Spectral Super-Resolution Reconstruction of Multispectral Remote Sensing Images via Clustering-Based Spectral Feature

Hyperspectral (HS) remote sensing imagery has broad applications but faces challenges in acquiring high-quality wide-swath data due to sensor hardware limitations. To address this, we proposed a jointly fused convolutional neural network for spectral super-resolution (JF-CNNSSR), leveraging spectral reflectance variations across different land cover types. The method employs paired overlapping multispectral (MS) and HS images to develop two spectral mapping networks: a 1-D-CNN spectral mapping network and a 3-D-CNN spectral mapping network, trained class-specifically on clustered spectral samples. A spectral distance-based joint weighting mechanism integrates outputs from both networks to synthesize HS imagery spatially consistent with the input MS data. Experiments on four benchmark datasets demonstrate that JF-CNNSSR outperforms existing spectral super-resolution methods, achieving significant improvements across key quantitative metrics including peak signal-to-noise ratio, spectral angle mapper, structural similarity index measure, ERGAS, and MSE. Practical validation through grassland coverage estimation in China’s Three-River Source Region further confirms the superiority of reconstructed HS imagery over original MS data in ecological application accuracy.