Optimized convolutional neural network using bacterial colony optimization for plant leaf disease detection

Abstract

The currently growing effects that plant diseases have on global agriculture require the creation of highly intelligent and accurate detection systems. Convolutional Neural Networks (CNNs), as deep learning, have proved effective in the detection of plant diseases based on images. The CNN performance, however, is very sensitive to hyperparameter tuning, which usually requires manual, sub-optimal tuning. The study suggests a new method to detect plant diseases with an optimized CNN architecture optimized by the Bacterial Colony Optimization (BCO). The BCO algorithm replicates the adaptive foraging behavior of bacterial colonies to automatically determine optimal CNN hyperparameters, such as the number of filters, kernel size, pooling methods, learning rate, and dropout probability of the CNN. The experiment conducted on the PlantVillage dataset demonstrated that the proposed BCO-CNN achieved an accuracy of 96.68% with a false alarm rate (FAR) of 4.05% outperforming other heuristic-fitted models such as particle swarm optimization (PSO) - CNN, CNN with support vector machine (SVM), and CNN, VGG16 in accuracy, precision, recall, and F1-score. The given work offers an automated, scalable approach to the accurate, early detection of the diseases of the plant, contributing to better yields of crops and sustainable agriculture.

Plant phenotyping relevance

植物葉画像から病害状態を推定するCNNのハイパーパラメータ最適化手法を開発・評価しており、植物フェノタイピング手法が研究の中心である。

Code and data availability

The paper's plant-phenotyping experiments (CNN/BCO leaf disease detection) were run on the public Kaggle Plant Disease (PlantVillage) dataset, explicitly cited with its public URL. No author code or models are reported as available.