The diseases affecting tomato leaves are a significant problem to agricultural production and, thus, early and precise diagnosis is inevitable. Convolutional Neural Networks (CNNs) are popular in detecting plant diseases automatically, but their performance is very sensitive to the selection of all hyperparameters. In this paper, a comparative analysis of three bio-inspired metaheuristic algorithms, which include the Whale Optimization Algorithm (WOA), Ant Lion Optimization (ALO) and Dragonfly Algorithm (DA) is carried out to tune the parameters of CNNs to use in the classification of tomato leaf diseases. A baseline CNN was able to achieve test accuracy of 95.28, whereas metaheuristic optimization was able to boost the performance considerably with adaptive tuning of the learning rate, dense layer units, and dropout rate. The experimental findings regarding the V2 tomato leaf disease data set (10 classes) prove that DA obtained the highest accuracy with 98.04% and closely WOA with 97.94% and ALO with 96.61%. The analysis of ablation showed that dropout rate is a factor that had the strongest impact on decreasing overfitting, and the joint optimization of all hyperparameters gave the highest generalization. This work indicates the possibility of metaheuristic-based optimization to improve the CNN efficiency of precision agriculture and describes future research in the hybrid optimization and real-time application.