Synthetic Abundance Maps for Unsupervised Super-Resolution of Hyperspectral Remote Sensing Images

A team from IP Paris and Télécom Paris has developed a novel, unsupervised AI framework for enhancing the spatial resolution of hyperspectral remote sensing images. The core innovation addresses a major bottleneck in the field: most existing super-resolution methods are supervised and require high-resolution ground truth images for training, which are often impossible to obtain for satellite or aerial imagery. Their proposed method cleverly sidesteps this need by first 'unmixing' a low-resolution hyperspectral image into its fundamental components—spectral signatures (endmembers) and their spatial distributions (abundances).

Instead of training on real high-resolution data, the team trains a neural network to perform super-resolution on synthetically generated abundance maps. These synthetic maps are created using a statistical 'dead leaves' model, whose parameters are derived directly from the characteristics of the low-resolution input image and the known blurring function (point spread function) of the sensor. This ensures the synthetic data is statistically representative of the real task. Once trained, the network enhances the resolution of the original image's abundances, which are then recombined with the endmembers to produce the final high-resolution hyperspectral image.

The method's effectiveness was rigorously tested, demonstrating strong performance across three different datasets, three scaling factors (2x, 3x, 4x), and multiple evaluation metrics. This proves both the viability of using synthetic data for training and the overall robustness of the pipeline. By removing the dependency on scarce ground truth, this work significantly advances the practical application of AI for analyzing satellite imagery, enabling higher-resolution insights into agriculture, environmental monitoring, and geology where perfect training data doesn't exist.