Next-Generation Hurricane Intensity Estimation Using Enhanced TIMM Architectures

Early cyclone prediction is vital to prevent damage and loss of lives. Enabling governments and organizations to issue timely warnings allows the people to evacuate without risk and adopt necessary precautions. Accurate predictions help communities prepare effictively, reduce fatality, and limit damage to property, infrastructure, and crops. Advances like satellite imaging and predictive models through AI have greatly facilitated the prediction of cyclones early and with high accuracy. This study has sought to contribute to better intensity estimation of cyclones through using pretrained TIMM(Pytorch image Models) deep learning models accessed on GitHub towards strengthening early warning systems and decision-making. The model utilizes a dataset from Hugging Face, which includes 114,634 hurricane wind speed satellite images and wind speeds between 15 knots (1 knot = 1.852 km/h) and 300 knots. In this study, a few of the Pretrained TIMM models such as Resnet 18, Effientnet B1, InceptionNext, Regnety 008 and ResnetV2 50 were employed for the process of feature extraction, and for evaluting the model performance various performance metrics were calculated including RMSE(Root Mean Squared Error), R2 Score(Coefficient of determination), MAE(Mean Absolute Error), EVS(Explained Variance Score), RMSLE(Root Mean Squared Logarithmic Error), MAPE(Mean Absolute Percentage Error), and Median Absolute Error.In this paper the data is divided in two ways i.e..the data set is trained and tested on 80% train set and 20% test set, and another split is trained and tested on 90% train set and 10% test set. For 90-10 split data, the InceptionNext have better perfomance with the RMSE is 4.021, R2 Score is 0.985, MAE is 2.985, EVS is 0.973, MAPE is 8.067%, RMSLE is 0.075, Median Absolute Error is 2.134.This research identifies the possibility of transforming disaster management by integrating state-of-the-art AI-based predictive models, paving the way for more robust early warning systems and more informed decision-making processes. The findings have significant implications for enhancing global resilience to climate-related hazards.