Self-Supervised Contrastive Learning Classification of Remote Sensing Images Based on SimCLR

Abstract

Large amounts of annotated data are typically needed for remote sensing picture classification, and pixel-level and scene-level annotations are relatively expensive due to the necessity for expert interpretation of land-cover categories, acquisition conditions, and spatial context. A useful technique for developing transferable representations of unlabeled remote sensing pictures is self-supervised contrastive learning. Because the conventional random cropping, color jittering, and negative sampling may change land-cover semantics or treat visually comparable landscapes as unreliable negatives, it is not optimal to directly apply SimCLR to remote sensing photos. A spectral-spatial SimCLR framework for classifying remote sensing images is presented in this research. Estimate spectral-spatial affinity between samples, create geospatially limited positive views, extract encoder-projection representations, and reduce uncertain hard negatives in contrastive optimization. The class-relevant spatial arrangement is maintained by adding a multi-scale consistency term. After self-supervised pretraining, a limited-label classification model is produced through prototype-assisted fine-tuning. The suggested approach outperforms supervised training, ImageNet pretraining, MoCo-style pretraining, BYOL-style pretraining, and vanilla SimCLR, achieving 94.28% overall accuracy, 93.71% macro-F1, and 0.936 kappa under 10% labeled fine-tuning, according to replaceable experimental results on a 45-class remote sensing scene dataset with 31,500 image tiles. The suggested remote-sensing-aware contrastive architecture enhances label efficiency, feature separability, and classification resilience based on the aforementioned investigation.