DEVELOPMENT OF A TROPICAL CYCLONE PREDICTION SYSTEM BASED ON DEEP LEARNING USING SATELLITE IMAGE DATA

• Main Author: Sirait, Rinaldi
• Format: Theses
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/86186
• Institution: Institut Teknologi Bandung

Tropical cyclones are significant natural phenomena that can cause extensive damage to regions they pass through, especially in tropical areas. Due to the potential adverse impacts, accurate prediction of tropical cyclones is crucial for disaster mitigation. Tropical cyclone prediction involves forecasting the cyclone's trajectory and intensity. The initial approach used to predict tropical cyclones relied on traditional methods such as numerical weather models. Although these models provide an understandable insight into atmospheric dynamics, they have limitations in terms of complexity and reliance on the quality of input data. Moreover, these methods require substantial computing resources and take a long time during the inference process. This has led researchers to explore using alternative methods. Currently, researchers are beginning to use artificial intelligence-based approaches, particularly deep learning. This method has proven to provide much faster results compared to numerical methods, while accuracy is still being developed. This research will develop a deep learning method to predict tropical cyclones. SimVP-gSTA, as the latest model addressing spatiotemporal issues, will be used as the baseline model for predicting tropical cyclones. This model falls into the sequence to sequence prediction category. This research will use two types of data in the experiments: infrared data and heatmap data. Infrared data is constructed based on the IBTrACS history and Himawari-8, while heatmap data is built based on the historical center of tropical cyclones. The prediction results from the Sim VP­gSTA model will be forwarded into the cyclone center estimation model and the cyclone intensity estimation model to produce trajectory estimation and tropical cyclone classification. Therefore, this study will experiment with using the Sim VP­gSTA model with two different data inputs, then evaluate its performance using appropriate metrics for image prediction, namely RMSE, SSIM, and PSNR. Additionally, this study will also conduct experiments for estimating the trajectory and classifying tropical cyclones and evaluate the trajectory estimation results using Euclidean distance and evaluate the classification estimation results by calculating accuracy. The research results show that the use of the SimVP-gSTA model with infrared data produces an average SSIM score of 0.632 and a PSNR of 25.89 for 3-hour predictions and an average SSIM of 0.623 and PSNR of 22.94 for 24-hour predictions. Meanwhile, with heatmap data input, the performance results of the SimVP-gSTA model show an average SSIM and PSNR scores of 0.8 and 35.11 for 3-hour predictions and 0.695 and 20.89 for 24-hour predictions. In addition, the trajectory prediction results with these two data inputs are also evaluated. For a 3-hour forecast, the SimVP-gSTA model with infrared data input results in an error of 2.39 (240 km) whereas with heatmap data input it results in an error of 0.4 (40km). For the 24-hour forecast, results with infrared data input result in an error of 5.17 (500 km) whereas with heatmap data input it has an error of 3.58 (360 km). Moreover, this research also explores the estimation of tropical cyclone intensity, using an architecture called EfficientNet. Experiments are conducted on two tasks: multi-class prediction and binary prediction. Performance results of the model when predicting multi-class show an accuracy of 0.53 with the EfficientNetB3 model whereas in predicting binary class, accuracy reaches 0. 767. This study also explores the use of Stochastic Weight Averaging (SWA) technique on the EfficientNet model to increase prediction accuracy. Overall, the SWA technique can increase accuracy but not significantly. The best model in this intensity estimation experiment achieves an accuracy of 0. 7 68 in binary classification with the EfficientNetBl model version. Overall, this research shows that the deep learning approach, especially the SimVP-gSTA model, can be used to predict tropical cyclones. To predict the trajectory, using heatmap data is better than infrared. Meanwhile, for estimating tropical cyclone intensity, binary class prediction performs better than multi-class prediction. This demonstrates the great potential of using deep learning in mitigating the adverse effects of tropical cyclones by producing more accurate and faster predictions.