5i 2009 Manzo Delgado Etal IJRS

This article was downloaded by: [Inst De Geografia]
On: 13 November 2009

Access details: Access Details: [subscription number 908627206]
Publisher Taylor & Francis
Informa Ltd Registered in England and Wales Registered Number: 1072954 Registered office: Mortimer House, 37-
41 Mortimer Street, London W1T 3JH, UK
International Journal of Remote Sensing

Publication details, including instructions for authors and subscription information:
http://www.informaworld.com/smpp/title~content=t713722504
Assessment of seasonal forest fire risk using NOAA-AVHRR: a case study

in central Mexico
L. Manzo-Delgado a; S. Sánchez-Colón b; R. Álvarez c
a
Instituto de Geografía, Universidad Nacional Autónoma de México, Ciudad Universitaria, Coyoacán,
Mexico DF b Coordinación Regional Para América Latina y El Caribe, División de Evaluación y Alerta
Temprana, Programa de las Naciones Unidas para el Medio Ambiente, Colonia Lomas Virreyes,
Mexico, DF c Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad
Nacional Autónoma de México, Ciudad Universitaria, Coyoacán, Mexico DF
To cite this Article Manzo-Delgado, L., Sánchez-Colón, S. and Álvarez, R.'Assessment of seasonal forest fire risk using
NOAA-AVHRR: a case study in central Mexico', International Journal of Remote Sensing, 30: 19, 4991 — 5013
To link to this Article: DOI: 10.1080/01431160902852796
URL: http://dx.doi.org/10.1080/01431160902852796
PLEASE SCROLL DOWN FOR ARTICLE
Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf
This article may be used for research, teaching and private study purposes. Any substantial or
systematic reproduction, re-distribution, re-selling, loan or sub-licensing, systematic supply or
distribution in any form to anyone is expressly forbidden.
The publisher does not give any warranty express or implied or make any representation that the contents
will be complete or accurate or up to date. The accuracy of any instructions, formulae and drug doses
should be independently verified with primary sources. The publisher shall not be liable for any loss,
actions, claims, proceedings, demand or costs or damages whatsoever or howsoever caused arising directly
or indirectly in connection with or arising out of the use of this material.
Vol. 30, No. 19, 10 October 2009, 4991–5013
Assessment of seasonal forest fire risk using NOAA-AVHRR: a case

study in central Mexico
L. MANZO-DELGADO*†, S. SÁNCHEZ-COLÓN‡ and R. ÁLVAREZ§

†Instituto de Geografı́a, Universidad Nacional Autónoma de México, Ciudad
Universitaria, CP 04510 Coyoacán, Mexico DF
‡Coordinación Regional Para América Latina y El Caribe, División de Evaluación y
Alerta Temprana, Programa de las Naciones Unidas para el Medio Ambiente, Blvd de los
Virreyes No. 155, Colonia Lomas Virreyes, CP 11000, Mexico, DF
§Instituto de Investigaciones en Matemáticas Aplicadas y en Sistemas, Universidad
Nacional Autónoma de México, Ciudad Universitaria, CP 04510, Coyoacán, Mexico DF
Downloaded By: [Inst De Geografia] At: 23:18 13 November 2009
A logistic model was constructed to assess the risk of forest fire and tested over the
central region of Mexico. The model incorporates both static and dynamic
predictive variables: elevation, aspect, slope, vegetation type, precipitation,
Normalized Difference Vegetation Index (NDVI), land surface temperature
(LST), and cloud cover. The latter three variables were derived from National
Oceanic and Atmospheric Administration (NOAA) Advanced Very High
Resolution Radiometer (AVHRR) images from the four months (November to
February) before the fire seasons. Actual forest fires were detected on
NOAA-AVHRR images from the four fire seasons (March to May) from 1997
to 2000. Variables included in the model were chosen following a stepwise strategy.
Statistically, the January NDVI, the February LST, vegetation type and slope had
the greatest influence on the distribution of forest fires; however, elevation and
precipitation were also included in the final model. The probability of forest fire
occurrence for each fire season from 1997 to 1999 was mapped. The accuracy of the
model was estimated to be 79.8% with reference to sensitivity, specificity and
receiver operating characteristic (ROC) curves. Model predictions were validated
against data from the 2000 fire season. The fire occurrence probability map is
useful for designing large-scale management strategies for wildfire prevention not
only in the test area of this study but also in regions where the static and dynamic
variables can be similarly defined.
1. Introduction
Forest fires are disturbances that act as a selecting force and lead to adaptation of
species; they also bring about changes in community richness and diversity (Pickett
et al. 1989, Christensen 1993, Bond and van Wilen 1996). The occurrence of forest
fires is regulated by physical (topography and climate), biological (community struc-
ture and functioning) and anthropogenic (often caused by the cultural use of fire,
tourism and transport) factors. However, in recent decades climatic anomalies have
been more frequent, such as droughts and rainfall deficit induced by the El Niño
Southern Oscillation (ENSO) phenomenon, which, when present, has contributed to
the increment of forest fire occurrence, especially in the tropics (Leighton and
*Corresponding author. Email: llmanzo@igg.unam.mx

