From Vern's Vision to Computer Vision: The Evolution of Tropical Cyclone Intensity Estimation from Satellite Data

The contribution of Vern Dvorak’s technique for estimation of Tropical Cyclone (TC) intensities cannot be overstated. From the BAMS article chronicling the method Velden et al. 2006 stated, “The Dvorak Techniques practical appeal and demonstrated skill in the face of dynamic complexity place it among the great meteorological innovations of our time. It is difficult to think of any other meteorological technique that has withstood the test of time and had the same life-saving impact.” The Dvorak Technique’s (DT) enduring usefulness stems from both the concept’s insightful premise, that organizing convective features are related to intensity, and the method’s ability to teach analysts to discern complex cloud patterns and morphology over a broad range of TC intensities. Minor changes have been implemented over time to account for rapid intensity changes, address some biases and improve the DT’s performance in various ocean basins. Nearly 50 years later, the method remains a primary tool used by TC warning agencies throughout the world.

The DT method was created in the 1970s and 1980s, when operational weather satellites were limited to visible and IR channels viewing clouds and profiling the atmosphere in clear regions. Analysts can only see the uppermost cloud top, not critical mid and lower-level storm structure, and this often leads to uncertain center locations and intensity estimates which are dependent on accurate storm position. Low earth orbiter passive microwave (PWM) imagers and sounders emerged in the 1980s and opened the door to utilize the best of both worlds: the frequent GEO IR/VIS scans and the “into the cloud” TC structure from PMW. In addition, active microwave sensors (scatterometers and synthetic aperture radar) came online in the 1990s and 2000s that enabled the analysts to see through rain to the ocean surface and depict ocean surface wind vectors providing horizontal surface structure.

More recently, automated algorithms such as the Advanced Dvorak Technique and Digital Dvorak have been developed in an attempt to ameliorate some of the inherent analyst subjectivity in the DT method. These automated techniques benefit from much improved spatiotemporal resolution of the global geostationary satellite sensors, and the fact these estimates take only minutes to run computationally and can produce intensity estimates at much greater frequency. They can be applied by anyone who has access to the data and do not require many months of training to produce skillful estimates the way the DT does. The ADT has matured to operational status and now sits alongside the DT as a tool routinely used by global warning agencies.

In the present age we see a transformation underway within the satellite community to harness the power of deep learning tools to analyze satellite imagery and extract key parameters such as TC intensity. Deep-learning based tools now have the ability to “see” and analyze TCs in ways that Dvorak did, and more. The techniques are being applied to a wide spectrum of satellite frequencies including infrared, visible and microwave and are leading to improvements in satellite-based TC intensity estimates. Despite the adoption of these state-of-the-art methods Vern’s technique developed in the 1970s remains in use today. This presentation will explore the path from Vern’s ubiquitous DT to the modern day application of deep learning methods to satellite imagery and how they might shed light what it was that Vern saw in those thousands of images he analyzed with his talented vision and what the future may hold for his method.