Paper Title
Developing AI/ML Geospatial Fusion Framework Using Earth Observation Satellite Data

Abstract
This paper presents a deep learning framework for geospatial data fusion of upper tropospheric humidity (UTH) observations from multiple Earth observation satellites. The approach leverages convolutional and transformer-based neural network architectures to reconstruct spatially complete humidity fields with improved accuracy. Experiments were conducted on UTH data collected from Metop-A, Metop-B, NOAA-18, and NOAA-19 sensors over the Indian region during August 2018—a period characterized by significant cloud cover and monsoon activity. The proposed models were rigorously evaluated using metrics such as Mean Squared Error (MSE), Structural Similarity Index Measure (SSIM), and Peak Signal-to-Noise Ratio (PSNR). Results demonstrate that the nested U-Net++ architecture achieved superior reconstruction quality compared to both the baseline U-Net++ up sampling model and the transformer-based model, achieving a mean PSNR of 44.99 dB and SSIM of 0.9963. The findings highlight the potential of deep learning for reliable and scalable fusion of geospatial satellite measurements to supportatmospheric science applications.