ISSN 0143-1161 print/ISSN 1366-5901 online # 2009 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160902852796
4992 L. Manzo-Delgado et al.
Wirawan 1986, Cochrane et al. 1999). The 1997–1998 ENSO phenomenon was
considered the strongest of the last century (Wolter and Timlin 1998, McPhaden
1999), causing catastrophic forest fires in Southeast Asia (Indonesia), Central
America and Mexico, South America (Brazil, Peru and Bolivia), Europe (Greece,
Italy and Spain), Russia (Khabarovsk Krai) and China (Rowell and Moore 2000).
Data from the US National Oceanic and Atmospheric Administration (NOAA) for
the 10 strongest ENSO events of the past century show that, since 1970, their
frequency and intensity have increased. The 1997–1998 ENSO is an example of the
progressive evolution of this phenomenon. The frequency of ENSO may be changing
as a result of the increases in greenhouse-gas concentration (Timmermann et al. 1999).
According to the forecasts for ENSO events, catastrophic forest fires could become
even more frequent, although there is no agreement in the scientific community about
this topic. ENSO, when present, can contribute in an additional, anomalous way to
forest fire occurrence. To be ready to cope with these disturbances requires the
assessment of forest fire risk, particularly during the drought season, when the
vegetation is most vulnerable to fire. If ENSO events do become more frequent and
intense, catastrophic forest fires are expected to become more frequent. Preparation
for these disturbances requires an early warning system such as the assessment of
seasonal forest fire risk, to allow identification of areas and critical time periods in
advance of their occurrence.
In the test area of this study, the central portion of Mexico, forest fires are an
annual event and have followed an increasing trend during the past three decades.
During the 1970s, the mean number of fires that occurred each year was 5129, with a
standard deviation (SD) of 955, burning a mean of 158 204 ha (SD = 47 763 ha),
increasing during the 1980s to 6736 fires (SD = 2659) and 257 980 ha (SD = 155 045),
and during the 1990s to 10 042 fires (SD = 5470) and 267 689 ha (SD = 248 390). That
tendency of fires and burning has been influenced not only by human activities but
also by climatic anomalies, such as ENSO and hurricanes of great magnitude. A
comparative analysis during the past three decades between ENSO anomalies and the
total area burning in Mexico each year (figure 1) revealed that the catastrophic
episodes occurred after the 1986–1987 ENSO (first semester of 1988) or during the
Figure 1. Sea-surface temperature (SST) anomalies (solid line) in the central Pacific and total
area burning in Mexico (broken line) from 1997 to 2001. The wildland fire episodes during 1988
and 1998 were largely influenced by the drought caused by ENSO. Note the delay between the
temperature maxima in the SST and the maxima of fire occurrences.
ForestSAT 2007 4993
final stage of the 1997–1998 ENSO (from March to June). After the first strong
ENSO, in 1988, 10 942 fires were recorded, burning 518 265 ha. The subsequent
event, in 1998, was even stronger, with 14 300 fires that affected 849 632 ha, more
than twice the average in the 1980s (SEMARNAT 2002), with the resulting smoke
travelling hundreds of kilometres and reaching the USA and several countries in the
Caribbean region (Rodrı́guez-Trejo and Pyne 1998). The areas most impacted by fire
were the highlands of the Sierra Madre Occidental, the central region of Mexico and
the tropical mountains of the states of Oaxaca and Chiapas. In an area of the central
region, some 6000 fires were reported (46% of the total), affecting 32 000 ha
(SEMARNAP 1998). Hence, the central region of Mexico can be considered a valid
test ground for designing strategies for the assessment of forest fire risk that could
subsequently be adapted to other regions’ characteristics.
1.1 Current estimates

