FEATURE EXTRACTION AND TRACKING SYSTEM FOR TROPICAL CYCLONES

Abstract

There are numerous sub-continents in the world where cyclones yearly hit a certain region. Since cyclones directly affect people's lives and homes, their prediction is crucial to preventing the loss of life and property. There are many ways of techniques that is able to forecast tropical storms systems such as Dvorak technique, different kinds of time series analysis, Artificial Neural Network (ANN), numerical weather prediction system (NWP) model, machine learning, etc. Machine learning theory such as regression analysis is still challenging for forecasting tropical cyclone’s track. It is very useful and suitable for predicting and great impact on independence and random data for time series. Tropical cyclones that occurred in the Northern Indian Ocean affected Myanmar Land. Historical datasets are obtained from Joint Typhoon Warning Center (JTWC) and provided from 1945 to 2022 years. Feature extraction has a critical role in machine learning theory and also strong features impact the outcome of the cyclone trajectory. The movement of the cyclone trajectory points out the value of Latitude and Longitude. In this research, these values are changing in the direction and magnitude of the movement. The main contribution is stand on the correlation coefficient value of the direction and magnitude of the historical trajectory data and test data. Not only the latitude and longitude of the cyclone but also metrological data such as wind speed and sea level pressure are also used the input data to extract the features. Features of Direction and Movement are extracted to build the model based on similar cyclones and tested one. Logistic regression method is used to forecast the latitude and longitude of a cyclone's location 24 hours ahead of time by using the last twelve hours of observations (two positions, at six hourly intervals, and the current position). The threshold value is also an essential decision-maker or forecaster of the system. According to the value, the accuracy of the system can also change. Three threshold values of the sigmoid function are tested which is based on two similar and three similar cyclones are tested. For the evaluation of the system, three matrixes are selected such as mean absolute percentage error (MAPE), mean absolute error (MAE), and root mean squared error (RMSE). By adding a maximum wind speed and minimum sea level pressure from the historical dataset, performance evaluation is gradually improved for these regression methods.