Compressed Domain Remote Sensing Scene Classification Using Vision Transformer based Autoencoding and Data Efficient Image Transformers

Deep transformer architectures are increasingly being used in remote sensing scene classification as they capture scene contextual dependencies on a global scale, but since they use full-resolution images this means that they require large storage requirements and high computational costs reducing their ability to scale in large scale and resource constrained systems. To meet this, this paper presents a compression-aware transformer architecture, which is called VITAE-DeiT, which collaboratively learns both semantic compression and classification through a single learning pipeline. The suggested method provides a Vision Transformer-based autoencoder to create small but semantically discriminative latent representations, which then allows performing scene classification in the compressed space, using a small transformer. In contrast to the traditional approaches, which consider compression and recognition as two distinct processes, the presented framework explicitly incorporates into the compression process semantic preservation and thus reduces as much as possible information loss, which is important in the classification process. Decades of benchmark remote sensing dataset experiments show that the suggested approach yields a test accuracy of 94.44, good precision (95.32), recall (94.50), and MCC (0.9433), which is superior to the recent CNN, hybrid, and transformer-based baselines. In addition, the compression-conscious architecture can be used to minimize the effective complexity of inference, attaining 21.3ms per image using moderate model complexity. Ablation and complexity tests also support that it is the synergistic integration of transformer-based semantic compression and model global context that contributes to the performance benefits. These findings make VITAE-DeiT a scalable and effective model of high-accuracy, remote sensing, scene classification at storage and computation constraints.