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

Abstract A deep learning-based method augmented by prior knowledge of tropical cyclones (TCs), called DeepTCNet, is introduced to estimate TC intensity and wind radii from infrared (IR) imagery over the North Atlantic Ocean. While standard deep learning practices have many advantages over conventional analysis approaches and can produce reliable estimates of TCs, the data-driven models informed by machine-readable physical knowledge of TCs could achieve higher performance. To this end, two approaches are explored to develop the physics-augmented DeepTCNet: (i) infusing the auxiliary physical information of TCs into models for single-task learning; (ii) learning auxiliary physical tasks for multi-task learning. More specifically, augmented by auxiliary information of TC fullness (a measure of the radial decay of the TC wind field), the DeepTCNet yielded a 12% improvement in estimating TC intensity over the non-augmented one. By learning TC wind radii and auxiliary TC intensity task simultaneously, the model’s wind radii estimation skill is improved by 6% over only learning four wind radii tasks, and by 9% over separately learning a single wind radii task. The evaluation results showed that the DeepTCNet is in-line with the Satellite Consensus technique (SATCON) but systematically outperforms the Advanced Dvorak Technique (ADT) at all intensity scales with an averaged 39% enhancement in TC intensity estimation. The DeepTCNet also surpasses the Multi-platform Tropical Cyclone Surface Wind Analysis technique (MTCSWA) with an average improvement of 32% in wind radii estimation.