Publications

Hou, JL; Huang, CL; Zhang, Y; You, YH (2022). Reconstructing a Gap-Free MODIS Normalized Difference Snow Index Product Using a Long Short-Term Memory Network. IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, 60, 4304914.

Abstract
Atmospheric disturbance, sensor malfunctions, and other factors can cause serious gap pixels in the Moderate Resolution Imaging Spectroradiometer (MODIS) normalized difference snow index (NDSI) products. In this article, MODIS NDSI gap pixels are reconstructed in a highly heterogeneous area with drastic snow accumulation and melting changes using a long short-term memory (LSTM) network. Three LSTM-based MODIS NDSI gap pixel reconstruction schemes, i.e., forward, backward, and bidirectional LSTM networks that separately use earlier, subsequent, and integrated earlier and subsequent timestamp information, are developed. NDSI information for the gap pixel is restored using the long-term spatiotemporal information for this pixel and its adjacent pixels. A case study of NDSI reconstruction in the source area of the Yellow River, northwestern China, during the 2018-2019 snow season, demonstrates that all three LSTM-based schemes can reliably generate spatiotemporally continuous NDSI data with an accuracy comparable to that of the original MODIS NDSI products under clear-sky conditions. The bidirectional LSTM-based scheme, which has the best performance, can achieve a desirable overall accuracy of 89.93%, with an omission error of 3.82% and a commission error of 6.25%, in terms of dichotomous evaluation based on in situ snow depth observations. The R-2, average RMSE, overestimation error, and underestimation error are 0.95, 5.13%, 5.39, and 6.40%, respectively, in terms of the continuous value assessment based on the gap pixels assumption. Our results demonstrate the reliability and feasibility of the LSTM-based schemes in recovering the missing values in MODIS NDSI products by deeply excavating the spatial continuity and long-time series dependence of the snow cover.

DOI:
10.1109/TGRS.2022.3178421

ISSN:
1558-0644