Estimation and analysis of missing temperature data in high altitude and snow-dominated regions using various machine learning methods.

Abstract

Considering the importance of limited natural resources, accurately recording and evaluating temperature data is critical. The daily average temperature values obtained for the years 2019-2021 of eight highly correlated meteorological stations, characterized by mountainous and cold climate features in the northeast of Turkey, were analyzed by an artificial neural network (ANN), support vector regression (SVR), and regression tree (RT) methods. Output values produced by different machine learning methods compared with different statistical evaluation criteria and the Taylor diagram. ANN6, ANN12, medium gaussian SVR, and linear SVR were chosen as the most suitable methods, especially due to their success in estimating data at high (> 15 ℃) and low (< 0 ℃) temperatures. All the methodologies and network architectures used produced successful results (NSE-R2 > 0.90). Some deviations have been observed in the estimation results due to the decrease in the amount of heat emitted from the ground due to fresh snow, especially in the -1 ~ 5℃ range, where snowfall begins, in the mountainous areas characterized by heavy snowfall. In models with low neuron numbers (ANN1,2,3) in ANN architecture, the increase in the number of layers does not affect the results. However, the increase in the number of layers in models with high neuron counts positively affects the accuracy of the estimation.