Following the extreme fire events of 1998, fire management agencies in tropical
countries became aware of the need for an integrated fire information system.
Therefore, at the request of such countries, the Canadian Forest Service developed
an operational prototype system for Mexico and Southeast Asia based on the
Canadian Forest Fire Danger Rating System (CFFDRS; Stocks et al. 1989), largely
because of its simplicity and its strong interpretive products. The CFFDRS supplies
cartographic information for the assessment of forest fire risk from two primary
subsystems: the Canadian Forest Fire Weather Index (FWI) System and the
Canadian Forest Fire Behavior Prediction (FBP) System. These data are used as a
guide in a wide variety of fire management applications such as monitoring national
fire danger and wildfire activity on a daily basis (Lee et al. 2002). The Mexico Fire
Information System was developed in cooperation with the Secretariat of the
Environment and Natural Resources (SEMARNAT) of Mexico (Goldammer
et al. 2003). This system was launched on the Web in the spring of 1999 (http://
app1.semarnat.gob.mx/dgeia/compendio06/Reportes/D1_desastres/D1_r_desastre00_
02.htm). The website offers daily maps of fire weather and fire behaviour potential for
Mexico based on real-time weather data observed in Mexico and disseminated by the
World Meteorological Organization through the Canadian Wildland Fire Information
System.
Daily registers of forest fire probability are available from SEMARNAT-CFFDRS
through a cooperative programme (Goldammer et al. 2003). However, they do not
provide a seasonal evaluation of such a risk in advance. To produce the seasonal
predictions, we have developed a statistical model that involves remote sensing
measurements to obtain and integrate the necessary variables and to predict poten-
tially catastrophic situations in a given region. We consider that both types of
evaluations are necessary and complement each other.
1.2 Logistic model

In this paper, the forest fire risk is defined as the probability for a fire to occur in a
particular area (Chou et al. 1993). Previous work on the assessment of forest fire risk
has been based on meteorological data (Viegas et al. 2000), vegetation types and fuel
build-up (Burgan et al. 1998). The logistic regression model (O’Brien 1992, Crawley
1993) is particularly suitable for modelling the probability of forest fire occurrence
(Vega-Garcı́a et al. 1999, Dı́az-Avalos et al. 2001). Logistic regression is a mathema-
tical modelling approach that can be used to describe the relationship of several
independent variables (quantitative or categorical) to a binary response variable
(dichotomous dependent variable). The logistic function describes the mathematical
form on which the logistic model is based. That function can be specified as:
f ðzÞ ¼ 1=ð1 þ ez Þ (1)
When z is –1, the logistic function f(z) equals 0, and when z is +1, then f(z) equals 1.
Thus, the range of this function is from 0 to 1. This is the primary reason why the
logistic model is so popular and it is often the first choice when a probability is to be
estimated, such a risk of occurrence. Another characteristic of the logistic function is
the S-shape, which represents an index that combines the contributions of several risk
factors (Kleinbaum and Klein 2002). Its parameters are estimated by the maximum
likelihood method. Test of significance for model fit, regression parameters and the
value of including another predictor variable in a model are based on model deviance
(Crawley 1993). The concept of the coefficient of determination (R2) can be general-
ized to regression methods for which maximum likelihood is the criterion of fit (Cox
and Snell 1989). Automated stepwise procedures are often used for variable selection.
Consequently, logistic regression is by far the most accepted modelling procedure to
evaluate forest fire risk (Chou et al. 1993, Vega-Garcı́a et al. 1999, Diaz-Avalos et al.
2001).
Fire occurrence can be modelled as a binary response variable (present or
absent), whereas topography, vegetation type and meteorological data are pre-
dictive variables that can be related to forest fire risk. The weight of each
predictive variable is a function of its correlation with forest fire occurrence.
Predictive variables in a logistic model may be quantitative (e.g. elevation and
temperature) or categorical (e.g. vegetation type), static (e.g. slope) or dynamic
(e.g. vegetation condition). The possibility of representing the spatial and tem-
poral variability of the predictive variables enables us to use a geographic infor-
mation system (GIS) to display the model’s predictions in a geographic context.
A map of the probability of forest fire occurrence may be useful for designing
large-scale management strategies for fire prevention (Chou et al. 1993, Chuvieco
and Salas 1996, Chuvieco et al. 1999).
1.3 Remote sensing data

