Teknos

View Original

Staring Into the Eye of the Storm: Machine Learning for Automated Hurricane Eye Detection

Staring Into the Eye of the Storm: Machine Learning for Automated Hurricane Eye Detection

Josh Gong Thomas Jefferson High School for Science and Technology

This article was originally included in the 2020 print publication of the Teknos Science Journal.

Flooded streets, uprooted trees, and demolished homes greeted me when I turned on the TV. The raw destruction was shocking. In the wake of Hurricane Maria, I wondered why evacuation orders were not issued earlier. Turning to the Internet, I learned that the hurricane had surprised forecasters by undergoing a meteorological process known as rapid intensification (RI) [1]. In only fifteen hours, it had strengthened from a minimal category one storm to a deadly category five storm, leaving many Puerto Ricans, including a close family friend, with insufficient time to evacuate. Through an internship at National Oceanic and Atmospheric Administration (NOAA), I aim to find a solution to this problem using machine learning.

Every year, hurricanes are responsible for widespread loss of life and significant damage to property. As such, the accurate monitoring and prediction of hurricanes is crucial to reduce the socioeconomic impacts of these natural disasters and assist cities in preparing for landfall. Traditional forecasts make use of numerical weather prediction (NWP), computer models that incorporate multiple data sources and initial conditions to predict future states of the atmosphere. These forecasts have improved significantly since first introduced in the 1920s, but their downside is that they do not take into account high-level details, such as past events or visual patterns [8]. In addition, the lengthy data assimilation process involved in NWP leads to difficulty anticipating RI because intensity conditions can fluctuate dramatically as hurricanes near landfall, propagating inaccuracies in input data throughout the model. Moreover, RI is the result of many dynamic interactions between physical and environmental factors, such as wind shear and eyewall convection, that are difficult to quantitatively account for. While track forecasts have continually improved in accuracy over the years, intensity forecasts have lagged behind largely for these reasons [7].

One consistent indicator of RI is the formation of an eye in a storm [6]. Many theories exist as to why this occurs, but it is generally acknowledged that eye formation is necessary for hurricanes to achieve top wind speeds. Experienced forecasters can detect and analyze visual storm features (such as the eye) and make general judgments on a hurricane’s development. Nevertheless, this process is often subjective and time-consuming. To address this issue, I have spent the last several months at NOAA creating a machine learning model that automates the detection of a hurricane’s eye. 

A subset of artificial intelligence, machine learning is the use of computers to find patterns in data and make predictions without explicit programming. The ability to solve problems that conventional programs cannot, such as image classification, has resulted in an increased application of machine learning to many disciplines. Machine learning is also a viable solution to the analysis of the vast archives of aerial and satellite imagery, as models generally require training on large volumes of data for accurate results [4]. For example, researchers from Texas A&M University have created a neural network for analyzing drone and helicopter footage to improve disaster response [2]. By training computers to identify damaged infrastructure, first responders and other organizations can act with greater efficiency following natural disasters. In the realm of hurricane forecasting, satellite cloud images have been used to develop a predictive model for the direction of hurricane movement [3]. These models have been retroactively compared to forecasts, and they have shown similar or more accurate results. Recently, researchers have also successfully applied machine learning models to estimate the intensity of hurricanes, automating the manual Dvorak technique used by forecasters (Rogers et al., 2019). Predicting RI remains a problem, but researchers are optimistic. Professor Robert Rogers (personal communication, Jan. 24, 2020) of the NOAA Hurricane Research Division states, “All of this work holds the potential to improve six to twelve hour intensity forecasts, including hopefully RI. That's an area where machine learning techniques I think could be useful as well.” 

My research project at NOAA addresses this area, as I apply machine learning to satellite images with the goal of anticipating RI. Training on thousands of annotated temperature maps of past hurricanes, the model learns to recognize the onset of eye formation, which is often an indicator that RI will follow. To achieve this, I used a regional convolutional neural network, a model that creates bounding boxes and segmentation masks, which can be compared to the original annotations for accuracy. The first stage of the network identifies possible regions of interest, and the second stage refines those guesses to produce a final bounding box around the hurricane’s eye. 

My model trains on the NOAA-funded S4 at the University of Wisconsin-Madison, one of the fastest supercomputers in the world. Some of the more complex weather simulations run on S4 can take consumer laptops more than a year to complete; training my model took only 12 hours. I browsed through the output logs and examined the results before evaluating the model on new images it has never seen before. The final model runs in less than ten seconds per image and consistently achieves 88% accuracy or better on test data, efficiently and reliably automating the first step of what forecasters currently do subjectively.

By incorporating data and patterns from past storms, machine learning algorithms complement NWP’s quantitative nature with high-level visual cues. Though it is unlikely that machine learning models will replace human knowledge and experience in weather prediction, they serve as a powerful tool to improve the accuracy and efficiency of forecasts. As climate change continues to worsen, the strength and frequency of hurricanes are projected to increase [5]. Improved monitoring of hurricanes and detection of  RI has become more important than ever. Perhaps one day, my project may provide those facing an imminent hurricane the extra thirty minutes they need to safely evacuate.


References

[1] National Oceanic and Atmospheric Administration National Hurricane Center. (2019, February). Hurricane Maria (R. Pasch, A. Penny, & R. Berg, Authors). Retrieved from https://www.nhc.noaa.gov/data/tcr/AL152017_Maria.pdf

[2] Pi, Y., Nath, N., & Behzadan, A. (2020). Convolutional neural networks for object detection in aerial imagery for disaster response and recovery. Advanced Engineering Informatics, 43. https://doi.org/10.1016/j.aei.2019.101009

[3] Rajesh, K., Ramaswamy, V., Kannan, K., & Arunkumar, N. (2019). Satellite cloud image classification for cyclone prediction using dichotomous logistic regression based fuzzy hypergraph model. Future Generation Computer Systems, 98, 688-696. https://doi.org/10.1016/j.future.2018.12.042

[4] Rogers, R., Welden, C., Zawislak, J., & Zhang, J. (2019). Tropical cyclones and hurricanes: Observations. Reference Module in Earth Systems and Environmental Sciences. https://doi.org/10.1016/B978-0-12-409548-9.12065-2

[5] Varotsos, C., Krapivin, V., & Soldatov, V. (2019). Monitoring and forecasting of tropical cyclones: A new information-modeling tool to reduce the risk. International Journal of Disaster Risk Reduction, 36. https://doi.org/10.1016/j.ijdrr.2019.101088

[6] Velden, C. (2020). GOES-R series applications to hurricane monitoring. In S. Goodman, T. Schmit, J. Daniels, & R. Redmon (Eds.), The GOES-R series (pp. 95-102). https://doi.org/10.1016/B978-0-12-814327-8.00009-3

[7] Zhang, G., Perrie, W., Zhang, B., Yang, J., & He, Y. (2020). Monitoring of tropical cyclone structures in ten years of RADARSAT-2 SAR images. Remote Sensing of Environment, 236. https://doi.org/10.1016/j.rse.2019.111449

[8] Zhang, Y., Wistar, S., Li, J., Steinberg, M., & Wang, J. (2017). Storm detection by visual learning using satellite images. IEEE Transactions on Geoscience and Remote Sensing, 55(2), 1039-1052. https://doi.org/10.1109/tgrs.2016.2618929