Crop-Specific Effects of Data Augmentation on Classification Accuracy in Sentinel-2 Imagery

This study explores the impact of data augmentation techniques on crop classification using Sentinel-2 multispectral imagery. The DACIA5 dataset (2020-2024) comprising 12 crop classes and high-quality in situ labels was used to train a ResNet18 model adapted for 12 -channel multispectral input to perform crop classification. An imbalanced data sampling strategy was employed to address class disparities, and various augmentation techniques were tested to assess their impact on model performance. The outcomes of this research show that the augmentation effects vary significantly by crop type, with some techniques improving classification for certain crops while degrading performance for others. These findings contribute to our understanding of the challenges associated with crop-specific classification in remote sensing applications and underscore the need for tailored augmentation approaches.