Fuzzy-guided hyperparameter optimization for efficient CNN-based plant disease classification

Abstract Deep Convolutional Neural Networks (DCNNs) have exhibited outstanding performance in plant disease classification. Nevertheless, their usefulness is very susceptible to hyperparameter selection. Efficient Hyper-Parameter Optimization (HPO) remains a fundamental challenge in deploying deep learning models for plant disease classification, as implementing extensive search techniques such as grid search, Bayesian Optimization (BO), and evolutionary algorithms requires considerable computational cost. Such conventional HPO approaches frequently rely on single-objective fitness criteria and demand substantial computational resources. Moreover, these approaches often lack interpretability in decision making. In this study, we propose an HPO based on a Fuzzy Inference System (FIS) approach that integrates a Mamdani FIS into the hyperparameter selection process of a DCNN. It maintains classification accuracy, while speeding up the hyperparameter selection process. The candidate configurations are assessed based on several factors, including validation accuracy, validation loss, and training time, rather than being assessed based on just one factor. These factors are combined into a single, crisp fitness score using fuzzy linguistic rules, providing transparent, interpretable, and multi-objective decision making. In order to minimize computational cost, candidate hyperparameters are first assessed using a quick partial training method that makes use of frozen convolutional layers and warm-started weights. A Mamdani FIS is then used to aggregate performance metrics including accuracy, loss, and training time, producing a crisp score for candidate ranking. After that, training of the chosen hyperparameters is exposed to fully end-to-end training, resulting in superior classification performance with respect to the state-of-the-art exhaustive search techniques. Experiments have been executed on four public plant disease datasets, using an AlexNet-based architecture, where the first experiments have been performed on a baseline model with a randomly-chosen hyperparameter configuration before the fuzzy-guided optimization process is initiated. Then, the model is performed with the best optimized hyperparameter configuration selected by the fuzzy-guided optimization process, demonstrating up to 29% improvement in accuracy, precision, recall and F1-score, along with reliable and comprehensible hyperparameter ranking and optimization. The achieved results show that while keeping an acceptable computing cost, the introduced fuzzy-guided technique consistently outperforms baseline model configurations with respect to all evaluation metrics, yielding greater classification accuracy and enhanced generalization. Hence, our suggested approach provides a lightweight, comprehensible, and efficient solution for hyperparameter tuning in the plant disease classification domain when compared to conventional HPO techniques, offering a practical direction for resource-efficient model optimization in precision agriculture that enables quick deployment of deep learning systems.