A Physics-guided NN-based Approach for Tropical Cyclone Intensity Estimation

Abstract

Abstract In this paper, a regression neural network pGCNN-TC is designed for estimating tropical cyclone (TC) intensity using satellite images, which is an essential step for TC forecasting. The proposed model employs features derived from convective cores near TC centers, group-equivariant convolution layers that learn rotation equivariant representations from data, and an adjusted MSE loss function to alleviate the underestimation tendency on high intensity TCs. We evaluate our proposed model on a benchmark dataset TCIR, where TCs during 20152016 are used as testing cases. Our proposed model outperforms all other start-of-the-art models including NN-based models and objective operational techniques in overall performance. Case studies also suggest that convective core features are especially helpful for high intensity TCs when eye structures are not clear in satellite images. * This work was supported by the National Key Research and Development Program of China (2022YFC3004102) and Qinghai Kunlun Talents Program. Ying Zhao (yingz@tsingua.edu.cn) is the corresponding author.

Publication
Proceedings of the 2023 SIAM International Conference on Data Mining (SDM)
Wenguang Chen
Wenguang Chen
Professor
(教授)