Reconstruction of GRACE terrestrial water storage anomalies using Multi-Layer Perceptrons for South Indian River basins.

Abstract

The Gravity Recovery and Climate Experiment (GRACE) satellite mission began in 2002 and ended in June 2017. GRACE applications are limited in their ability to study long-term water cycle behavior because the data is limited to a short period, i.e., from 2002 to 2017. In this study, we aim to reconstruct (1960-2002) GRACE total water storage anomalies (TWSA) to obtain a continuous TWS time series from 1960 to 2016 over four river basins of South India, namely the Godavari, Krishna, Cauvery and Pennar River basins, using Multilayer Perceptrons (MLP). The Seasonal Trend Decomposition using Loess procedure (STL) method is used to decompose GRACE TWSA and forcing datasets into linear trend, interannual, seasonal, and residual parts. Only the de-seasoned (i.e., interannual and residual) components are reconstructed using the MLP method after the linear trend and seasonal components are removed. Seasonal component is added back after reconstruction of de-seasoned GRACE TWSA to obtain complete TWSA series from 1960 to 2016. The reconstructed GRACE TWSA are converted to groundwater storage anomalies (GWSA) and compared with nearly 2000 groundwater observation well networks. The results conclude that the MLP model performed well in reconstructing GRACE TWSA at basin scale across four river basins. Godavari (GRB) experienced the highest correlation (r = 0.96) between the modelled TWSA and GRACE TWSA, followed by Krishna (KRB) with r = 0.93, Cauvery (CRB) with r = 0.91, and Pennar (PCRB) with r = 0.92. The seasonal GWSA from GRACE (GWSAGRACE) correlated well with the GWSA from groundwater observation wells (GWSAOBS) from 2003 to 2016. KRB exhibited the highest correlation (r=0.85) followed by GRB (r=0.81), PCRB (r=0.81) and CRB (r=0.78). The established MPL technique could be used to reconstruct long-term TWSA. The reconstructed TWSA data could be useful for understanding long-term trends, as well as monitoring and forecasting droughts and floods over the study regions.