Fire statistics necessary for evaluating forest fire risks are often incomplete, and in
many cases they lack proper georeference of each event. In addition, meteorological
data such as temperature, precipitation and wind, in combination with fuel build-up,
are dynamic predictive variables useful for modelling the probability of forest fire
occurrence. However, weather stations in forest regions are often scarce, which
hinders the assessment of fire risk. Alternatively, information can be sought from
other sources such as climatic maps and remote sensing data.
For several years, remote sensing techniques have provided a reliable means of
estimating forest fire risk (López et al. 1991, Illera et al. 1996, González-Alonso
et al. 1997, Maselli et al. 2003). The information obtained from satellites
(e.g. NOAA-AVHRR, LANDSAT and MODIS), in combination with that
contained in thematic maps (e.g. topography and vegetation maps) or recorded
ForestSAT 2007 4995
at weather stations (e.g. temperature and precipitation), can be used as dynamic

predictive variables.
Data from the Advanced Very High Resolution Radiometer (AVHRR) operated
by NOAA have been used in several studies of forest fire (CONABIO 1999, Palacio-
Prieto et al. 1999, Stroppiana et al. 2000, Leblon 2001, Galindo et al. 2003). Their
success stems from their high temporal frequency (twice daily for a single satellite),
adequate spatial resolution (,1 km2), low cost, and multispectral character (one red
and four infrared bands). These features allow detection of active fires as hotspots
(Flasse and Ceccato 1996), the monitoring of vegetation condition from the
Normalized Difference Vegetation Index (NDVI) (Reed et al. 1994, Chuvieco et al.
2003) and estimation of land surface temperature (LST) (Price 1984, Coll and Caselles
1997), as well as evapotranspiration (Vidal et al. 1994).
The relationship between spectral data, particularly the NDVI, and fire probability
is indirect and based not only on a decrease in photosynthetic activity but also on
variations in canopy moisture and water stress (Maselli et al. 2003). Hence, the NDVI
decreases during the dry season, whereas dry conditions increase the probability of
forest fire (Burgan et al. 1998). Several studies have demonstrated a close relationship
between this decrease in NDVI and the occurrence of forest fires (López et al. 1991,
Illera et al. 1996, González-Alonso et al. 1997).
Because LST and evapotranspiration are good indicators of plant water stress, they
have also been included as factors of fire risk. Note, however, that not many studies
have included them in the NDVI analysis. During the major ENSO event of the winter
1997–1998, LST increased markedly, and there was a dramatic increase in forest fire
occurrence (Manzo-Delgado et al. 2004). It has also been shown that a decrease in the
evapotranspiration/potential-evapotranspiration ratio is related to an increase in
forest fires (Vidal et al. 1994).
The aim of this paper was to construct a statistical model to predict the probability
of forest fire, with a set of biological, topographical and meteorological variables,
both dynamic and static, obtained from NOAA-AVHRR images and thematic maps,
and test it in an area of central Mexico.
2. Test area
The test area is located in central Mexico (figure 2; 18 18¢ to 20 17¢ N and 98 27¢ to
100 30¢ W). It encompasses the States of Morelos and Mexico, as well as the Distrito
Federal, along with small portions of the surrounding states. Its surface area is 41 800
km2 and its elevation ranges from 400 to 5400 m. The area is mountainous with several
plateaus. Almost 95% of the annual precipitation falls during the summer months
(July to September) because of Trade Winds and hurricanes. In early autumn
(October), the Westerly Winds are drier than the Trade Winds, and the precipitation
gradually decreases. In winter (January to March) and early spring, 5% of the annual
precipitation occurs (Garcı́a 1981).
In the southern part of the study area, warm climates dominate at elevations below
1800 m, with annual temperatures of 18–26 C and precipitation from 700 to 1200 mm,
favouring the development of dry tropical forest and grassland; in the northeast part,
semi-dry climate prevails from 2200 to 2800 m, with annual temperatures of 14–18 C
and precipitation between 400 and 600 mm, enabling the growth of xerophytic scrub-
land and grasslands; in the central portions, at higher altitudes, temperate and cold
climates prevail, with annual temperatures in the range 5–12 C and precipitation
Figure 2. Test area (41 800 km2) in central Mexico. Contours at 1000 m intervals, solid lines.
State limits, broken lines.
between 600 and 1800 mm, enabling the growth of temperate forest and grasslands
(Garcı́a 1996, SEMARNAT 2001).
Most fires in central Mexico occur during the spring before the summer rains and
the greening of the vegetation (Rodrı́guez-Trejo and Fulé 2003). The fire season lasts
from January to May, and is most severe from April onwards (Rodrı́guez-Trejo
1996). Forest fires occur in dry tropical forest, grassland and temperate forest
(Rzedowski 1986), and most of them are caused by the cultural use of fire, tourism
and transport. Forest fire reports usually lack the exact location of each event, hence
there are no accurate spatial and temporal records.
In the study area, some attempts have been made to investigate fire behaviour
(based on fuel type, elevation and weather), particularly in the temperate forest. Some
of the dominant tree species, such as Pinus hartwegii, P. montezumae and P. teocote,
display fire-resistant traits (serotinous cones, thick bark and self-pruning)
(Rodriguez-Trejo and Fulé 2003), whereas other abundant species such as Abies
religiosa, Cupressus lindleyi and Quercus spp. are less tolerant to fire (Rzedowski
1986).
3. Extracting variables from NOAA-AVHRR 14 imagery

