Deep Learning (DL) methods have been found successful for plant disease classification on several datasets and environments. However, research on the development of DL algorithms requires extensive experimentation to propose accurate and feasible solutions. Moreover, complex DL models require a long training time to achieve training convergence for each epoch. There are methods available to reduce the training time by making the DL models lightweight. This was achieved by reducing the number of layers and the complexity of the model. This study analysed the capability of metaheuristic algorithms to select the most relevant subset of features from high-dimensional data while maintaining the high accuracy of the DL model. In this regard, the latest metaheuristic algorithms were considered, including Lemurs, Osprey Optimization Algorithm, and Tuna Swarm Optimization. The proposed ensemble approach involves data preprocessing, feature selection using metaheuristic algorithms, and training/testing of DL models such as EfficientNetV2, RegNetY-8GF, and MaxViT. The proposed method was tested on a plant disease dataset named NZDLPlantDisease-v1 which contains 20 classes of healthy and diseased plants in multiple plant organs, such as leaves, fruits, and stems, for five different fruits. The results showed a reduction in the training time to 2.9 min, while improving the F1-score by 1.62%. The robustness of the proposed methodology was analysed using an external testing dataset containing images curated from other agricultural environments.