The impact of an ever-changing territory on wildfire hazard assessment

EarthzineArticles, Earth Observation, Original, Sections, Themed Articles, Wildfires Theme


Wildfires are a common occurrence in mainland Portugal, not only during summer months, when the conditions are historically more fire prone, but also during other seasons when temperature rises and rainfall is lower than normal. Several attempts have been made to assess and model wildfire hazards, and it is possible to achieve very satisfactory results with models of low complexity by correlating themes that have high impact on wildfires. One of those themes is land cover, obtained through Earth observation methods. This paper aims to assess how changing from Corine Land Cover 2000 to 2006 impacts predictive capacity.

1. Introduction

Wildfires are a common phenomenon in mainland Portugal, as in other Mediterranean countries. åÊPrescribed burns, an easy and affordable method of fuel management, may cause wildfires, as can climatic factors and human negligence. Wildfires are notåÊ exclusive to Portugal, and as such many authors have studied how to best model them (Chuvieco and Congalton, 1989; Viegas et al., 1999; Vasilakos et al., 2007; and Verde and ZÌ»zere, 2010 among others). Two-thirds of Portugal consists of forested spaces. Even greater areas are susceptible‰ÛÒ including more than 8.4 million hectares of rural landscapes-. The widespread risk makes it particularly important to assess wildfire susceptibility.

Wildfires consumed in excess of 3.5 million hectares in Portugal from 1980 to 2012.Not all fires represent a total loss. Some might be useful for pasture renewal and fuel management. Even so, it is difficult to tame wildfires with more than 20,000 ignitions per year, as has been recorded from mid-1990s onward (Fig.1).

    Figure 1 ‰ÛÒ Ignitions and burnt areas in mainland Portugal from 1980 to 2012 (2012 is pending publication of final results). Source: ICNF, 2013.

Figure 1 ‰ÛÒ Ignitions and burnt areas in mainland Portugal from 1980 to 2012 (2012 is pending publication of final results). Source: ICNF, 2013.

Attempts have been made to model susceptibility by means of different methods, as in Amatulli et al. (2007), by applying interpolation techniques to map lightning- or human-caused wildfires; or DurÌ£o et al. (2010), whose work deals with the Canadian FWI system that assesses the probability of fire in a given region by running simulations. Pereira et al. (2005), Trigo et al. (2006), and Le Page et al. (2008) have explored the correlations of wildfires and weather conditions, hinting at possibilities for wildfire prevention and driving efforts for the better prediction of those conditions that favor fire spread. This paper focuses on susceptibility as a property ofåÊ a given territory, under the assumption that even though a land parcel is not static in nature, it is sufficiently stable to support a structural approach. A previous paper by Verde and ZÌ»zere (2010) has demonstrated that a model of low complexity is capable of providing good results in mapping wildfire hazard and financial risk, when economic data is considered. In fact, that paper has shown that land cover, slope and wildfire historical data together are capable of yielding useful predictions.

2. Conceptual Framework

In Section1, we established that wildfires pose a problem for a considerable area of mainland Portugal. For purposes of clarification, and considering how the concepts on hazard and risk can be differently understood depending on the knowledge of the reader, it is necessary to establish what the authors understand by susceptibility, hazard and risk. A consensus on the concept of wildfire risk still eludes the scientific community. Bachmann and Allg̦wer (1999) allude to this fact when stating that ‰ÛÏthe somewhat inconsiderate use of the various terms ‰Û÷danger’, ‰Û÷hazard’, and ‰Û÷risk’ may result in misunderstandings that can have fatal consequences‰Û . The conceptual framework we adopt in this paper is the same used to predict other hazardous phenomena, like mass movements and floods, following the proposal of UNDRO (1979) and Varnes (1984) and the definition of risk developed by Bachmann and Allg̦wer, namely (1999): ‰ÛÏthe probability of a wildfire to occur at a specified location and under given circumstances and its expected outcome as defined by the impacts on the affected objects.‰Û In this paper, wildfire susceptibility is åÊdefined as the terrain’s propensity to suffer a wildfire or to support its spreading, given the terrain’s intrinsic characteristics (e.g., elevation, slope, and vegetation cover). Further, we consider wildfire hazard to mean the probability of a wildfire occurring.

