Long before there were “tech kids” by any definition, I was never more content than in the company of the “huge mainframe” computer residing at the local university computing center while still in high school. (I even crafted an elaborate plan to spend the night there—but that story is for another time.) Later on, during graduate school, I combined my interest in software engineering and hydrology and have remained keenly interested in all developments relating to those two fields ever since.
Many factors have impacted the latest explosion in hydrologic computing, but a primary one undoubtedly originates in the data boom. As is often noted, this is the age of “big data,” and hydrologic analyses can draw on plenty of sources, ranging from optical and microwave satellite remote-sensing data, LIDAR, ultrasonic sensors, synthetic aperture radar (SAR) to drone footage and Twitter data, among others.
Machine learning (ML) is a computational process that has been around for decades but has resurfaced of late to leverage big data in the context of equally big global issues such as flood-risk forecasting. ML applications in hydrology are constantly evolving, and long-term success will depend on the quality of the working parts on which they rely. The availability of high-quality, relevant data and the transparency and validation of results are some issues that come to mind. Now and for the foreseeable future, this tool is becoming increasingly more significant to protect against extreme flood risks.
ML on the Job
In extremely simplistic terms, an ML-based model trains an algorithm by iteratively processing a vast number of data points to derive functional relationships among parameters. An ML model trained on a particular dataset brings its learned recognition of key relationships to a test dataset; its ability to detect patterns and nonlinearities as well as generalize conclusions based on new data points are key, giving the model its predictive power. Let’s simplify how this is done: Google Image Search Engine can recognize a cat in a photo by the training that went into its ML image-recognition model by showing the ML thousands of cat pictures.
ML as a flood forecasting tool has taken off, and cities across the globe are using this technology to process storm data and gauge possible flood risk. Google launched its AI-based forecasting technologies a few years ago and now claims to be able to accurately predict flood risk days in advance of a storm for more than 80 countries as well as deliver updates in real time through its Flood Hub platform. It presumably achieved this by first running pilot projects in India, where it used a variety of site-specific data to train the model, later combining results with basic hydrologic flood-simulation models and Long Short-Term Memory networks (LSTMs). Interestingly, an initial focus just on AI tools evolved to integrate output from basic flood-simulation models, ultimately producing results more in line with realistic flood-inundation patterns.
Hybrid systems combine two different modeling approaches: the process-based models grounded in the laws of physics (USACE’s HEC-HMS stormwater runoff and HEC-RAS flood-modeling software, for example)—which have driven our hydrologic modeling to date—and the statistical approach, represented by data-centric ML techniques.
MIT researchers (bit.ly/MIT-GAN) recently used a type of ML model based on a generative adversarial neural network (GAN) to generate satellite images of a region in Houston before and after a flood event. One network was trained on real satellite images before and after a hurricane. The other network was trained to identify real imagery from synthetic imagery. The feedback between the two networks acted to improve the predictive competencies of each.
Using this AI technology alone generated realistic flood images, but the model still hallucinated, projecting flooding in areas impossible to flood. Interestingly, when combined with a physics-based model that integrated dynamics associated with storm surge and other flood- and climate-specific parameters, the model’s results improved in accuracy. ML applications in hydrology seem increasingly likely to evolve into hybrid systems. Integrating two different methodologies may ultimately enhance the overall reliability and predictive power of the system.
A Partner in Flood Management
Notwithstanding a lifelong fixation with tech, I’m painstakingly cautious when considering model outputs. ML applications have significant potential, but also are experimental, prone to error and highly dependent on countless unknown variables. Model evolution is an iterative process, as it should be, and these early systems, each developed and configured differently, are the pioneers for systems likely to grow more common—and necessary—with time. In a field characterized by complex feedback signals, significant nonlinearities and uncertainty, it’s clear that a lot could go wrong.
One fundamental question an experienced professional always needs to ask: Does it make sense? Thinking critically about tech-generated solutions is paramount, and given the ubiquity and “flashiness” of AI right now, it’s easy to defer all judgment to the machine. (Think of all the Tesla Autopilot car crashes that have occurred where drivers assumed the car knew what it was doing.) Input of professionals is important and the more interdisciplinary the better. A critical, experienced and trained eye yields a technically stronger tool in the end.
The post Thoughts From Engineers: Machine Learning and Gauging Flood Risk first appeared on Informed Infrastructure.