Daily NOAA-AVHRR 14 images, both night-time and afternoon-time, for the period
November–May from 1996 to 2000 were obtained at the TeraScan ground receiving
station in Mexico City (Álvarez et al. 1998). These were corrected for atmospheric and
radiometric effects (Lauritson and Nelson 1979) and georeferenced (Lambert
Conformal Conic projection), with a ground spatial resolution of 1.1 km. Images
were clipped (cropped) and registered to the extreme coordinates of the study area.
The clip covers 200 · 190 pixels. The response variables such as forest fires were
ForestSAT 2007 4997
extracted from the night-time images. The dynamic predictive variables such as
NDVI, LST and clouds were calculated using the afternoon-time images.
3.1 Response variables

Forest fire temperatures range from 570 to 1800 K (297–1527 C) and emit infrared
radiation accordingly (Robinson 1991, Kennedy et al. 1994); these emissions can be
detected in the middle infrared (AVHRR channel 3) and in the thermal infrared
(AVHRR channels 4 and 5). Although, in principle, channel 3 should be sufficient for
forest fire detection, there is in fact the possibility of confusion with other bodies that do
not emit in channel 3 as intensely as fires (apart from volcanoes); it is instead solar
reflection at channel 3 wavelengths, also known as sunglint, that can be mistaken as fires.
To avoid this possible confusion an additional criterion is used, involving channels 4 and
5. In this way the possible ambiguity in forest fire detection is circumvented.
Forest fires were detected on the night-time images as hotspots for each March–May
period from 1997 to 2000 by using the method of Cabrera et al. (1998), which is a
modified version of the contextual algorithm developed by Flasse and Ceccato (1996).
The performance of this fire detection contextual algorithm has been successfully
adjusted to several environments (Eva and Flasse 1996, Boles and Verbyla 2000). The
algorithm consists of three tests based on thermal channels T3 (3.55–3.93 mm) and T4
(10.3–11.3 mm). The first test identifies areas that could be masked (omitted). For
instance, clouds or water surface can be identified and subsequently masked from the
analyses. The second test selects hotspots that could be potential fires: if T3 – T4 8 C
(8 K) and T3 25 C (298 K). The third test confirms or denies that each pixel selected in
the second test is definitely a fire. This decision was made calculating both mean and
standard deviation for a context-window 5 · 5 pixels around the selected hotspots.
The hotspots of each image were transferred to the Arc/Info GIS (ESRI Inc. 1996),
constructing a cumulative fire map for each season (March–May). When several fires
occurred at the same pixel, only one event was recorded. During the 1997–2000
seasons, a total of nearly 1000 hotspots were detected in the study area: 40 in 1997,
730 in 1998, 76 in 1999, and 143 in 2000. More than 50% of the fires in each season
occurred in temperate forest and dry tropical forest, whereas the rest took place in
grasslands and agricultural areas. Most hotspots (65%) occurred on temperate forest
and between the 2000 and 3000 m elevation contour lines (figure 3).
The accuracy of hotspots was assessed by using the national report of fires by
municipality, elaborated by SEMARNAT: 85% (836) of hotspots were validated as
forest fires as they occurred in municipalities’ fire records; 15% (153) did not appear in
the fire records. These last were reviewed carefully, overlaying topographic and
vegetation maps: 50 hotspots were not accepted as forest fire as they occurred in
industrial spaces or thermal volcanic anomalies (e.g. Popocatépetl volcano); the rest
were considered as unrecorded forest fires. The final numbers of hotspot pixels
accepted as forest fires were: 34 in 1997, 704 in 1998, 66 in 1999, and 135 in 2000;
however, one or more forest fires could be included in one pixel.
The 804 fires recorded for the 1997–1999 seasons were used as the basis for
modelling the probability of forest fire occurrence. To build the model, concurrent
data from locations where no fires occurred were also required; these data points are
referred to as non-fire events. For each season, non-fire events were selected at
random: 45 for 1997, 740 for 1998, and 84 for 1999. These events could have been
located at any place, including areas that could not sustain a fire (e.g. water or urban
Figure 3. Forest fires (hotspots) detected from night-time NOAA-AVHRR 14 images, super-
posed on a vegetation/land-use map (National Forest Inventory 2000, scale 1:250 000;
SEMARNAT, 2001).
lands). The only condition for accepting the non-fire events was that they could not
occur in a hotspot pixel. The 135 fires and 132 non-fire of the 2000 season were
subsequently used to validate the model predictions.
3.2 Dynamic predictive variables

