A Meteorological Challenge in the Comoros
It was around April of last year when Stefan Ligtenberg told me: “The project got funded, now we need to do what we promised!” It was good news, but also not unusual. As consultants in meteorology, our work at Weather Impact can often be summarised as implementing the ideas we wrote in donor funded requests for proposals. This time, however, what we had pledged was meteorologically ambitious, making me feel excited and slightly concerned in equal measure.
As part of a consortium led by Haskoning, we were asked to map rainfall and wind speed extremes, both for current climate conditions and for the SSP3-7.0 future climate scenario in 2050 and 2080. The exciting part was the target region: the Comoros Islands, a tiny archipelago in the southwest Indian Ocean, northwest of Madagascar, made up of three main islands (Grande Comore, the largest, Anjouan, and Mohéli). The largest of them is about 60 km long and 50 km wide (roughly the size of a city like Los Angeles). The ambitious part was that we were asked to produce maps at 1×1 km resolution, something highly advanced in meteorology, especially over small mountainous islands where ground observations are extremely scarce.
For context: the most advanced operational weather model covering that region has a resolution of 9×9 km, but it cannot be used for multi-year climate analysis. For historical climate, the ERA5 reanalysis at about 31×31 km is considered the gold standard, while the highest-resolution future climate data sits at around 44×44 km, though 100×100 km is far more common.
This project became one of the most rewarding one, since I started working at Weather Impact. It has been a perfect opportunity to learn new expertise with an inspiring goal, producing results which can concretely help decision-makers prioritize resources. This is why I decided to share this research in this article (~10 minutes of reading time).
A Paradise with Sharp Edges
Close your eyes and imagine a tropical island: lush vegetation, a dominant volcano rising from the sea to over 2300 meters, smaller islands ringed by reefs, tropical forests and steep terrain. Unfortunately, there was no fieldwork for me in this project, so I had to imagine it just as you are doing now. But my luckier colleagues who did visit it described not just the beauty of the islands, but the meteorological realities of living there.
Being remote islands in the middle of an ocean, the Comoros are hit by the full force of the elements. Coastal areas are particularly exposed to storm surges driven by extreme winds. The islands also sit along the path of some of the most powerful storms on the planet: in 2019, Grande Comore was struck by Cyclone Kenneth, with wind speeds reaching 240 km/h in the open ocean. On the main island, the 2300 meters high volcano amplifies intense rainfall, which then flows quickly toward the inhabited areas at its base, often in unpredictable paths. The volcano is geologically young, and water has not yet had time to carve deep, stable river valleys down its flanks.
The urgent need for detailed and reliable meteorological hazards information became clear and paved the way for the Comoros Post-Kenneth Recovery and Resilience Project, funded by the The World Bank Group , of which this research is part of.
Resolution is Everything
To map extreme weather events over the three islands at a resolution of 1 km, we first had to overcome a significant resolution gap. Going from grid cells of 31 km per side to just 1 km doesn’t mean finding 31 times more detailed information. It means finding 961 times more (31×31). Think of it as going from a single pixel to a 31×31 pixels image: in the first, you have one pale red square; in the second, your brain has enough details to recognise a bright red letter “A” with a shadow and a white background.
This information gap matters even more in meteorology, because weather and climate models only report one value per grid cell. This means that if the 31×31 km cell has a value for hourly rain of 5 mm, the actual amount of rainfall ‘produced’ by the model in 1 hour equals a thin layer of water, 5 mm deep, over the entire square of 31 km per side. That’s about 4.8 billion liters, a lot of water! Of course, in reality it is almost never the case that rain is spread evenly over such a large area. Often, especially at the tropics, the thunderstorms can discharge a huge amount of water in a very localized area.
Taking the example of before, imagine that you can now zoom within the 31x31km square, and distinguish cells of 1km per side. You might find a thunderstorm affecting only 10% of all the surface. Doing some math, it turns out that if 90% of the area experienced no rain, the remaining 10% of it experienced 50mm in 1 hour. For those unfamiliar with rainfall measurements, 50 mm/h is an intense downpour, likely causing flooded underpasses and flash floods in steep terrain with poor drainage. Figure 2 shows a real example of the rainfall downscaling during the passage of cyclone Kenneth.
“Mapping extreme events at high resolution is therefore essential to understand when and where weather conditions can lead to potentially dangerous situation, where investments in water management or transport infrastructure need to be prioritised, and what areas have the highest exposure to such meteorological risks.”
The figure above shows, on the left, the coarse input data with resolution of 31x31km, in which the island is represented by 4 pixels. On the right, the data are downscaled by Weather Impact at resolution of 1x1km (masked on land only). Grande Comore is now represented by approximately 1000 pixels.
All Sorts of Downscaling
So, how do you go from 31km to 1km of resolution? With something called downscaling. In meteorology, there are many different techniques, but one distinction is especially important: dynamic versus statistical downscaling.
- In dynamic downscaling, you run a high-resolution physical climate or weather model over a smaller area, using the coarse-resolution model as input at the boundaries. The model explicitly solves the equations of atmospheric motion, moisture, and energy, allowing mountains, coastlines, and storms to physically shape weather and rainfall at finer scales. It is physically consistent and highly detailed, but also computationally expensive. In a project like this one, dynamic downscaling requires simulating hundreds of variables (e.g. temperature, humidity, wind, pressure, clouds) at every level of the atmosphere and every point of the ocean, even though we only needed two of them (surface wind and rainfall) over land.
- In statistical downscaling, instead of running a new physical model, you use statistical relationships learned from observations (or from high-resolution simulations) to translate large-scale climate information into local-scale conditions. It is much cheaper and faster to apply, but it relies on assumptions that past statistical relationships between large-scale patterns and local weather remain valid in the future.
