Deriving daily water levels from satellite altimetry and land surface temperature for sparsely gauged catchments: A case study for the Mekong River

Abstract

Space borne radar altimeters often complement in-situ water levels to provide greater insights to hydrodynamic models in sparsely gauged catchments. However, 10-day or 35-day water levels derived from satellite radar altimetry are generally too infrequent for flood forecasting or hydrodynamic modelling purposes. This paper proposes a new approach to find daily water levels for areas where in-situ river heights are not available. The new approach is based on the relationship between river height and difference between daytime and night time land surface temperatures (LST). This relationship is first demonstrated using in-situ gauge data to explore appropriate statistical models to predict river height using LST differences. The approach is then applied to a number of virtual stations at the intersections between the Mekong River and the ground-tracks of Jason-2 satellite altimetry which gives 10-day water level time series.The LST difference from the thermal infrared (TIR) observations of Moderate Resolution Imaging Spectro-radiometer (MODIS) is shown to have a strong relationship with in-situ water levels and a good relationship with the Jason-2 water levels. The models included a simple linear regression model which was then extended to firstly include seasonal terms and secondly assimilate satellite altimetry data. A regression model tree (M5) was also considered but was found to be inferior to the seasonal linear model. The developed regression models were used to predict daily water levels to infill the 10-day Jason-2 altimeters. RMSE of modelled daily water levels at gauges is between 0.3 m to 0.6 m, while RMSE of modelled daily water levels without using in-situ data is higher varying from 1.4 m to 1.9 m. The temporal correlation between the modelled water levels using satellite altimeters at virtual stations and in-situ water levels at adjacent gauges ranged from 0.72 to 0.86. These results show the potential of the proposed approach to produce high temporal resolution water levels for flood models or other applications using only remotely sensed data.