The NDVI, LST and cloud cover for the four dry months (November to February)
before each fire season, from 1996 to 2000, were calculated to develop the dynamic
predictive variables.
3.2.1 NVDI. This variable was calculated from the near-infrared (NIR; 0.72–1.10 mm)
and red (R: 0.58–0.68 mm) channels of NOAA-AVHRR 14, and the results were
rescaled to set a positive interval from 0 to 255, using NDVI = [(NIR – R)/
(NIR + R) + 1] · 127. Clouds were removed from the NDVI imagery by means of
the Saunders and Kriebel (1988) algorithm. NDVI-Maximum Value Composite
(NDVI-MVC) images were constructed for 10-day periods (Holben 1986).
Previous work (Manzo-Delgado and Meave 2003) on multitemporal analyses of
the NDVI for each class of vegetation in the same study area and period revealed
that each vegetation class followed a decreasing trend (figure 4) during the dry
season (November to April). During the 1998 dry season the decrease in NDVI
was more severe and there was an increase of forest fires (704), particularly in the
temperate forest. However, the 1997 dry season was very irregular and there was a
minimum number of fires (34). These data suggested that the trend in the NDVI
from the beginning of the dry season (November to February) up to the start of
the fire season (March) could be used as a variable to model the probability of
forest fire occurrence.
In this context five dynamic predictive variables were developed based on the
NDVI-MVC images. The first variable was defined as the difference between the
maximum (NDVImax) and the minimum (NDVImin) values recorded in the NDVI-
MVC images for the period from November to February: NDVInov–feb = NDVImax
– NDVImin. The second variable was the NDVInov–feb ratio, defined as the ratio
ForestSAT 2007 4999
Figure 4. NDVI trends for four classes of vegetation (temperate forests, dry tropical forest,
grasslands and xerophytic scrubland) in central Mexico from 1996 to 2000.
NDVImin/NDVImax. The three remaining variables were NDVIdec, NDVIjan and

