CSIRO PUBLISHING

www.publish.csiro.au/journals/ijwf

International Journal of Wildland Fire, 2006, 15, 197–202

Modelling the effects of distance on the probability of fire detection from lookouts Francisco Castro RegoA and Filipe Xavier CatryA,B A Centro

de Ecologia Aplicada Prof. Baeta Neves, Instituto Superior de Agronomia – Tapada da Ajuda, 1349-017 Lisboa, Portugal. B Corresponding author. Email: [email protected]

Abstract. In the management of forest fires, early detection and fast response are known to be the two major actions that limit both fire loss and fire-associated costs. There are several inter-related factors that are crucial in producing an efficient fire detection system: the strategic placement and networking of lookout towers, the knowledge of the fire detection radius for lookout observers at a given location and the ability to produce visibility maps. This study proposes a new methodology in the field of forest fire management, using the widely accepted Fire Detection Function Model to evaluate the effect of distance and other variables on the probability that an object is detected by an observer. In spite of the known variability, the model seems robust when applied to a wide variety of situations, and the results obtained for the effective detection radius (13.4 km for poor conditions and 20.6 km for good conditions) are in general agreement with those proposed by other authors. We encourage the application of the new approach in the evaluation or planning of lookout networks, in addition to other integrated systems used in fire detection. Additional keywords: detection; forest fires; lookout towers; Portugal; visibility.

Introduction Reducing the severity of loss associated with wildfire is highly dependent on measures such as fire prevention, detection and fire fighting. When fire prevention fails, fire-fighting steps must be initiated, but this can only happen following detection. The best way of reducing fire fighting costs and fire losses is therefore early detection and rapid extinguishment (Chandler et al. 1983). In Portugal, lookout towers remain the principal system of organized detection with a permanent network of 237 observation points, generally operating 24 h per day during the ‘normal’ fire season (from 1 June to 30 September). As in many other countries the lookout system is regarded as a characteristic feature of fire control (Davis 1959) that operates on an annual budget of ∼2.5 million euros (Galante 2001). This annual budget, combined with the fact that the percentage of first detections made in Portugal by lookouts is relatively small (16.5% in 2001, 12.4% in 2002, and 10.9% in 2003), highlights the need to study the factors involved in the probability of detecting fire by the network. Factors such as distance, topography, air transparency, time of the day, and existence of other detection systems all play an important role in determining efficiency of lookouts, and thus should be considered equally when planning an integrated system for © IAWF 2006

fire detection. This is especially relevant, for example, whenever decisions are made determining whether existing towers should be decommissioned or new towers are to be established (Brown and Davis 1973; Ruiz 2000; Catry et al. 2004). The present paper addresses the effects of distance on the probability that a fire is detected by a lookout tower, taking into account the indirect effects of topography and air transparency. Other factors known to affect detection probability are discussed and a review of the current state of the art is provided. State of the art The time between the start of a fire and the time suppression forces actually begin to deal with it can be subdivided into four periods: discovery time, report time, get-away time of the initial combating force, and travel time of the combating force (Buck 1938). Discovery time is certainly a very important component of the total time between fire initiation and fire fighting; however, it is very difficult to evaluate as, in reality, no information is generally available before the fire is detected. As most forest fires are detected by the human eye, studies have to take into account the accuracy and limitations of visual detection, as shown in previous studies relating fire detection 10.1071/WF04016

