Evaluating Preprocessing and Data Augmentation for Soybean Leaf Pest Classification with Deep Learning

Soybean is one of the most important crops in the world, and its productivity is often reduced by pests that attack the leaves. Detecting these problems early is important for effective crop management. In this work, we evaluate three deep learning models, ResNet-50, ViT-B16, and ConvNeXt-Tiny, for the classification of soybean leaves as healthy or infested by Caterpillar or Diabrotica speciosa. The experiments were carried out using the public "Images of Soybean Leaves" dataset collected in real field conditions. We tested two preprocessing strategies (Resize and Random Resized Crop) and three levels of data augmentation (none, mild, strong). Results show that ConvNeXt-Tiny obtained the most stable performance, with its best results under Random Resized Crop and strong augmentation (accuracy of 0.9594, precision of 0.9555, and F1-score of 0.9576). ResNet-50 also performed well with cropping strategies, while only ViT-B16 benefited from simple Resize preprocessing. Macro and weighted metrics showed that the models kept good performance even with class imbalance. These results confirm the potential of deep learning for automatic pest detection in soybean leaves. Future work includes testing with larger datasets and using more advanced augmentation methods to improve robustness in real field conditions.