MSC-D3Net: A Resilient Multi-Scale Learning Framework for Adaptive Cross-Domain Scene Understanding in Remote Sensing

Abstract

Remote sensing scene classification is important for accurate earth observation, environmental monitoring, and geographic analysis. However, intra-class diversity and domain specific variations make this classification quite difficult. This study aims to design a model that can generalize well on cross domain datasets without accessing target domain data. For this purpose, a novel deep learning model “MSC-D3Net” is proposed that combines CNN and ViT features within hierarchical domain disentanglement and cross scale semantic alignment. An adversarial domain discriminator module, along with uncertainty calibration, is also integrated. The model is trained and tested on three publicly available datasets and achieved very high in domain accuracies above 98% and cross domain accuracies above 93% on all datasets. The model also demonstrates low uncertainty errors and outperforms existing architectures by a large margin.