Comparative analysis of optimization algorithms for maxvit-based tomato leaf disease detection in precision agriculture
Keywords:
AdamW, Multi-axis Vision Transformer, Optimization algorithms, SophiaG, Tomato leaf diseasesAbstract
The detection methods used to identify crop diseases require both accuracy and speed because they play a vital role in achieving better crop yields and sustainable agricultural practices. This study conducts an extensive assessment of six optimization algorithms, which include AdamW, Adam, Stochastic Gradient Descent (SGD), Adagrad, RMSprop, and SophiaG, to evaluate their performance within the Multi-Axis Vision Transformer framework for classifying tomato leaf diseases. In this study, a subset of the PlantVillage dataset was employed for model training and evaluation. The experiments were conducted in PyTorch on Google Colaboratory with NVIDIA Tesla T4 GPU acceleration. The MaxViT-Tiny model used pretrained ImageNet-1k weights to establish its initial state, which developers used to train the model through different optimizers with fixed hyperparameters. Model performance was evaluated using F1-score, precision, recall, and accuracy metrics calculated with scikit-learn, along with confusion matrix analysis and convergence curve visualization. AdamW achieved the best performance results among all tested optimizers because it reached an F1 Score and Recall combined with an overall accuracy of 99.89%. SophiaG and Adam achieved similar results, while Adagrad showed the least classification success with 98.32% accuracy, which demonstrated how standard adaptive optimizers fail to manage high-dimensional visual tasks in controlled environments. The results demonstrate that advanced optimization techniques enable technological progress in agricultural smart systems and deep learning capabilities used for plant disease research, which depends on transformer models.