseNet121, where the VGG16 model achieved the highest accuracy of 95.02%. Finally, we provided a thorough class-wise performance and error analysis to show the capability of the model under different scenarios.