Stochastic Dynamical Modeling for Time-series Zero-Degree Isotherm Heights Prediction over Tropical Climate

Abstract The accurate prediction of zero-degree isotherm height plays a significant role in weather forecasting and climate modelling, which have applications in agriculture, aviation, telecommunications, and disaster management. Climate action as one of the Sustainable Development Goals (SDGs) can only be achieved continuous study and understanding of our complex atmosphere. The dynamics of the atmosphere is a hot issue in research due to the random process or random probability distribution of the parameters associated with the environment, in which rain height is a key factor. Advanced Deep Learning (ADL) has achieved reasonable results in many fields and the associated applications, especially in the accurate prediction of time-series events. In this study, we adopted stochastic dynamical models to forecast Zero-Degree Isotherm Height (ZDIH) in Akure, a tropical region located in the Southwestern Nigeria. Existing methods, including traditional statistical approaches and ITU-R recommendations, have shown limited accuracy and generalizability due to their inability to capture intricate patterns in atmospheric data. One of the goals of this research is to identify gaps in the existing methods, such as their reliance on shallow learning techniques and their inadequacy in processing large volumes of data with high variability. This research proposes a new approach by applying the Seasonal AutoRegressive Moving Average (SARIMA) and two ADL models, namely; the Long Short-Term Memory (LSTM), and the hybrid LSTM, referred to as Long Short-Term Memory - Convolutional Neural Network (LSTM-CNN) model. Six years (2005-2010) of archived data from the National Aeronautics Space Administration (NASA) are used for the prediction. The result shows seasonality and a fairly stable trend over the yearly analysis. The ZDIH is generally low throughout the rainy season with a mean value of 4.75 km while mean of 4.94 km was observed during the dry season. Validation and comparison of the models revealed that SARIMA (2, 1, 1)(1, 1, 1, 12) model has a better performance over the other two models with RMSE and r 2 of 46.3 and 0.89, respectively. The LSTM provides a fair performance, while the hybrid LSTM-CNN model performed poorly. The outcome of the research suspects the seasonality feature (i.e parameter “s”) of SARIMA could be responsible for its higher prediction accuracy over the ADL models. The results obtained demonstrate that the model significantly improves forecasting accuracy and it is hereby recommended for prediction of ZDIH the tropical region.