Critical factors for the use of machine learning to predict lake surface water temperature

Models based on Machine Learning (ML) are pervading all fields of science and practical applications, including the problem forecasting water temperature in lakes, a crucial variable for ecosystems proxy climate change. Here, we review most used ML algorithms this field highlight some physical constraints that should be carefully considered when adopting black-box approach. To illustrate them, refer to an artificial case study representing temperate lake simulated by means physically model, which take full control input output variables, restrict analysis surface (LSWT). Three main factors relevant successful prediction LSWT models: choice predictors (mostly, meteorological variables), their pre-processing (we tested three approaches), specific algorithm (nine different algorithms). We show selecting suitable inputs plays important role. In our study, is product numerical model not real lake, minimum amount information needed obtain acceptable results consider air (AT) day year. The use additional does substantially improve performances (the relative improvement RMSE was 7.75% test data set). also demonstrate better than normal obtained either averaging them over time window or values from previous days as model. Considering recent history forcing allows one comply with fact large mass makes lakes acting “filters” thermal response (thus, influenced AT days), changes depending lake’s depth. Eventually, did find definite answer about single optimal using same (although neural network had slightly results), suggesting insight into dynamics still factor exploitation ML.