Dynamic Downscaling
In this project, we did both. We began with dynamic downscaling of selected events where we expected extreme weather conditions to happen. We used the records of tropical cyclone passages, we analysed historical weather data and the observations of the few weather stations present on the islands. This allowed us to simulate past weather conditions experienced by the Comoros in fine details. Watching a cyclone taking shape in the simulation was honestly terrific (and terrifying to see it hit the island).
But playing with the weather, as mentioned, is incredible computationally expensive. One of the reasons is that you cannot feed a weather model data at 31 km resolution and expect it to generate 1 km output in one step: you need to run three nested simulations, at 9 km, 3 km, and finally 1 km, each feeding into the next (Figure 3). Using dynamical downscaling, we simulated 70 days of weather conditions over the Comoros, totaling 875 hours (or 36 days) of simulation time and producing 1080 GBof data.
Statistical Downscaling
Statistical downscaling, less fun but much more efficient, was introduced to simulate tens of years of weather conditions at 1km resolution. For this, we used a supervised learning technique called Random Forest: a powerful algorithm for recognising patterns in large datasets. In simple terms, Random Forest learns the relationship between a set of predictors and a target variable, and can then make predictions when presented with new, unseen predictor data.
As an analogy: imagine providing it with two days of observed weather (predictor) and the weather on the third day (target). Repeat this by training it with a large number of such predictor-target examples. When you ask Random Forest to predict today’s weather from the past two days, it will match patterns across the many examples given, and it will likely provide you with a quite reasonable result. The reliability of the results depends on many factors, such as the amount and quality of the given predictor-target examples, and by the nature of the weather situation to predict. We applied a similar logic to both rainfall and wind extremes.
Starting with the wind data, we trained a different Random Forest for each 1x1km pixel, using as predictors the wind data at coarser resolution, alongside variables such as elevation, terrain exposure, and time of day. The target was the dynamically downscaled wind output from the WRF model. Once the wind data was downscaled (Figure 4 is an example), we repeated the process for rainfall, in this case training one Random Forest algorithm per island. In this way, a computationally inexpensive algorithm was trained to reproduce the results of a far more expensive physical simulation. The mean error was approximately 5 km/h for wind speed and 1.9 mm for rainfall, well within acceptable bounds given that the extremes we were studying are one or two orders of magnitude larger.
On top of the traditional validation method for statistical downscaling, what looks promising of this previous map is the high spatial coherence of the downscaled with field. A different Random Forest model was trained for each pixel of the 1x1km grid, however the overall picture is high coherence as if there was one single model applied to the entire domain.
The Results
The downscaling exercise gave us a very detailed description of wind and rainfall conditions across the islands. The weather observed in a specific location, at a specific time, is the realization of only one of the possible weather trajectories stemming from similar initial conditions. When downscaling with weather models, we let the weather evolve artificially to one specific realization of the possible ones. It is therefore not advisable to investigate individual extreme events using downscaled data, especially if that event has not been downscaled dynamically, but only statistically.
Return Times
To characterise extremes, we used instead the statistical concept of return times, the average interval you would expect between recurrences of a given event. A 100-year flood, for instance, has a return time of 100 years, meaning there is roughly a 1% chance it occurs in any given year (not that it happens exactly once per century). With approximately 40 years of downscaled data at our disposal, we used statistical distributions, specifically the Pearson Type III, to extrapolate beyond the observational period and estimate the intensity of events with return times up to 100 years. In this way, we can infer what are rainfall amounts or wind speeds corresponding to a return time longer than the amount of years we have data for.
Figure 5 shows the intensity of an extreme rainfall event associated to 100-year return time, under current climate conditions. Each map shows high heterogeneity across the islands, with locations near the summits of both Grande Comore and Anjouan receiving almost 1000 mm in 24h, while other locations along the coast experiencing ‘only’ up to 400 mm (already an exceptional daily amount by any standard).
In my view, one of the most valuable aspects of these maps is not the absolute numbers themselves, but what they reveal about the relative spatial distribution of risk. Above a certain threshold, the meteorological hazard becomes so severe that the difference between 700 mm and 1,000 mm in a day matters less than knowing wherethe event is most likely to strike.
“No country, and certainly no small island state, can implement equal adaptation measures everywhere. Knowing which locations face the greatest exposure allows decision-makers to prioritise resources accordingly.”
In this study, hazard is only one piece of the puzzle. Risk is ultimately determined by combining hazard with exposure and vulnerability, work that Haskoning will carry forward in the next phase of the project.
A Personal Note
Of my two and a half years at Weather Impact, this project has been the most rewarding on a personal level. It drew on virtually everything I know, from coding and running weather simulations, to processing climate model outputs, analyse and visualize climate data, building supervised learning routines, and managing relationships with project partners. I did not do this alone, and I want here to thank Stefan for the many updates meeting in which his critical eye as a former climate scientist made this a better piece of research.
Climate Consultancy
This experience also taught me what it means to do technical work in the context of climate adaptation consultancy. At university, at least the two I attended during my education, climate research was presented as an exclusively academic activity. Consultancy was never really part of the conversation. Looking back, the work we did here could easily form the basis of a master’s thesis or a PhD project. In academia, achieving similar results might take several years and a much larger budget. We did it in roughly nine months, with one person working approximately one day a week.
I don’t say this to diminish academic research, but quite the opposite. Both paths contribute meaningfully to society, and rigorous, slower-paced research lays the fundamentals of much of what applied work like ours depends on. What I do wish is that I had known earlier that there was a way to put a scientific education to work in this way: on real problems, with concrete constraints, and with results that feed directly into decision making. If you are a student or early-career scientist wondering whether consultancy is “serious” work, I hope this project is a small answer to that question.
