Development of Machine Learning Models for the Nearcasting of the Duration of Low-Visibility Events in the Indo-Gangetic Plains Regions of India

Abstract

The prediction of the duration of the occurrence of low-visibility events in a calendar day is a difficult process because of the complex and chaotic mechanisms of the onset and dissipation of the low-visibility events. However, it is most useful for the operation of airport services (scheduling of aircraft, optimal operations of the airports) and the planning of any activities (travel, tourism, agriculture, etc.). This research tries to build the best dynamic weighted ensemble of the best combination of base machine learning( ML) models (Light Gradient Boosting Machine (Light GBM), Random Forest (RF), and Support Vector Regression (SVR)) to accurately nearcast the duration of low-visibility events( fog (surface visbility<1000 m) and dense fog (surface visbility<200 m))  for a calendar day based on the initial conditions of 1500 UTC (Universal Time Co-Ordinate). Conditions such as surface meteorological parameters (air temperature, dew point temperature, relative humidity, wind (every 3 hours), rainfall (daily), and sunshine (daily)) and upper air meteorological parameters (wind, temperature, and relative humidity of 1000, 925, and 850 hPa (every 3 hours)) were taken into account to find the best set of explanatory factors for the accurate nearcasting of the duration of the low visibility events. The  Pearson correlation coefficient and Spearman's rank correlation coefficient were used to choose the final set of model explanatory variables. The datasets were thoroughly examined using supervised  ML algorithms at the various stages of training, testing, modelling, and cross-validation. All the best combination models' accuracy was evaluated and compared using performance measures, namely MAE (mean absolute error), RMSE (root mean square error), and R²( R squared error). Based on the coefficient of determination( R²), it can be observed that the suggested dynamic weighted ensemble model exhibits the best level of prediction accuracy, specifically 0.89 and 0.88 for the duration of fog and dense fog for a given lead time of a day. This surpasses the accuracy of LightGBM (0.79), RF (0.78), and SVR (0.76) for the prediction of the duration of fog. Therefore, this study highlights the potential of machine learning in facilitating the advancement of automation in airport scheduling and optimising the operations of airports, specifically in the fog-prone Indo-Gangetic Plains(IGP).