Accurate identification of medicinal plants is essential in traditional medicine practices. It also plays an important role in biodiversity conservation. This work aims to improve data augmentation, class weighting, and fine-tuning strategies in MobileNetV3 to accurately identify Indian medicinal leaves of 56 classes. The dataset is slightly imbalanced as the number of samples per class ranges from 69 to 177. Our experiments showed that the baseline model with no augmentation and no class weights achieved a test accuracy of 88.99%. The experiment in which we applied both targeted augmentation and class weights achieved a test accuracy of 90.59%. We also performed partial fine-tuning of 15 and 30 layers. The 15-layer fine-tuning showed competitive performance with faster convergence. By comparing these different strategies, we were able to understand how these strategies contribute to better performance. Overall, this study provides actionable insights for researchers to optimize transfer learning strategies in fine-grained leaf classification tasks.