3. Susceptibility assessment on a changing land cover

Susceptibility assessment, as seen in the model proposed by Verde and Z̻zere (2010), integrates widely used variables in wildfire hazard modeling, such as elevation, slope, land cover, average annual rainfall, average number of days with a minimum temperature greater than or equal to 20 degrees Celsius, and past burn scar mapping (transformed into simple probability). Variables best used for dynamic mapping, such as wind speed and direction, were not considered, as they are of most use when a fire is already progressing, a circumstance outside the scope of this work.

A sensitivity analysis was performed in order to assess the best variable combination regarding the prediction capacity. It demonstrated that a simple model, comprising only three variables, is capable of good prediction results (Figure 2). Indeed, the first 20 percent of susceptible area account for more than 50 percent of the burnt area in the prediction block (which was 1995‰ÛÒ2004, predicted against a modeling block of 1975‰ÛÒ1994).That number exceeds 80 percent if the susceptible area increases to 40 percent. These figures represent the two top-most susceptibility classes used in Portugal, ‰ÛÏVery High‰Û and ‰ÛÏHigh.‰Û

The methodological procedures leading to wildfire susceptibility assessment in mainland Portugal took place in a raster-based Geographic Information System, after preparation and transformation of the available information in vector format. Raster processing allows for greater ease of calculation and lower processing power requirements, given that a relatively small logical area has been chosen, of only 0.64 hectares per pixel.

The assumptions that guide this work are unchanged from previous research by Verde (2008) and Verde and Z̻zere (2010). Specifically, this work assumes (1) The probability or likelihood of burned areas can be assessed quantitatively using statistical relationships between the areas burned in the past and a set of spatial databases; and (2) Wildfires, assessed by their burned areas, occur under conditions that can be characterized by the layers of the aforementioned databases, which will then be considered as conditioning factors (or predisposition factors) and integrated into the assumed model.

The adopted pixel size was 80 meters (0.64 hectares), a limitation imposed not only by the digital elevation model from which slopes were derived (, but also by the scale of Corine Land Cover, which, at 1:100,000, made 80 meters an adequate pixel size.åÊ Following our previous work, the model combined slope, land cover and historical data as a probability, derived from annual mapping of burnt areas, for the period of 1975 to 2004.

With a fairly long time series, favorability scores for the various classes of each variable in the model can be computed. With the exception of probability, for which model input values are percentages, favorability scores were calculated for all other variables for modeling (Chung and Fabbri, 1993, Fabbri et al., 2002).

The calculation of favorability scores was made by weighting the number of burnt pixels against the number of pixels available to burn. For operational convenience, the result of this weighting is multiplied by 100 and rounded, thus avoiding performing GIS operations with decimal values, as seen in equation ‰Û܉ÛÜ1. Probability, as defined earlier, has favorability scores computed via equation 2.

the equation

In equation 1, above, Sfx represents the favorability score for variable x, with umAx being the total number of burnt pixels for variable x and x the pixel universe for raster variable x. In equation 2, f is the number of times each pixel has been burnt, and N is the number of years in the data series.

In the process of data integration, the susceptibility score of each and every raster unit, or pixel, was obtained through the multiplication of the favorability scores of all the variables present in that pixel. Following the scores computed for the variables previously presented, the data was reclassified in such a way that all pixels in a certain class would assume their score. Historical data, transformed into an annual probability (as discussed earlier), was integrated into the model along with the scores of all other variables, as shown in equations 3 and 4:

unique conditions

Where F() is the favorability function (previously described in [1]), pa is the annual probability, and Sf is the favorability score for each model variable.

For data analysis and support of some choices made in this work, success and prediction rates were calculated, and their curves were also plotted (Chung and Fabbri, 2005). Areas under the curve have also been calculated (Bi and Bennett, 2003, Liu and Li, 2005).

Success rates were computed by cross tabulation of the resulting maps and the burn scars considered in the model. For each favorability value, the number of burnt pixels was computed, assuming that the highest favorability score should correspond to the highest burnt area.åÊ These scores were placed in descending order and afterwards plotted, crossing the percentage of burnt area and the percentage of total area. Prediction rates were computed in the same way, with the exception that burnt areas were those not considered in the model runs, for which they allow an independent validation.

