GP-VAE DATA FUSION FOR INTEGRATING HETEROGENEOUS REMOTE SENSING IMAGES WITH GIS PRIORS

The fusion of remote sensing data from diverse platforms, such as satellites and drones, presents a unique opportunity to enhance geospatial analysis by leveraging their complementary spatial and spectral resolutions. However, misalignment issues caused by inherent disparities in resolution and cover- age often limit the effectiveness of existing fusion techniques. This study introduces a novel methodology that integrates Gaussian Processes (GPs) for spatial regression and Variational Autoencoders (VAEs) for spectral feature encoding, guided by GIS priors to address these discrepancies. Using moderate-resolution Planet Dove imagery (RGB bands) and high-resolution Maxar WorldView data (8 bands), we demonstrate how this hybrid approach improves the consistency of reconstructed maps. Experimental results show a 43% improvement in SSIM scores and significant gains in MAE and LPIPS metrics compared to traditional GP-based methods. By addressing spatial and spectral discrepancies, our framework facilitates the integration of multi-source remote sensing data, enabling enhanced accuracy and reliability in applications such as environmental monitoring, urban planning, and resource management.