Deep Learning for Enhanced Representation of Multispectral Imagery in Remote Sensing

The increasing availability of multispectral satellite data has opened new possibilities for Earth observation, yet its full potential is often limited by the intrinsic trade-off between spatial resolution and spectral richness. This thesis investigates deep learning methods to enhance the spatial and spectral representation of multispectral imagery in remote sensing, addressing both supervised and self-supervised learning paradigms. In the supervised setting, the Swin2-Mixture-of-Sparse-Experts (Swin2-MoSE) model is introduced for Single-Image Super-Resolution (SISR) of satellite images. Building upon hierarchical vision transformers, Swin2-MoSE integrates sparsely gated Mixture-of-Experts layers to dynamically allocate representational capacity across spatially heterogeneous regions. A hybrid objective combining Normalized Cross-Correlation and Structural Similarity metrics is employed to preserve fine details and maintain cross-band consistency. Experiments across multiple datasets demonstrate consistent gains over existing transformer-based SISR models, particularly at higher upscaling factors. In the self-supervised domain, the Wavelet Masked Autoencoder (WaveMAE) is proposed to learn robust, transferable representations from large unlabeled multispectral archives. WaveMAE replaces pixel-space reconstruction with frequency-domain reconstruction by predicting discrete wavelet coefficients, thereby encouraging the encoder to capture multi-scale spectral and spatial correlations. The model incorporates a Geo-conditioned Positional Encoding based on spherical harmonics to exploit geographical context, improving feature alignment across diverse scenes. Evaluations on the PANGAEA benchmark confirm that WaveMAE significantly outperforms prior self-supervised methods on downstream tasks such as segmentation and regression. Together, these contributions advance the representation quality of multispectral imagery and highlight the complementary strengths of supervised and self-supervised paradigms for remote sensing. The findings establish a foundation for future large-scale, frequency-aware, and geographically conditioned foundation models for Earth observation.