NDVIfeb, corresponding to MVC-NDVI for December, January and February,
respectively.
3.2.2 LST. In this work, this parameter was obtained using the split-window algo-
rithm (Becker and Li 1990); emissivity values (e) were obtained from averages for soil
and vegetation classes (Rubio et al. 1997). Pixels with clouds were also masked. LST-
MVC images were also constructed for 10-day periods.
A previous study (Manzo-Delgado et al. 2004) using multitemporal analysis of LST
in the same region for the full dry season (November to April) revealed that the LST
followed an increasing trend from the beginning of the dry season to its end. During
the 1996–1997 and 1997–1998 (under the influence of ENSO) dry seasons, the
monthly LSTs ranged from 35 C to 46 C and from 33 C to 51 C, respectively; during
1998–1999 (under the influence of La Niña) and 1999–2000 the LSTs were lower,
ranging from 28 C to 47 C and from 28 C to 41 C, respectively. These results
suggested that the trend in the LST during the first four months of the dry season
could also be used as a variable to model the probability of forest fire occurrence
(figure 5). Consequently, four additional dynamic predictive variables were developed
based on the LST-MVC images (LSTnov–feb = LSTmax – LSTmin), as well as the
maximum LST registered in December (LSTdec), January (LSTjan) and February
(LSTfeb).
3.2.3 Cloud cover. During the dry period before each fire season, cloud cover
was also considered as a potential dynamic predictive variable of forest fires. It
was described in terms of the percentage occurrence of clouds (as detected in the
NOAA-AVHRR 14 images) during the four-month period (November to
February) before each fire season, by dividing the number of times that each
pixel was covered by clouds by the total number of images available for the
period.
Figure 5. Monthly composites from November to February of land surface temperatures

(LSTs) in central Mexico. These show the differences in the LST at the beginning and the end
previous to each fire season.
4. Static predictive variables

Fire initiation and propagation in forested areas is conditioned by the topography
(elevation, slope and terrain orientation), climate and vegetation type. In statistical
terms, these are considered as static predictive variables that represent an additional
value for forest fire risk. For instance, in areas of greater slope the fire propagation
speed is incremented; in the northern hemisphere, areas oriented to the south receive
greater insolation, inducing lower humidities and consequently a greater fire possibi-
lity; and the type of vegetation defines the combustion properties, which vary among
evergreen, deciduous, shrub or herbaceous species. Finally, meteorological conditions
(draught, rainfall and wind speed) play a fundamental role for fire risk evaluation in a
given region (Hernandez-Leal et al. 2004).
Hence, five additional static predictive variables were included: elevation, slope and
aspect (terrain orientation), mean annual precipitation, and vegetation type. The first
three were derived from a 90-m resolution digital elevation model (INEGI 2000),
mean annual precipitation was derived from the 1:1 000 000 climate maps by Garcı́a
(1996) and vegetation type from the 1:250 000 scale National Forest Inventory 2000
ForestSAT 2007 5001
(SEMARNAT 2001). The numerous vegetation/land-use categories in the National

Forest Inventory were reclassified into seven broad classes: temperate forest, dry
tropical forest, xerophytic scrubland, grassland, agricultural land, water, and urban
areas or areas devoid of vegetation (figure 3).
5. Construction of the GIS

All of the variables were converted into raster layers with the same spatial resolution
and projection as the NOAA-AVHRR images, and assembled into a GIS using
Idrisi32 software (Eastman 2000). The system included the response variables (forest
fires and non-fire events maps for each fire season from 1997 to 2000), and dynamic
(NDVInov–feb, LSTnov–feb, NDVInov–feb ratio, NDVIdec, NDVIjan, NDVIfeb, LSTdec,
LSTjan, LSTfeb and cloud cover) as well as static (elevation, slope, aspect, mean annual
precipitation and vegetation type) predictive variables. Having the data distributed in
layers of the GIS allows for a faster analysis including determining the relationships
between them. These analyses were organized in frequency distribution tables to
analyse their variability. A database was also constructed containing the characteristics
of each fire event, which was queried at various modelling stages. Finally, the results of
the model were also fed to the database and their spatial representation was derived
from the GIS.
6. Modelling the probability of fire occurrence

The logistic model is suitable for modelling the probability (P) of fire occurrence
(Chou et al. 1993, Vega-Garcı́a et al. 1999). Formally, the logistic model can be
expressed as:
eb0 þb1 x1 þb2 x2 þþbk xk

