Investigating the Crucial Factors Impacting the Accuracy of Machine Learning Models for Lake Surface Water Temperature and Ice Thickness Prediction

Climate change is having a profound impact on freshwater ecosystems, with lakes being highly sensitive to environmental changes. In recent years, the increasing popularity of machine learning techniques in this field can be attributed to their ability to analyze intricate patterns and relationships in meteorological data, lake dynamics and their response to climate change. This study aims to explore how meteorological variables impact lake surface water temperature and ice thickness. First, we evaluate nine machine learning algorithms for lake surface water temperature prediction in synthetic lakes, comparing their performance and investigating the effects of input variables on accuracy. Our study, based on a numerical model (distinct from a real lake), indicates that considering air temperature and day of the year suffices for acceptable outcomes. Additional predictors provide minimal improvement. Furthermore, better results can be achieved by pre-processing air temperature through time averaging or incorporating past values specially for deep lakes. Despite exploring various machine learning algorithms with the same inputs, no single optimal choice emerged (although artificial neural networks exhibited slightly better results). Second, we select a machine learning technique, artificial neural network, to model the influential factors of lake surface water temperature response on 2024 lakes worldwide, based on the CCI Lakes dataset. Our analysis reveals that, in general, the day of the year is the most relevant factor, suggesting that the mean (climatological) year is already a good approximation. Removing it from the set of predictors, air temperature, shortwave and downward longwave radiation and relative humidity gain the predominant roles and the incorporation of other meteorological variables could significantly or moderately improve the models' performance across different climatic zones. Third, we investigate the influence of various meteorological variables on ice thickness prediction in two distinct lakes in Sweden using artificial neural network. Among the input variables, the day of the year assumes a significant role in simulating ice thickness. Additionally, shortwave radiation and specific humidity prove to be pivotal predictors. During the period of ice formation, aside from the day of the year, the negative degree days linked to negative air temperature also stand out as influential predictors. Our findings demonstrate that machine learning techniques offer a promising avenue for studying lake dynamics and their response to environmental changes because of being flexible to change the input variables and to analyze their importance on the model leading to understand the physics behind it. Through the random regeneration of each feature, we can assess its impact on the model by measuring the extent to which it reduces the model's performance compared to the model incorporating all variables. The choice of meteorological variables plays a critical role in model performance, emphasizing the need to select relevant input variables for optimal results.