Optimized Deep Learning Framework for Potato Leaf Disease Classification Using EfficientNetB3

Abstract

This research proposes a new deep learning-based technique to detect diseases of potato leaves using an EfficientNetB3 CNN. Plant Village dataset is used in this research which contains 2152 images subdivided into three classes: Potato Early Blight, Potato Late Blight and Healthy Potatoes. The objective of the project is to design an image classification model with optimal performance for accurate detection of the mentioned conditions to improve crop management through early intervention. The workflow includes complete data preprocessing steps where Image Data Generators from TensorFlow library are used for sample augmentation to improve generalization of the model. The EfficientNetB3 model is modified by the removal of the top layers and addition of new dense layers with dropouts to mitigate overfitting and one SoftMax output layer for multi-class outputs. Training of the model is done with the best optimization strategies, for example: Adam optimizer, early stopping, and reducing the learning rate when the model plateaued, to achieve stable convergence. One of the most important tasks to be done in the pipeline is computing the class weights to compensate for the dataset lacking healthy potato images with class imbalance.