Physics-augmented Deep Learning to Improve Tropical Cyclone Intensity and Size Estimation from Satellite Imagery

Abstract:

Current empirical or conventional statistical analysis approaches to tropical cyclone (TC) intensity and size estimation are limited. In this study, a physics-augmented deep learning model, called “DeepTCNet”, is developed to estimate TC intensity and wind radii from infrared (IR) imagery over the North Atlantic Ocean. While standard deep learning practices have achieved reliable estimates of TCs, informing this data-driven model some physic - in the form of physical information/relationship in this study - could greatly improve performance of the deep learning models. Three ways to augmenting the DeepTCNet by incorporating physical knowledge and/or physical relationships of TCs are proposed: (1) infusing the auxiliary physical information of TC into model; (2) introducing sequential IR images which appends a physical continuity of the intensity change of TC; (3) learning the auxiliary physical task. By augmented with auxiliary physical information of TC fullness, the intensity estimation performance of the model is improved by 19% than the no-augmented one. Tracking TCs from sequential IR images over 18 hours also can improve the intensity estimation by 12% than a single IR image. Jointly learning the auxiliary task of TC intensity, the skill of wind radii estimation of the model is enhanced by 12% than merely learning the tasks of wind radii, and is advanced by 20% than single task learning. The evaluation results showed that the DeepTCNet is in-line with SATCON, but outperforms ADT by 39% in TC maximum wind intensity estimates for TCs in all intensity categories. It also outperforms the subjective DT even though it was trained with best data that are mainly contributed with the subjective DT. DeepTCNet also surpasses MTCSWA in TC size estimates of TCs by 28% in average. As another demonstration of the capability of the DeepTCNet, the models trained with merely Atlantic Ocean data are also used to analyze both intensity and size of TCs in Western and Eastern North Pacific Oceans, showing that it has good performance in other basins.