Reducing uncertainty on wildfires

On a recent article published on the 88th bulletin of the portuguese National Authority for Civil Protection (in portuguese only, accessible here), I wrote that there is a map capable of reducing uncertainty on land management and, after a fashion, also on suppression efforts, in that analyzing prediction curves for wildfire susceptibility maps one realizes that in mainland Portugal 70% of all burnt areas are contained in just 30% of susceptible territory, differently put, 70% of all burnt areas are contained in just 30% of those areas that can actually be affected by fire.

I consider susceptible all those territories that can be affected by a wildfire, leaving out urban areas and inland waters (rivers, lakes, …) which explains why the map I present below has blanks. Those blanks are areas where it makes no sense to talk about wildfire susceptibility in that they do not have fuel to support a wildfire.


On a red map, you’ll get it right everytime

Risk management is about reducing loss. On a simple note, a risk is a consequence – usually negative – on an exposed element, in the presence of a certain hazard in that that element at risk is contained in a perilous area. Regarding wildfires, exposed elements or elements at risk are such as fuel itself, either forest stands or cultivated land or shrubs, of variable value depending on wether those are economically viable areas, service and infrastructure (electricity, communications, water, …), dispersed buildings and… people.

Reducing uncertainty, managing risk, is creating ways of minimizing losses. It means less economical loss to those living off rural produce, and greater safety to those living, working or standing at rural areas. It is easily understood how reducing risk is useful, and that is why spatial modelling is made and models are run to help understand how the phenomenon at hand relates to the territory that supports it. I have done it before on my MSc thesis, I published a paper on the matter, and more recently, on my PhD dissertation, after multiple model runs, I demonstrated how different wildfire susceptibility models behave, allowing me to safely state what I did opening this post.

Mainland Portugal has roughly 8.9 million hectares of which 8.5 million hectares are wildfire susceptible, and if spatial modelling isn’t done, wildfire risk management has to be equally done to those 8.5 million hectares, assuming they can be equally affected by wildfires. That is just not the case.

If you paint your map in red, all of it (the colour most used to denote the most dangerous areas), it is easy to be right all the time. You’ll always nail it! Wherever it burns, it’s red, you’re just too good. On the other hand, if there is no map, you might feel tempted to say it just happened, it was a matter of chance, nobody knew. But losses have to be cut, people’s safety has to be assured. Risk has to be managed!

One of the modelling blocks I studied the most, for the period  1975 – 1994, using 1995 onward for independent validation, as mentioned before, 70% of all burnt areas fit in just 30% of susceptible territory, which is equivalent to about 2.5 million hectares. Those are still a lot of hectares, but 2,500,000 hectares are, nonetheless, significantly less than the 8,500,000 hectares that in mainland Portugal can sustain wildfires. And these are somewhat conservative numbers. If the modelling block is different, such as 1975 – 2012, only 18% of all susceptible areas are needed to fit 70% of burnt areas, but let’s stay with the most conservative numbers for now.

Defining priorities, protecting value, safeguarding people

The map shown below is the wildfire susceptibility map for the modelling block 1975 – 1994 in mainland Portugal. It is the same map as published on the National Authority for Civil Protection’s 88th bulletin, and whose methodology principle is the same as the map that shows in the annual Directive for wildfire suppression. In itself, it is quite a useful map; in a universe of 8.5 million susceptible hectares, it creates five classes of susceptibility, informing on propensity and severity. As it doesn’t have a scenario and probability it cannot be called a hazard map, and since it also doesn’t have information on vulnerability and value, it definitely cannot be called a risk map. It is a susceptibility map for 8,500,000 hectares. Period. Just not a full stop! It does help define priorities. Let’s take a look at the prediction curve, right after the map.

The prediction curve is one of the instruments that allows us to assess map’s quality. The area under the curve (AUC) for this prediction curve (in blue) is 77%. The higher the AUC, the better the model. Besides, this curve also allows us to know how much susceptible area (x-axis) is needed to fit a certain amount of burnt area (y-axis). In this case, the modelling interval 1975-1994 needs just 30% of susceptible land to fit 70% of burnt areas, and 30% of susceptible land translates into just 2.5 million hectares. What else has been burnt was dispersed beyond those 2.5 million hectares, but the bulk of it, the most demanding wildfires, most likely happened on those 30% of the territory. A smaller area than can be easily identified.

We look at the map above and still it looks as too big an area, but I assure you that what’s painted in black are those 30% of the susceptible land, those 2.5 million hectares, not the total 8.5 million hectares that mainland Portugal has to offer wildfires. Areas painted in black are good candidates for priority action, for landscape management, for fuel breaks and mosaics, for added surveillance or suppression effort concentration, knowing that 7 out of 10 wildfires will most likely occur in those areas and start not deep within those patches of land, but probably on their interfaces.


Integrating knowledge

I wrote, above, my numbers are conservative. Mainland Portugal has one of the best – if not the best – databases on wildfires, both on tabular data and on cartographic coverage. This richness allows researchers like myself to run a myriad of models with different criteria and modelling intervals. I have observed that a wildfire susceptibility model with the modelling block of 1975 – 1994 is a solid model with excellent results, but it is equally possible to model with a whole series or different combinations of 10, 20 or 30 years. The models I studied for mainland Portugal are stable and I am not expecting significant changes on model behaviour. Integrating this kind of knowledge on organizations is a must for defining priorities and knowing where to start.

Overlaying what I have called priority areas (the 2.5 Mha) on top the wildfire susceptibility map, I find no surprises. Priority areas mostly overlap the two highest susceptible classes, high and very high. Many of the medium (yellow) areas are left off, just as most the green areas. This does not mean, mainly in the medium class, that there aren’t wildfire prone areas. It just probably means that in the modelling interval those fueled areas haven’t been touched by fire in the same way, that they are economically viable and protected, or that the cultural use of fire is not the same as in some northern areas.

In a context of limited resources, what is available doesn’t actually need to equally cover the 8.5 Mha. In a combined effort of land management and suppression efforts, over a period of time (one year just isn’t enough), knowing exactly what are the priority areas, I believe I have demonstrated it is possible to know where to start acting yesterday. It is true that empirical knowledge somewhat already told us where to look, but the millions of Euros that are poured over this every year are just incompatible with empiricism. The science exists. This text shows it.


Additional resource

To browse an online wildfire susceptibility map for mainland Portugal, be sure to check this link.