Benchmarking Vision Transformers for Satellite Image Classification based on Data Augmentation Techniques

This article evaluates the implications on transformer model performance in satellite image classification by means of numerous data augmentation techniques using the Eurosat dataset. We examine the Swin-tiny, Swin-small, Convit-small, and Crossvit-small models under several augmentation methods including MixUp, CutMix, Geometric, WGP-GAN, and DCGAN. Our findings demonstrate that Mixup and WGP-GAN augmentations significantly enhance model performance; Swin small achieves 99.26% test accuracy the best. CutMix was more helpful for Swin-small than for Swin-tiny; geometric augmentation improved Swin-tiny and Crossvit-small. DCGAN behaved differently on many models. These results highlight the significance of selecting appropriate augmentation techniques suited for certain model architectures to increase performance in assignments needing satellite image classification.