Areas under the curve (AUC) can be used to verify which curves show the best results. As success and prediction curves are presented with percentages on their axis, the AUC can also be presented and interpreted as a percentage. To calculate the AUC for success and prediction curves, they were deconstructed into smaller rectangular areas and added together.

The original work fell short, however, of demonstrating what the effect would be if the land cover layer

Figure 2 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010).

Figure 2 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010).

changed. The first studies were conducted with only Corine Land Cover 2000, and since then the 2006 coverage has become available, allowing studying the impact of approximately six years of land cover changes. Land cover was, indeed, the most volatile layer of the three considered in the model. Major changes in slope are not expected to occur, but land cover may change very quickly from year to year. The speed at which land cover mapping is produced does not follow the potential for those changes to occur, making it impossible to map wildfire susceptibility (and hazard, or even risk) against an ever-changing territory. It is, then, useful to study how land cover change impacts the model.

Corine Land Cover was considered an adequate predictor of land cover coverage in terms of its scale and purpose. ThisåÊ wildfire hazard model was not intended for local (or very high scale) assessments; it was intended for nationwide analysis, for which Corine Land Cover’s scale of 1:100,000 is adequate. Furthermore, this land cover model could åÊbe extended to other European countries if desired, provided that the fundamental assumptions behind the proposed model were still met. Those assumptions are 1) The probability of occurrence of burnt areas can be quantitatively assessed by statistical relationships between past burnt areas and a spatial dataset; and 2) Wildfires, assessed by their respective burnt areas, occur under conditions that can be characterized by the layers in the aforementioned spatial dataset, thus considered conditioning (or predisposal) variables to be integrated in the prediction model.

Wildfire susceptible territories are not only those considered as forested areas. In a broader sense, any rural territory (forest and agriculture) is susceptible and demands intervention and fire suppression efforts, on most occasions. For that reason, in the first level, major categories 1, 4 and 5 of Corine Land Cover were left out of the model. When Corine Land Cover 2006 was published, the 2000 coverage was also released with revisions. Table 1 shows what classes have been taken into account.

Table 1 ‰ÛÒ Thematic layers and favorability values of variables. Burnt pixels for the period 1975‰ÛÒ1994

table 1

4. Result and Discussion

Figure 3 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010). Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB ‰ÛÒ Success, Corine Land Cover 2006, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB ‰ÛÒ Prediction, Corine Land Cover 2006.

Figure 3 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010). Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB ‰ÛÒ Success, Corine Land Cover 2006, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB ‰ÛÒ Prediction, Corine Land Cover 2006.

Computing unique condition scores of favorability and presenting them in descending order when crossed with burnt areas allowed the plotting of two types of curve: success and prediction rate curves. The success rate curve results from the cross tabulation between the model results and the burnt areas used to build the model, which were those of 1975 to 1994. Therefore, this curve is able to evaluate the degree of model fit. The prediction rate curve results from the cross tabulation between the model results and an independent set of burnt areas that was not used in the model, from 1995 to 2004. Hence, prediction rate curves can be used to predict the future behavior of wildfires.

In Figure 3, success and prediction curves for the two Corine Land Cover layers are shown. In Table 2 the percentage of burnt area out of the total area for each curve is shown.

Table 2 ‰ÛÒ Area under the curve marks for success and prediction curves. Higher values in bold. Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB ‰ÛÒ Success, Corine Land Cover 2006, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB ‰ÛÒ Prediction, Corine Land Cover 2006.

another table

The difference for the modeling set of 1975‰ÛÒ1994 between the coverage of 2000 and 2006 is not very significant. Even though the more recent coverage is generally better in predicting future wildfires ‰ÛÒ and it should be noted that Corine Land Cover 2006 is on the uppermost part of the prediction block for 1995 to 2004 åʉÛÒ åÊthe results are very similar, usually around 1 percent in difference, which leads to the conclusion that changing land cover layers would not be important for the overall predictive capacity of the model. In fact, one could easily wonder if the gain in prediction would balance the time investment in preparing new data and running a new model for such a low increase in the results. Still, if the difference between 2000 and 2006 is small, it could be that the territory has not changed that much in this six-year period. It may also be that the most-changed territories are not those where wildfires usually occur, and, as such, the land cover classes that best support wildfires in the model are those that will probably take the most time to change.