1049-8001/06/020197

198

Int. J. Wildland Fire

with visibility of the smoke column (Buck and Fons 1936; Buck 1938; Byram and Jemison 1948; Davis 1959). It was recognized early on that for the question ‘How far can smoke be seen?’ no answer could be given without first asking ‘What kind of a smoke?’The time required for a lookout to discover smoke of rapidly increasing proportions is clearly of more use in practical terms than the detection radius for smoke of constant size and density (Bruce 1941). In order to understand which factors could influence the discovery time of forest fires by lookout observers, Buck (1938) studied 200 test fires that were set and allowed to burn for 10–30 min in ponderosa pine forests of northern California. He concluded that under similar conditions of position of the sun and atmospheric visibility, the discovery time could be predicted as a function of distance and fire rate of spread. Haze and air-light were also found to be very important in the pioneering studies of Byram and Jemison (1948) on the principles of visibility and their application to forest fire detection. Theoretically, the effect of distance in a perfectly transparent atmosphere is only to control the visual angle subtended by the smoke column at the eye of the observer. Therefore, at lower distances average discovery times may be considered as a straight-line function of distance. However, for distances greater than 30 km, the effect of distance on discovery time decreases because fires allowed to burn long enough to become visible at such distances normally begin producing smoke at a greatly accelerated rate (which become visible at relatively long distances within a very short time). This combined influence was expressed by Buck (1938) as the ratio between distance and area burned per unit time. To complicate matters even further, variables affecting visibility and detection should also include personal factors associated with the observer. The fact that a fire is visible to a person of normal vision does not necessarily mean that a particular observer will detect it. A number of variables that affect detection by human observers are: quality of eyesight, searching techniques, attention and concentration, position, observation conditions, and equipment aids (Davis 1959). Finally, even the definition of visibility is not simple. Visibility distance is used to denote the maximum distance at which a small amount of smoke can be seen and identified as being smoke (Byram and Jemison 1948). There is an obvious need, however, for defining a standard against which to measure detection, as proposed for example by Davis (1959). In spite of all of the above complications the visibility radius of lookout towers has been used to produce cartography of the seen/unseen areas and consequently for planning and managing networks of lookout posts (Brown and Davis 1973; Ruiz 2000). Detection function The difficulties associated with the analysis of visual detection of objects are not unique to the fire problem. There are

F. C. Rego and F. X. Catry

other fields of science where similar problems of detection as a function of distance are already considered and solutions proposed (Buckland et al. 1993). Distance sampling using detection functions has been used traditionally in the analysis of biological populations. Studies of birds represent a major use of these methods, in addition to the successful application to population studies of many terrestrial mammals (ranging from mice to primates), reptiles, amphibians, beetles and spiders, marine mammals (from dolphins to whales), fish (in coral reefs) and crabs estimated from underwater survey data. Furthermore, many inanimate objects such as nests and burrows have been surveyed using distance sampling that could also be applied to plant populations or even to non-traditional applications such as the number of mines anchored to the seabed in mine fields (Buckland et al. 1993). Within modeling detection functions, the point transect sampling method seems particularly appropriate to deal with the detection of objects (fires) from fixed points (lookout towers). In this method, points are set in the field and the observer records the distance to detected objects of interest. The sample data are the set of sighting (radial) distances between the point and the detected objects and any other relevant qualitative or quantitative related variables. Modeling point transect data takes into account that the area of the ring of incremental width dy at distance y from the observer is 2πydy, and thus proportional to y. The theory also assumes that objects in the area close to the point are detected but that many of the objects will go undetected as there is a general tendency for detection to decrease with increasing distance from the point. Therefore, underlying the theory is the concept of a detection function (Buckland et al. 1993): g( y) = prob{detection | distance y}. Importance must also be given to the concept of an effective detection radius (EDR) where the distance for which unseen objects located closer to the point than EDR equals the number of objects seen at distances greater than EDR. The effective detection radius is therefore critical when planning distances between towers. In this study we propose the use of detection functions for analysing the effects of distance and other variables in the detection of forest fires. Methods Study case: the 2001 fire season in Portugal During the fire season of 2001 the Portuguese Forest Services (Direcção-Geral das Florestas, DGF) established a database with records detailing the approximate initial geographical coordinates of each fire occurrence and the lookout tower responsible for its detection. The coordinates of the towers were also provided by DGF. It was therefore possible to determine distances to a fire from each tower.

Effects of distance on forest fire detection

Int. J. Wildland Fire

The database used for detection function analysis consisted of 9434 fires detected by 207 towers, with an additional 30 towers showing no first detections. The database was analyzed using the software DISTANCE (Buckland et al. 1993; Thomas et al. 2002). DISTANCE DISTANCE is based on the standard work of Buckland et al. (1993) and updated by Buckland et al. (2001) who described conventional distance sampling in detail by covering aspects of survey design with several worked examples. The modeling process in detection function can be conceptualized in two steps. First is the choice of a ‘key function’ as the half-normal, the hazard-rate or the negative exponential. Second, there is the selection of a ‘series expansion’, used to adjust the key function, using perhaps one or two additional parameters. The series expansions considered in DISTANCE are the cosine series, the simple polynomials, and the Hermite polynomials (Buckland et al. 1993). In the present study Akaike’s Information Criterion (AIC) (Akaike 1973) was used as the quantitative method for model selection. The adequacy of the selected model was assessed using the Chi-square goodness of fit statistics and visual inspection of the estimated detection function. The selected key function was the half-normal with cosine adjustments of second order. In general, when histograms of distance data decline markedly with distance, the half-normal key often represents a good choice as a key function (Buckland et al. 1993). The selected model has the form: Key function: Half normal, exp(−y2/2σ 2 ); series expansion: Cosine  aj cos( jπy/w). One of the methods in which DISTANCE can handle heterogeneity is post-stratification of the data, as classes of associate variables. In the present study the variable terrain roughness was quantified for a circle with a 30 km radius around each lookout tower as the ratio (R) of the true 3-dimensional area and the corresponding planimetric 2-dimensional area (Jenness 2001). Classes were established in order to cover areas of similar sizes as: low roughness (R < 1.020), moderate roughness (R from 1.020 to 1.037) and high roughness (R > 1.037). The analyses were based

