Satellite-Based Object Detection With PSNR-Driven Image Enhancement and Structural Interpretability

Satellite image analysis faces persistent challenges in real-time processing, classification across heterogeneous terrains, and limited model interpretability. Ensuring input image quality before inference remains a critical bottleneck in operational remote sensing. To address these issues, we propose a unified deep learning framework that combines quality-assured preprocessing, efficient training, and interpretable multi-label classification for high-resolution imagery. Central to the framework is the Swin Vision Transformer, integrated within a GPU-accelerated pipeline featuring Albumentations-based augmentation and enhancement techniques (distortion correction, histogram equalisation, denoising), validated by structural Similarity Index Measure (SSIM) (0.9564) and Peak Signal-to-Noise Ratio (PSNR) (30.11 dB). Confidence-based prediction filtering improves reliability, while specialized modules for Land Use and Land Cover (LULC) and coastal region analysis enhance semantic understanding. Interpretability is achieved using SHapley Additive exPlanations (SHAP), Local Interpretable Model-Agnostic Explanations (LIME), and occlusion maps to visualize spatial attention. On a curated 9-class subset of the MLRS-Net dataset, the model achieves 99.74% validation accuracy with up to 99.9% confidence in real-time inference. The proposed framework delivers a scalable, robust, and explainable solution for real-time satellite monitoring applications.