Figure 4 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010). Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB' ‰ÛÒ Success, Corine Land Cover 2006 with 2000's scores, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB' ‰ÛÒ Prediction, Corine Land Cover 2006 with 2000's scores.

Figure 4 ‰ÛÒ Success and prediction curves for the model proposed by Verde and ZÌ»zere (2010). Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB’ ‰ÛÒ Success, Corine Land Cover 2006 with 2000’s scores, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB’ ‰ÛÒ Prediction, Corine Land Cover 2006 with 2000’s scores.

It could also be argued that using either Corine Land Cover 2000 or 2006 to model burnt areas of years other than those immediately before and after data capture would not be adequate, as there could not be a correspondence between the actual land cover of any given year with that of the year from which burnt scars are taken. However, in a previous work, Verde (2008) has shown that the effectiveness of the model was not affected when combining land cover of the year 2000 with burnt scars of the period 1975‰ÛÒ1994, which the author tested, and we believe the same applies to the land cover of 2006. In fact, that author has shown that, using land cover of the year 2000, the same model had an overall better behavior with older burnt scars (e.g. 1975‰ÛÒ1984) than with a block comprising the year the land cover was created with (1995‰ÛÒ2004). In all likelihood, this will not be the case for rapidly changing land covers. Only on somewhat stable landscapes can this be considered valid.

Given what has been presented so far, another test on the available data can be performed. How would the model behave if the scores that have

been computed for the Corine Land Cover 2000 coverage were used with Corine Land Cover 2006, assuming that any given variable within that theme retained its favorability, despite potential land cover changes? For that purpose, a new theme was created, picking the 2006 land cover and reclassifying the raster to the 2000 land cover scores. After running the model, with the same dataset (e.g., 1975‰ÛÒ1994 as modeled burnt areas and 1995‰ÛÒ2004 as an independent validation set), success and prediction curves were again plotted. The result is show in Figure 4, where the new curves are compared with the ones shown above in Figure 3.

The areas under the curves are displayed in Table 3, and in Table 4 the areas under the prediction curves are presented.

Table 3 ‰ÛÒ Area under the curve marks for success and prediction curves. Higher values in bold. Key: SA ‰ÛÒ Success, Corine Land Cover 2000, SB’ ‰ÛÒ Success, Corine Land Cover 2006, PA ‰ÛÒ Prediction, Corine Land Cover 2000, PB’ ‰ÛÒ Prediction, Corine Land Cover 2006

table 6

Table 4 ‰ÛÒ Areas under the curve for the prediction curves.

last table

5. Conclusions

Figure 5 ‰ÛÒ Wildfire susceptibility for mainland Portugal, adapted from Verde and ZÌ»zere, 2010.

Figure 5 ‰ÛÒ Wildfire susceptibility for mainland Portugal, adapted from Verde and ZÌ»zere, 2010.

Following the previous work by Verde and ZÌ»zere (2010) with the same methodology, burnt areas were tested against two versions of Corine Land Cover the 2000 and 2006 coverage. Differences between the two resulted in very small differences in the prediction capacity of the model when tested with a modeling block comprising the period 1975‰ÛÒ1994 and an independent block, for validation purposes, of the period 1995‰ÛÒ2004. There is more data currently available from burnt scar mapping up to 2012, but the authors chose to use the same year intervals for comparison.

Even though differences were small in this exercise, land cover layers should be made available as soon as they are revised and updated. Not doing so would leave users unaware of potentially hazardous changes in the territory. When revising the model, and even with very small changes in predictive capacity, researchers have an opportunity to become aware of how land cover is changing, sensing if the model retains its capacity. Any significant change detected in the success and prediction curves might show that the assumptions supporting the model are no longer valid, and that different or additional layers, or a different method, should be applied so that the results continue to serve the purpose of providing users with a useful susceptibility map (Figure 5).

As the results showed, Corine Land Cover 2000 scores were robust and could be maintained in the future, even with more recent land cover coverages, but it is useful to upgrade themes as they become available and rerun the model, so the results better represent the ever-changing land cover reality.

