A Novel Deep Learning-based Landsat 7 ETM+ Multi-Spectral to Hyperspectral Reconstruction Model: Application for Water Bodies in an Indian Region

Abstract

Hyperspectral (HS) remote sensing has the capacity to provide finer spectral information and better identification of objects while multispectral (MS) data are more readily available but with fewer bands. In absence of HS data, spectral reconstruction from MS to HS data can be considered in order to enhance the applicability of HS data. In this study, a deep learning-based Multi-Head Attention enabled Multi-Layer Perceptron (MHA-MLP) model is developed to reconstruct a scene of EO-1 Hyperion (HS) from a Landsat 7 ETM+ (MS) image. The reconstruction is also done using Multi-Layer Perceptron (MLP) and Transformer models. The reconstructed data from the three models is studied for water body locations in Betul, Madhya Pradesh, India. The results are analysed by comparative study of the spectra and calculation of standard statistical metrics. The reconstructed spectra from the MHA-MLP are found to follow the original HS spectra more closely than the other models and show the best values in the statistical metrics. Hence, The MHA-MLP model is found to be the best HS reconstructor model when compared to MLP and Transformer models. The reconstructed spectra are capable of capturing the 670 nm notch important for the study of chlorophyll concentration. This model can be used for water quality assessment applications and can also be extended for other applications. © 2024 IEEE.