Remote sensing and environmental monitoring have undergone transformative advances through the integration of advanced image processing techniques, particularly deep learning architectures. This systematic survey examines the evolution, current state, and comparative performance of image processing methods applied to remote sensing imagery. This review establishes a structured taxonomy categorizing techniques by approach (traditional vs. deep learning) and learning paradigm (supervised, unsupervised, and self-supervised). Prior work is critically synthesized; key strengths are identified, including the dominance of convolutional neural networks (CNNs), which achieve accuracies exceeding 97% on benchmark datasets such as EuroSAT, and limitations, such as computational complexity, data requirements, and generalization challenges. The experimental comparison section provides a detailed analysis of specific architectures (ResNet, U-Net, VGG, Vision Transformers) evaluated on standard benchmarks (EuroSAT, BigEarthNet, UC Merced, Sen12MS) with comprehensive performance metrics. Critical research gaps are identified, including the need for improved model efficiency, better handling of multispectral and hyperspectral data, enhanced generalization across diverse geographic regions, and development of interpretable models. This review serves as a comprehensive resource for researchers and practitioners in remote sensing, providing actionable insights for future research directions and practical applications in environmental monitoring. Key differentiators from existing surveys include a PRISMA-compliant search methodology, a unified taxonomy spanning traditional through diffusion-based methods, methodology-centric comparison tables, dedicated hyperspectral and diffusion model coverage, and explicit prioritization of research gaps.
Rao et al. (Thu,) studied this question.