on a raster digital elevation model for continental Portugal, with a spatial resolution of 25 m produced by the Portuguese Geographical Institute (Lisbon, Portugal). For the variable geographical location, two classes were established, coastal and interior, based on distance to sea. Other variables that could potentially affect air transparency and that were available on environmental maps (IA 2003) were included as candidate covariates: average elevation, temperature, relative humidity, rainfall, and solar radiation. Other factors that could potentially affect detection probability, such as year, season, or time of the day, were known to be less important than geographical location (Jemison 1940) and therefore were not considered in order to keep the model as simple as possible. As right truncation is often recommended for low frequency cases, data were omitted for distances over 50 km to reduce spurious noise (Buckland et al. 1993). Results Using the model selected by the AIC criterion (half-normal with a cosine series expansion) the analysis of the poststratification options showed that only roughness and distance to sea had significant effects on the parameters of the detection function. All other variables were discarded in subsequent analyses. The adjusted parameters for the six classes resulting from combinations of geographical location and terrain roughness are shown in Table 1. From the statistical comparison of the confidence intervals of the parameters of these classes, it was found that the only statistically significant different class was interior/low roughness, whereas all the other classes could be considered together. Therefore, for the interior/low roughness class, combination data were used for 21 towers and 341 detections. The estimator chosen based on minimum AIC was a model with a half-normal key, with cosine adjustments of order 2. For all other situations in the country, data were used for 186 towers and 9093 detections. The estimator chosen was also a model with a half-normal key, with cosine adjustments of order 2. The fit of the model as measured by the goodness-of-fit Chisquare test benefited from higher order adjustments; however, the AIC value only slightly decreased. As a consequence, and in order to keep the model as simple as and comparable with that developed earlier, it was decided that due to the very close

Table 1. Adjusted parameters for the resulting six classes Tower location

Terrain roughness

Coastal

Low Moderate High Low Moderate High

Interior

Parameters of the detection function a2 a1 = σ(105 ) 0.124 0.109 0.102 0.159 0.121 0.118

199

0.567 0.408 0.387 0.139 0.401 0.462

Effective detection radius (km) 13.4 12.8 12.3 20.6 14.1 13.4

200

Int. J. Wildland Fire

F. C. Rego and F. X. Catry

Table 2. Adjusted parameters (±s.e.) for the final two classes Tower location

Terrain roughness

Interior All other situations

Low All other situations

Parameters of the detection function a2 a1 = σ(105 ) 0.159 ± 0.0047 0.118 ± 0.0009

Detection probabilit

(a)

Effective detection radius (km)

0.139 ± 0.0869 0.467 ± 0.0196

20.55 ± 1.087 13.41 ± 0.078

(b)

1.0

1.0

0.9

0.9

0.8

0.8

0.7

0.7

0.6

0.6

0.5

0.5

0.4

0.4

0.3

0.3

0.2

0.2

0.1

0.1

0.0

0.0 0

10 000 20 000 30 000 40 000 50 000

0

10 000 20 000 30 000

40 000 50 000

Radial distance (m)

Fig. 1. Detection probability as a function of radial distance for towers situated in the interior in low roughness terrain (a) and all other situations (b). Observed values as histograms, fitted models as line.

