A Multi-Features-Based Model for Tropical Cyclone Intensity Estimation with Infrared Remote Sensing Observations

In this study, a Tropical Cyclone (TC) Intensity Estimation (TIE) was proposed by comprehensively utilizing the multi-features of infrared (IR) remote sensing images, namely MFTIE model. The new model comprises of three modules: a CNN module based on Residual Blocks to extract the graphical features, an optimized multilayer perceptron (MLP) module with residual structure to process the statistical features, and a fully connected output module to give the final estimation of TC intensity. With the HURSAT-B1 dataset a root-mean-square error of 8.91 kt on test set. Besides, the MF-TIE model was compared with a model that uses only images as inputs, and the results showed that the test accuracy of the proposed model improved by 7.67 %. Furthermore, the comparisons with existing models confirm the performance superiority of the proposed model. The findings suggest that utilizing multi-source features of remote sensing data, such as the dynamic and radiative characteristics of IR images, has great potential in the field of remote sensing inversion.