Prediction of Integrated Water Vapor Using a Machine Learning Technique

Abstract

Long-term measurements of GNSS receiver at Gadanki, India, have been used to develop a machine learning technique – lightGBM for the prediction of integrated water vapor (IWV) with different lead times. A variety of data sets related to IWV (representing source, sink and transport) and short-scale features of IWV (gradients, sinusoidal pattern) have been used to train the model. Model performance is validated in different seasons and also on storm days. The predicted IWV at different lead times (30–120 min) perfectly captures the temporal variability of measured IWV with a correlation coefficient >0.99. The RMSE of predicted IWV with 30 minutes lead time is less than 1 mm in all seasons. Nevertheless, the RMSE for predicted IWV with longer lead times increases with lead time but always remains <3 mm. The bias is slightly larger during the monsoon, mainly due to the higher occurrence of longer duration rainy events. Even in those days, the model is able to accurately predict the enhanced IWVbefore the rain occurrence. Sensitivity analysis and feature importance analysis on different predictors used in the model reveal that the IWV features are more important for short-scale prediction, like 30 min, whereas the importance of other predictors is high for longer lead time prediction (1-2 hours) and on storm days.