Predicting surface temperature in Lake Villarrica (Chilean Patagonia) using a long short-term memory model

Abstract

In this study, we analyze water-temperature time series was measured over 34 years, between 1986 and 2020, at the water surface at seven stations across Lake Villarrica (Southern Chile). The spring and summer seasons show an increment in the superficial temperature during the study period. The annual maximum temperature, ranging between 17.35 and 21.65 °C were observed in 1997 and 2009, respectively, while the annual minimum, ranging between 16.8 and 21.5 °C were observed in 2001 and 2009, respectively. In addition, we employ a machine learning based estimation model to predict surface temperatures in a South American lake spanning the period 1989 to 2021. Our model uses data in situ of physical, chemical, and biological parameters of lake quality water, along with meteorological data and spectral bands, including combinations of images from the Landsat 8 satellite, as input variables. The 7 lake monitoring stations were classified into 4 regions according to their geographical location: north, south, east, and west. Our findings demonstrate the exceptional performance of the long short-term memory (LSTM) model in accurately estimating temperatures across Lake Villarrica. The best results were obtained for the west region of the lake with good statistical metrics from the estimation model of RMSE = 2.79, Bias =−0.06, max error = 5.93, MSE = 7.83 and median absolute error (MedAE) = 2.13. This approach represents a significant advance in the integration of remote sensing and machine learning techniques to monitor and manage inland water systems.