agreement between observed and predicted values, the 2nd order adjustment should be kept. The adjusted parameters for the final two classes are shown in Table 2. The detection function used was as follows: f (y) = exp[−y2/(2a21 )] × [1 + a2 cos(2πy/w)]/(1 + a2 ), where a1 = 0.1588 × 105 , a2 = 0.1385, w = 50000, and y = distance (m). The graph of the detection probability as a function of radial distance is shown for the interior/low roughness situation in Fig. 1a. From the observation of the histogram it seems that the detection function fits badly at small distances but this can be due to a preventive effect of the existence of the tower, or to the large variability associated with the small areas closest to the tower. The fit of the model is not heavily influenced by such short distances as the goodness-of-fit Chi-square test did not show a significant difference between observed and predicted values. The graph of the detection probability as a function of radial distance for the main part of the country is illustrated in Fig. 1b. From the observation of the histogram and the fitted model it becomes apparent that detection probability

decreases much more sharply in this situation than in the interior/low roughness condition. The geographical location of the two different situations in the map of the Portuguese mainland allows for the application of each model in the corresponding area (Fig. 2). Discussion and conclusions Detection of fires from lookouts has been challenged by other techniques used for fire fighting. In North America, for instance, aerial detection has been favoured to ground detection for several decades (Brown and Davis 1973). The use of mobile detection ground units by patrolmen or rangers is another important consideration. In addition, significant studies have been made recently on the application of new technologies for fire detection, and the potential for improved effectiveness appears very promising (Kremens et al. 2003; Utkin et al. 2003). Finally, recent unpublished studies from Portugal and France show that in areas with a high human population density, the general public detects a large proportion of fires. However, in Portugal and in many other parts of the world, a significant part of the fire detection chore still falls to

Effects of distance on forest fire detection

Int. J. Wildland Fire

201

Fig. 2. Geographical location of the two visibility situations in Portugal: Model I, interior in low roughness terrain (26% of the country); Model II, all other situations (74%).

observers in fixed lookout positions in spite of the alternatives and of the increasing cost of maintaining a fixed lookout tower (Chandler et al. 1983). Therefore it is important to fully understand the capacities of the system in order to use it fully, to complement it, or replace it with other systems. The present study proposes a new methodology that is widely accepted and used in other fields of science in order to evaluate the effect of distance and other variables on the probability that a fire is detected by a lookout system. This methodology could also be applied to the analysis of the effect of distance on the probability of fire detection by cameras and other detection systems. The model selected for the lookout system in Portugal seems to be robust and thus could be applied to a wide range of different situations. The results of the effective detection radius (13.4 km for poor conditions and 20.6 km for good conditions) are of the same order of magnitude as those based either on smoke tests or proposed empirically for several different scenarios. In Portugal, early studies (although not for fire detection purposes) suggest that in some coastal weather stations, as in Lisbon and Porto, visibility during the fire season (June to September) generally falls in the range from 10 to 20 km, whereas in some interior areas, as in Évora, visibility is placed in the range from 20 to 50 km (INMG 1992; Catry 2002). In the USA, radii from 13 to 32 km have been used. In a large part of the western USA, a radius of 24 km is accepted

as the standard, as visibility is normally good during the fire season and smoke from a small fire can be seen up to this distance. In the south and south-east, where visibility is poorer, a 10–13 km radius is most often used (Davis 1959; Brown and Davis 1973). In Canada the range of towers in operation is determined by the degree of visibility, which is, on average, up to 18 km (Artsybashev 1984). Fire observation towers in the former USSR were constructed for an effective fire detection radius of 5–7 km. However, when situated on dominating elevations, the effective detection radius was increased 1.5–2 times (Artsybashev 1984). Finally, in Spain the visibility distance of a lookout observer with normal vision is established as 6–8 km for the worst conditions and in 30–40 km for optimal situations. As a rule, values of 10 km and 20 km are used for rough terrain and flat areas, respectively (Ruiz 2000). It can be concluded that the approach proposed in the present study clarifies the concept of the effective detection radius, yielding results that are consistent with the relevant literature for the subject. However, the approach of modeling detection functions has the unique advantage of allowing us to know the probability of when a fire that has started in a given geographical location will first be detected by a lookout observer at a given distance. This is a much more realistic approach than considering that all fires occurring within a fixed distance that are visible from a lookout tower will be detected, whereas all fires outside that range will not

