Supervised Machine Learning Techniques to Assess Tropical Cyclones in Bay of Bengal and Bangladesh

Bangladesh experiences extreme weather events such as heavy rainfall due to monsoon, tropical cyclones, and thunderstorms resulting in floods every year. Regular flood events significantly affect in agricultural industries and human lives for economic losses. One of the reasons for these weather phenomena to sustain is latent heat release from Bay of Bengal (BoB) and Southeast Tropical Indian Ocean (SETIO). As the country has limited observations from stations and oceans, modeling for numerical weather prediction (NWP) are challenging for local operations. For operational NWP, computational resources and time are also concerns for a developing country like Bangladesh. Besides, recent machine learning (ML) techniques are widely applied to study various meteorological events with efficient results. Therefore, this research aims to estimate predictability and accuracy of supervised ML for tropical cyclones by assessing air temperature at 2 meter (AT) and sea surface temperature (SST). For AT and SST, the study utilizes monthly data at 0.25 × 0.25^o horizontal resolution provided by the ECMWF reanalysis (ERA5). The gridded data is downscaled to area of interests such as coastal regions, BoB and SEITO with a study period of 40 years from 1979 to 2018. Furthermore, Bangladesh Meteorological Department (BMD) provides AT for 36 years from 1979 to 2015. The experiments segregate into two sections: (1) data normalizations via linear regression (LR) and multi-linear regression (MLR) and (2) supervised ML techniques applications in Matlab 2018b. The pre-processed data for LR show that AT from coastal regions such as Chittagong (CG), Barishal (BR), and Khulna (KL) divisions have stronger correlations (R) to SST in BOB with R = 0.910, 0.850, and 0.846 respectively than SEITO (R = 0.698, 0.675 and 0.678 respectively). Moreover, for these three regions, the correlation of MLR is 0.916 and 0.745 for BoB and SEITO with residual standard error (RSE) 1.312 and 1.218 respectively. For supervised ML applications, coarse decision tree (CDT) predict SST based on AT with train (80%) and test (20%) of the ERA5 data. Finally, the results from CDT model indicate that SST predictions are possible with 98.5% accuracy based on coastal stations. The trained CDT also validated model prediction utilizing observed AT (BMD observations) to forecast monthly SST and found 85% accuracy for monthly time series. In conclusions, CDT can predict SST from station data and assess if there is any possibility for tropical cyclone formation. The future works include further assessment for various categories of tropical cyclone and predict their intensity based on SSTs. This research aims to contribute in disaster mitigation by improving early warning systems. The possibility of cyclone formations will help for preparedness in saving property damages in Bangladesh.