Multi-Label Satellite Image Classification using Lightweight Vision Transformers

To address the shortcomings of traditional deep learning applications for multi-label satellite image classification, this work shifts focus toward lightweight Vision Transformers (ViTs) that involve fewer parameters and can be deployed in resource-constrained devices. ViT offers a promising alternative to Convolutional Neural Network (CNN) due to its unique architecture which is designed to capture long-range dependencies more effectively. While CNN relies on convolutional layers to extract features, ViT divides images into patches and processes them as token sequences, similar to the manner text is handled in language processing, which is comparatively way better than CNN. This approach enables ViT to capture broader and more detailed features across the entire image, essential for dealing with multi-label satellite images that involve varying categories, sizes, and spatial arrangements. In our study, we employ six lightweight ViT models, namely, ViT-Small, ViT-Tiny, ViT-Base, Swin-Tiny, DeiT-Tiny, and DeiT-Base, to effectively capture different sets of labels associated with Amazon rainforest satellite imagery, leveraging extensive prior training to better recognize complex patterns and relationships within the data. The ViT-Tiny model achieves the best F-beta score (beta=2) of 0.9252, surpassing the performance of the baseline ViT model and other lightweight ViT architectures.