Estimating tropical cyclone intensity using dynamic balance convolutional neural network from satellite imagery

Abstract

Accurate estimation of tropical cyclone (TC) intensity helps to understand the evolution of TCs throughout their life cycle and plays an essential role in mitigating TC impact. Although TC intensity estimation methods based on deep learning have made significant progress, the existing techniques do not apply good methods to overcome the intensity overestimation and underestimation problems caused by the unbalanced distribution of TC data. Therefore, we propose a dynamic balance convolutional neural network to overcome these issues. The model consists of two branches, one branch is the learning of the raw data, and the other is the learning of strong (weak) TCs that account for a few data samples. Finally, the model is dynamically adjusted by adaptive trade-off parameters, gradually from the learning of the raw data to the learning of strong (weak) TCs, thus reducing errors in underestimation (overestimation) of strong (weak) TCs. Furthermore, an attention mechanism is employed to obtain the correlation between channels to improve the accuracy of TC intensity estimation further. We used globally 1285 TC cases from 2003–2016 to train the model and globally 94 TC cases from 2017 as independent test data. The results showed that the root-mean-square error of TC intensity estimation was 8.32 kt, 35% lower than that of the advanced Dvorak technique and 26% lower than that of the deep learning method visual geometry group (VGG). For a subset of 482 samples (from East Pacific and Atlantic) analyzed with reconnaissance observations, a root-mean-square intensity difference of 7.95 kt is achieved. Finally, we explored the model’s feature learning process and the contribution of each component of the satellite image to the TC intensity estimation through model visualization.