Joo Verde Ph.D. RISKam Environmental Hazard and Risk Assessment and Management, Centre of Geographical Studies (CEG), Institute of Geography and Spatial Planning, University of Lisbon (IGOT-UL),


Amatulli, G., Per̩z-Cabello, F. and de la Riva, J. 2007. Mapping lightning/human-caused wildfires occurrence under ignition point location uncertainty, Ecol. Model., 200, 321-333.

Bachmann, A. and Allg̦wer, B. 1999. The need for a consistent wildfire risk terminology, in: Proceedings of the Joint Fire Science Conference and Workshop, Boise Idaho, E.U.A., 15-17 June.

Bi, J., Bennett, K. 2003. Regression Error Characteristic Curves. Proceedings of the Twentieth International Conference on Machine Learning (ICML-2003), Washington DC, 8 p.

Chuevieco, E. and Congalton, R.G. 1989. Application of remote sensing and geographic information systems to forest fire hazard mapping, Remote Sens. Environ., 29, 147-159.

Chung, C.F., Fabbri, A. 2005. Systematic procedures of landslide hazard mapping for risk assessment using spatial prediction models. In: Glade , T., Anderson, M.G., Crozier, M.J. (Eds.), Landslide Hazard and Risk. Wiley, Chichester, 139‰ÛÒ174.

Chung, C. F., Fabbri, A. 1993. The representation of geoscience information for data integration, Nonrenewable Resources, 2 (2), 122‰ÛÒ138.

Duṛo, R.M., Pereira, M.J., Branquinho, C. and Soares, A. 2010. Assessing spatial uncertainty of the Portuguese fire risk through direct sequential simulation, Ecol. Model., 221, 27-33.

Fabbri, A., Chung, C. F., Napolitano, P., Remondo, J., and ZÌ»zere, J. L. 2002. Prediction rate functions of landslide susceptibility applied in the Iberian Peninsula, edited by Brebbia, C. A., Risk Analysis III, Series: Management Information Systems, WIT Press, Southampton, Boston, vol. 5, 703‰ÛÒ718.

ICNF. 2013. RelatÌ_rios de IncÌ»ndios Florestais,, accessed June 2013.

Liu, H., Li, G. 2005. Testing Statistical Significance of the Area under a Receiving Operating Characteristics Curve for Repeated Measures Design with Bootstrapping. Journal of Data Science, 3, 257-278.

Pereira M.G., Trigo R.M., DaCamara C.C., Pereira J.M.C. and Leite S.M. 2005. Synoptic patterns associated with large summer forest fires in Portugal, Agr. Forest Meteorol., 129, 11-25.

Trigo R.M., Pereira J.M.C., Pereira M.G., Mota B., Calado M.T., DaCamara C.C. and Santo F.E. 2006. The exceptional fire season of summer 2003 in Portugal, Int. J. Climatol., 26, 1741-1757.

Le Page, Y., Pereira, J.M.C., Trigo, R.M., da CÌ¢mara, C., Oom, D. And Mota, B. 2008. Global fire activity patterns (1996-2006) and climatic influence: an analysis using the World Fire Atlas, Atmos. Chem. Phys., 8, 1911-1924.

UNDRO. 1979. Natural Disasters and Vulnerability Analysis, Report of Expert Group Meeting 9-12 July 1979, Office of the United Nations Disaster Relief Coordinator, Geneva.

Varnes, D.J. 1984. Landslide hazard zonation: a review of principles and practice, UNESCO, Paris.

Vasilakos, C., Kalabokidis, K., Hatzopoulos, J., Kallos, G. and Matsinos, Y. 2007. Integrating new methods and tools in •Âre danger rating, Int. J. Wildland Fire, 16, 306‰ÛÒ316.

Verde, J. 2008. Wildfire Hazard Assessment, MSc thesis, University of Lisbon (in Portuguese), Portugal.

Verde, J. C. and Z̻zere, J. L. 2010. Assessment and validation of wildfire susceptibility and hazard in Portugal, Nat. Hazards Earth Syst. Sci., 10, 485-497, doi:10.5194/nhess-10-485-2010.

Viegas, X., Bovio, G., Ferreira, A., Nosenzo, A. and Bernard, S. 1999. Comparative study of various methods of fire danger evaluation in Southern Europe, Int. J. Wildland Fire, 9 (4), 235‰ÛÒ246.