Flooding & climate risk: December Digest
Festive greetings!
As we frantically usher 2021 out of the door, in lieu of any Christmas holidays, I'll instead take you on an?ego?trip?as we look back on the research of Fathom & friends in 2021. It was another productive year, as we continue to rival the research output of most university departments.
The eagle-eyed amongst you will note the lack of a November Digest. As the sands of time slips through my fingers with a laxity akin to the spread of the omicron variant, perhaps 2022 will see an expectation-setting rebrand of this as a Quarterly Digest (I can already feel 2026's Biennial Digest approaching!). Either way, I promise I'll look at other peoples' work in that issue, but, for now, indulge me.
Enjoy!
It's only traditional to kick things off with a meme:
Model validation
2021 began with a bang as we published the science behind our new version of the US flood model in?Water Resources Research.?We ticked off a bunch of 'firsts' in our field in terms of model scale, scope, fidelity, and function; accounting for all major flood perils as well as providing a climate-changed view. As always, we present an extensive model validation section, where we compare the output to a bunch of different data sources and conclude the model can skilfully replicate local engineering-grade models.
Not enough model validation for you? Fine,?have some more. This time in EGU's?Natural Hazards and Earth System Sciences, we admit that model-to-model comparisons as evidence for one model's validity is arguably circumstantial. Models of different scales can still share the same biases, so the fact that they might look alike might not tell you a whole lot. Here, then, we try and simulate?real flood events, rather than the more esoteric?maps of made-up floods representing the same probability of flooding everywhere at once.
This comes with its own challenges. Sure, comparing to observations sounds preferable. But, in many cases, those observations are just as uncertain as the models! This means you have a pretty low ceiling for model interrogation: you cannot conclusively discriminate between competing models because the benchmark observations themselves are too uncertain to do so.
We found, then, that our US model can replicate field observations within error, but that perhaps is not saying very much.
(this gives me a great opportunity to re-visit a favourite meme of mine earlier in the year)
Model application
Self-deprecating memes aside, with some evidence that the model has?something?useful to say about US flood hazard, we then took a dive into a whole bunch of interesting applications.
In?Journal of Environmental Economics and Management, with friends Jake Bradt (Harvard) and Carolyn Kousky (Wharton), we looked at what motivates people to?voluntarily?obtain flood insurance in the US (it's mandatory in the FEMA 100-year Special Flood Hazard Area). We found that insurance uptake is greater in areas where we say there is flood risk, but FEMA do not. i.e., insurance is not mandatory, but folks are getting it anyway because they know there is risk there. Although not the central conclusion of the paper, it's a further neat bit of model validation, evidencing an alignment of local flood risk knowledge with the output of our model.
(I'm hoping I got enough credit for the self-critical memes to permit a self-indulgent one)
In?Natural Hazards, with friends at the University of Iowa, we examined the intersection of our modelled flood exposure with pockets of social vulnerability in the US.?We find?hotspots of this unjust collision in rural areas, across the US south, for mobile home communities, and for racial minorities.
We also worked with Mike Wehner (Lawrence Berkeley National Laboratory) in?Climatic Change?to understand the attributable fingerprints of climate change in Hurricane Harvey's flood hazard. While we're all pretty ok with more hot = more wet in the atmosphere, what does that mean for flood impacts on the ground? They find big nonlinearities in how more rainfall translates to flood hazard: best estimates suggest a 19% increase in Harvey precipitation due to climate change incurred a 14% (~$13bn) increase in flood losses.
领英推荐
Don't be this person (yes, I did a similar one earlier in the year: sue me)
We contributed to?an interesting review paper?in?Environmental Research: Infrastructure and Sustainability?led by Matt Brand (UC Irvine) on an emerging financial tool to deal with these environmental problems. These challenges are characterised by being mentally expensive and deeply uncertain, but also in addressing them you save?money?in the long run (let alone all the stuff that nobody cares about like clean air, clean water, resilient communities). The trouble lies in getting the capital together in the first place to invest in the solutions. That's where Environmental Impact Bonds come in: investors stump up some initial cash, and they get paid back based on the cost savings. Importantly, (as is the theme with lots of?Digests) uncertainty is treated as a tangible risk; and with risk comes reward. Investors get rewarded for investments in riskier/uncertainier projects with higher interest rates.?Our?Nature Sustainability?paper with The Nature Conservancy?in 2019 gives a flavour of this:
This would be an ideal EIB: get some capital together, solve problems, reward investors based on uncertain model results. Cool, huh?
Channel solver
One of the seminal technical developments of the year was?our new channel solver?in?Water Resources Research! Who'd've thunk that rivers are important to flood modellers? They convey the bulk of flood flows, and nobody has a clue how big they are globally! That's a problem!
We tie a channel size estimate to the particular discharge frequency we expect a river to convey when bankfull. That frequency is something like the discharge which is exceeded every other year. The neat thing about this is that the discharge value used is drawn from our extreme flow estimator, meaning any bias in this is dampened somewhat by the size of the channel. i.e., if we overestimate a given river flow, the channel would larger to account for this.
We devised a gradually varied flow solver to ensure that, when implemented in a hydrodynamic flood inundation model, the channels in the model flow out of bank at the correct frequency. It sounds nerdy and in-the-weeds, but it's a massive deal – particularly for small, frequent floods where a greater proportion of flood volumes remain in-channel. These events are the ones which, if modelled poorly, really drive annual modelled losses (by virtue of their frequency).
The method was applied in the aforementioned US flood model papers, and was further examined in?a?Hydrological Processes?paper?with the University of Concordia, where it was used in extremely high-accuracy models in Quebec.
Extreme flow estimation
With friends at the University of Bristol, we made big headway on methods for actually determining what the extreme flows you actually put into these channels should be for design flood mapping.
There are, generally speaking, two approaches for doing this a) hydrological modelling, and b) regionalised flood frequency analysis (RFFA). a) involves physically (most of the time) modelling the river flows that arise from rainfall and other stuff going on the atmosphere. b) involves working out what's going on in rivers that don't have gauges by building a statistical relationship between the flows where you do have gauges and a bunch of characteristics of those catchments, then transferring that relationship to the ungauged case.
Work led by Gang Zhao at UoB in?Hydrology and Earth System Sciences?involved building a new RFFA (b) using a bunch of recently available global data to predict gauged flows with machine learning techniques. We see considerable improvement in extreme flow estimation compared to?our previous RFFA methods, with low bias and prediction accuracy within typical?measurement?errors of gauge-based discharge. It's a hugely important development, that will now underpin all of our future fluvial models.
With Laura Devitt at UoB in?Environmental Research Letters, we see the new RFFA again outperformed the old version for extreme flows in the US, as well as all other global hydrological models (a). Obviously, the latter (a) have more functionality when it comes to examining river responses to different weather conditions (i.e., for climate change) and for understanding spatial correlation, but for the purpose of design flood mapping under baseline/historical (whatever that means) conditions, RFFA (b) appears to lead the way.
This work now creates a meaningful discriminant in evaluating large-scale flood models based on their boundary condition generation method.
I daren't test your patience any further. It's been a fun year! Plenty more to come in 2022, with some big papers on global DEMs, US climate risk, UK flood modelling, global stochastic, and more in the pipeline.
See you in January!