Rainfall Prediction using Spatial Convolutional Neural Networks and Recurrent Neural Networks

Rainfall is influenced by climate factors such as air temperature, humidity, rainfall, wind speed, and the Southern Oscillation Index (SOI). Microclimate allows local rain to occur, so it is necessary to consider climatic variables from some observation stations. This research involved multi variables of three stations for spatial analysis. Each variable is recorded in time series. So, this paper proposed spatial and temporal analysis in predicting weekly rainfall. Spatial information was obtained from climate variables of three adjacent Meteorological, Climatology, and Geophysics Agency (BMKG) stations: Tangerang Geophysics Station, Budiarto Meteorology Station, and South Tangerang Geophysics station, for twelve years (2010-2021). The 2D Convolutional Neural Network (CNN) and Recurrent Neural Network (RNN) methods were proposed to extract spatial-temporal features from climate data. As a result, the proposed model had the best accuracy of 87.80% compared to the 1D CNN model, with an average accuracy of 80.21%. This study shows that spatial features are essential to increase accuracy because the surrounding weather variables influence each other, and there needs to be a correlation in modeling. In addition, this research also compares the proposed model with the 3D CNN method. As a result, the accuracy of the 2D CNN-RNN model outperformed the 3D CNN by 12.46% higher because 3D CNN extraction was too dependent on the extraction of spatial features and lacked optimizing temporal information.