202

Int. J. Wildland Fire

F. C. Rego and F. X. Catry

be detected by that tower. This should be a very important consideration when planning a more effective system of fire detection networks. Acknowledgements We acknowledge Portuguese Forest Services for making available the official data concerning the location of lookout posts and detections made in 2001, and the Portuguese Geographical Institute for their help. We also acknowledge Rui Almeida for his significant help with the organisation of data, Miguel Galante and Miguel Cruz for making data available and Gillian Telfer, Susana Dias and Miguel Bugalho for revising the manuscript. References Akaike H (1973) Information theory and an extension of the maximum likelihood principle. In ‘International Symposium on Information Theory’, 2nd edn. (Eds B Petran, F Csàaki) pp. 267–281. (Akadèemiai Kiadi: Budapest) Artsybashev ES (1984) Detection of forest fires. In ‘Forest fires and their control’, Russian Translations Series 15. pp. 46–61. (AA Balkema: Rotterdam) Brown AA, Davis KP (1973) Forest fire detection. In ‘Forest fire – Control and use’, 2nd edn. (Eds WP Orr, N Frankel, S Langman) pp. 327–344. (McGraw-Hill: New York) Bruce HD (1941) Theoretical analysis of a smoke-column visibility. Journal of Agricultural Research 62, 161–178. Buck CC (1938) Factors influencing the discovery of forest fires by lookout observers. Journal of Agricultural Research 56, 259–268. Buck CC, Fons WL (1936) The effect of direction of illumination upon the visibility of a smoke column. Journal of Agricultural Research 51, 907–918. Buckland S, Anderson D, Burnham K, Laake J (1993) ‘Distance sampling. Estimating abundance of biological populations.’ (Chapman and Hall: London) Buckland S, Anderson D, Burnham K, Laake J, Borchers D, Thomas L (2001) ‘Introduction to distance sampling.’ (Oxford University Press: London) Byram GM, Jemison GM (1948) ‘Some principles of visibility and their application to forest fire detection.’ USDA Technical Bulletin No. 954. (Washington, DC)

Catry F (2002) Estudo da visibilidade em postos de vigia e da sua influência na vigilância de incêndios florestais. Graduation thesis, Escola Superior Agrária – Instituto Politécnico de Castelo Branco, Castelo Branco. Catry FX, Almeida RM, Rego FC (2004) Produção de cartografia de visibilidades para Portugal continental. A importância da sua utilização na vigilância contra incêndios florestais. Silva Lusitana 12, 227–241. Chandler C, Cheney P, Thomas P, Trabaud L, Williams D (1983) ‘Fire in forestry – volume II: forest fire management and organization.’ (John Wiley and Sons: New York) Davis KP (1959) Detection. In ‘Forest fire. Control and use’. (Ed. KP Davis) pp. 282–303. (McGraw-Hill: New York) Galante LM (2001) ‘Rede nacional de postos de vigia.’ (Direcção-Geral das Florestas: Lisboa) IA (2003) ‘Atlas Digital do Ambiente.’ Available at http://www. iambiente.pt/atlas/est/index.jsp [Verified 20 January 2006] INMG (1992) ‘Anuário Climatológico de Portugal. I Parte – Continente e Ilhas doAtlântico-Norte. Vol. VII –Anos de 1951 e 1953.’(Instituto Nacional de Meteorologia e Geofísica: Lisboa) Jemison GM (1940) The influence of locality, season, time of day, and year on the visibility of smoke columns. Journal of Forestry 38, 435–437. Jenness J (2001) ‘Grid surface area manual.’ Available at http://www. jennessent.com/arcview/arcview_extensions.htm [Verified 20 January 2006] Kremens R, Faulring J, Gallagher A, Seema A, Vodacek A (2003) Autonomous fiels-deployable wildland fire sensors. International Journal of Wildland Fire 12, 237–244. doi:10.1071/WF02055 Ruiz EM (2000) Detección. In ‘La defensa contra incendios forestales. Fundamentos y experiencias’. (Coord. R Velez) pp. 17.1–17.15. (McGraw-Hill/Interamaricana de España, SAU: Madrid) Thomas L, Laake J, Strindberg S, Marques F, Buckland S, Borchers D, Anderson D, Burnham K, Hedley S, Pollard J (2002) ‘DISTANCE 4.1, release 2.’ Research Unit for Wildlife Population Assessment. (University of St Andrews: St Andrews, UK) Available at http://www.ruwpa.st-and.ac.uk/distance/ [Verified 27 April 2006] Utkin A, Fernandes A, Simões F, Lavrov A, Vilar R (2003) Feasibility of forest-fire smoke detection using lidar. International Journal of Wildland Fire 12, 159–166. doi:10.1071/WF02048

http://www.publish.csiro.au/journals/ijwf