Methods of Spectral Matching for Remote Sensing Data

Abstract

Three methods of spectral matching for remote sensing data are studied: a pixelwise linear method, a pixelwise nonlinear method and a generalized nonlinear method. Nonlinear methods are implemented as a pair of multilayer perceptrons and a pair of convolutional neural networks respectively. Training and comparison of methods are performed using Landsat-8 and Sentinel-2 remote sensing images from 2021 IEEE GRSS Data Fusion Contest dataset. The root mean squared error (RMSE), the normalized mutual information (NMI) and the structural similarity index measure (SSIM) are used as metrics. A generalized nonlinear method demonstrates the best quality of spectral matching, achieving average values of RMSE = 0.048, NMI = 1.194 and SSIM = 0.887 over the testing set. A linear pixelwise method achieves RMSE = 0.075, NMI = 1.118 and SSIM = 0.847, a nonlinear pixelwise method achieves RMSE = 0.074, NMI = 1.117 and SSIM = 0.843. All methods show a significant improvement when compared to results without spectral matching (RMSE = 0.158, NMI = 0.119, SSIM = 0.585).