NB_Re3: A novel framework for reconstructing high-quality reflectance time series taking full advantage of high-quality NDVI and multispectral autocorrelations

Multispectral reflectance — signals reflected from the Earth's surface across different wavelengths — is a primary data source for most remote sensing applications. However, obtaining complete cloud-free multispectral reflectance time series during the vegetation growing season remains challenging due to cloud contamination and limitations of existing reconstruction methods. To address the challenge, this study proposed a novel NDVI guided Bi-directional Recurrent Reconstruction model for multispectral Reflectance time series (referred to as “NB-Re3”), which aimed to reconstruct dense time series of reflectance images by exploiting the dependence of NDVI on multispectral reflectance. NB-Re3 utilizes a Temporal Convolutional Network (TCN) to capture the temporal trends in the NDVI data and a Bidirectional Long Short-Term Memory (Bi-LSTM) to integrate the temporal features of the NDVI with the cloud-free reflectance data. The architecture establishes a robust dynamic NDVI-reflectance relationship while capturing temporal dependencies and multispectral autocorrelations of multiple spectral bands. We compared the performance of NB-Re3 with four representative methods (MNSPI, HANTS, STAIR, and U-TILSE) in reconstructing multispectral reflectance time series, ranging from the visible bands, near-infrared and short-wave infrared bands, at two challenging sites: the irrigated area of Colleambally, Australia, and the cultivated area of Rikaze on the Southern Tibetan Plateau, China. The result showed that NB-Re3 kept superiority with the lowest root mean square error values and highest correlation coefficients values. The effectiveness of integrating high-quality NDVI time series and using multispectral autocorrelation to improve reflectance time series reconstruction was further confirmed by the ablation experiments. It is concluded that NB-Re3 shows promise for generating long-term cloud-free reflectance time series products tailored for ecological and agricultural applications.