Better with fewer features: climate dynamics estimation for Van Lake basin using feature selection.

Area of Science:

• Environmental Science
• Data Science
• Climatology

Background:

• Machine learning (ML) and statistical methods are widely used for forecasting climatological parameters.
• Feature selection is often overlooked in existing ML-based forecasting models, potentially limiting their efficiency and accuracy.

Purpose of the Study:

• To develop and evaluate ML prediction models incorporating feature selection for one-step-ahead forecasting of temperature and evapotranspiration.
• To assess the performance of these models for long-term climate estimations (30 years).

Main Methods:

• Implementation of ML models with feature selection for climatological parameter estimation.
• Utilizing the Bayesian Ridge Regressor (BRR) and comparing its performance against other regressors.
• Experimental validation on data from three stations in the Van Lake Closed basin, Turkey.

Main Results:

• The Bayesian Ridge Regressor (BRR) demonstrated superior performance compared to other models, achieving R² scores between 0.961 and 0.988.
• Feature selection enhanced model performance, enabling comparable or better results with a reduced number of input features.
• The developed models require non-sparse and complete time series data for optimal performance.

Conclusions:

• Feature selection significantly improves the efficiency and accuracy of ML-based climate forecasting models.
• The BRR model, combined with feature selection, provides a robust approach for predicting temperature and evapotranspiration.
• The methodology's limitation lies in its requirement for complete time series data, posing challenges for sparse datasets.