Alteration mineral information extraction based on image super-resolution technology

Abstract

High-resolution remote sensing imagery is crucial for advancing earth science research. However, the scarcity of high-resolution data in specific non-visible spectral bands, such as short-wave infrared (SWIR) or thermal infrared (TIR), is challenge for various downstream tasks. To address this limitation, this study introduces a novel cross-band super-resolution (CBSR) method. This method improves the spatial resolution of targeted bands, specifically SWIR, within remote sensing images, and the methodology was tested with multi-band data from advanced spaceborne thermal emission and reflection radiometer (ASTER). The approach involves training a neural network on high-resolution visible and near-infrared single-band data, then the trained model is applied to generate high-resolution SWIR imagery. Validation was conducted over the Duolong Cu–Au porphyry district using ASTER imagery, focusing on mapping hydrothermal alteration. Through Bayesian optimization, the model achieved optimal performance with a peak signal-to-noise ratio of 43.17 dB, which is better than traditional methods. CBSR-reconstructed SWIR imagery, fused with principal component analysis, accurately delineated argillic alteration halos and ring structures. Furthermore, zero-shot transfer of the ASTER-trained model to Sentinel-2 imagery demonstrated the framework’s generalizability across sensor configurations. The proposed CBSR approach thus provides a robust, spectrally consistent mechanism for enhancing multi-resolution satellite data, with direct implications for mineral exploration, lithological mapping, and other domains reliant on high-fidelity SWIR information.