Ensemble based deep learning model for prediction of integrated water vapor (IWV) using GPS and meteorological observations

Abstract

Abstract Integrated water vapor (IWV) has been widely perceived through machine learning (ML) strategies. During this investigation, we employed IWV time series from weather stations to determine the oscillations and patterns with IWV across two latitudes namely VBIT, Hyderabad (India) and PWVUO station, Oregon (US). The GPS derived IWV and meteorological data such as pressure (P), temperature (T) and relative humidity (RH) dataset for the year 2014 has been taken from VBIT station and from PWVUO station for 2020. Five machine learning algorithms namely Optimized Ensemble (OE) model, Rational Quadratic Gaussian Process Regression model (RQ-GPR), Neural Networks model (NN), Cubic Support Vector Machine (CSVM) and Quadratic Support Vector Machine (QSVM) algorithms are used. The GPS derived IWV data revealed the maximum variation during summer monsoon period specifically in the month of July. The correlation analysis between GPS-IWV and optimized ensemble technique showed the highest correlation for the VBIT station with correlation coefficient as (ρ) = 99 % and at PWVUO station as (ρ) = 88 % for two different datasets. The residual analysis has also showed less variation to the optimized ensemble model. The performance metrics obtained for OE at VBIT station are mean absolute error (MAE) as 0.64 kg/m2, mean absolute percentage error (MAPE) as 3.80 % and root mean squared error (RMSE) as 0.94 kg/m2 and at PWVUO station the values are MAE = 1.91 kg/m2, MAPE = 11.76 % and RMSE as 1.97 kg/m2, respectively. The results explained that the OE method has shown a better performance compared to the remaining models.