P¼ (2)
1 þ eb0 þb1 x1 þb2 x2 þþbk xk
where x1, x2, . . ., xk are k predictive variables and b1, b2, . . ., bk are their corresponding
coefficients. The model parameters (bi) were estimated from the sample of 804 fires
and 869 random non-fire points. For each event, values for all the predictive variables
were extracted from the GIS.
The logistic model was fitted using the Generalized Linear Interactive
Modelling (GLIM) program (Baker 1987), following a forward-stepwise variable
selection strategy, as described by Crawley (1993). The vegetation type was
considered as a categorical predictive variable with seven classes (figure 3). Each
predictive variable was tested in turn and only those variables causing a signifi-
cant (P , 0.05) decrease in the residual deviance were included in the final model.
Pierce and Ferrier (2000) have shown that, when fitting logistic regression models,
a 0.05 level is a good compromise between the risk of including too many
variables and the risk of omitting important ones if higher or lower significance
levels are used, respectively.
The variables aspect, NDVInov–feb, NDVInov–feb ratio, LSTnov–feb, NDVIdec,
NDVIfeb, LSTdec, LSTjan and cloud cover were not included in the final model
because they were not statistically significant; all other variables were significant (P
, 0.05) because they induced large-amplitude variations in the model’s response
(table 1). The model obtained was:
Table 1. Estimates of the model’s parameters.
Term Coefficient SE p
Constant -8.072000 1.652000

NDVIjan 0.0571400 0.006735 0.00001
LSTfeb -0.0758700 0.016150 0.00001
Vegetation type
Agriculture 0.0 0.0 -
Temperate forest 0.6346000 0.1764000 0.00300
Tropical dry forest -0.1304000 0.2737000 0.63400
Xerophytic scrubland 1.0670000 0.5626000 0.05800
Grassland 0.5963000 0.2186000 0.00600
Water -3.4540000 4.2880000 -
Devoid of vegetation -0.5423000 0.6402000 0.39800
Slope 0.0378000 0.0101500 0.00004
Elevation -0.0002941 0.0001393 0.00522
Precipitation 0.0006602 0.0003272 0.04288
Total deviance 2316.7000
df 1672
Residual deviance 1553.9
df 1661
logitP ¼ b0 þðb1 · NDVIjan Þ þ ðb2 · LSTfeb Þ

þ ðb3 · vegetationtypeÞ þ ðb4 · slopeÞ
þ ðb5 · elevationÞ þ ðb6 · precipitationÞ (3)
The model includes two dynamic variables and four static variables. The first dynamic
variable selected was NDVIjan, indicating that vegetation conditions at the middle of
the dry season (January) determine the fire risk of the March–May period better than
those of November–February, December or February. This appears to reflect the
availability of combustible materials that may ignite. The second dynamic variable
selected was LSTfeb, which seems to operate as a suitable precursor of the conditions
that will prevail in the March–May period, and is better than those more distant of
December or January. The selected static variables confirm that the topographic
(elevation and slope) characteristics of the region, its vegetation and the meteorolo-
gical conditions (precipitation) are factors that influence fire risk occurrence.
This model accounts for 33% of the total deviance of the data (table 1) and predicts
a higher probability of fire, particularly at those sites experiencing a higher LST in
February, either covered by grasslands or temperate forests. The predictive ability of
the model was assessed with reference to sensitivity, specificity and the receiver
operating characteristic (ROC) curve, as described below. These elements were esti-
mated using the STATA (Statistics/Data Analysis) statistical program (Kohler and
Kreauter 2005).
ROC curves are a statistical method used to visually analyse the predictive ability of
some models with such a logistic regression, which include a binary response variable.
An ROC curve is a graphical plot of the sensitivity versus (1 – specificity) for a binary
classifier system. Sensitivity and specificity are two statistical measures widely used to
summarize confusion tables: sensitivity measures how well the classifier predicts class 1,
ForestSAT 2007 5003
or ‘event’, cases, and is also known as the true positive rate; specificity is the prediction
success rate of class 0, or ‘non-event’, cases, also known as the true negative rate.
Sensitivity and specificity provide information about the misclassification of both
classes. The points on the ROC curve correspond to a series of confusions matrices
calculated at different cutoff values. Choosing 0.5 as the cutoff is reasonable only if
the prior probabilities of class 1 and class 0 are the same in the population of interest.
The analyst can choose a particular cutoff along the curve on the basis of both
sensitivity and specificity. The optimum cutoff is reached when the proportions in
both event types are the same. In the ROC plot, the diagonal line going from (0,0) to
(1,1) represents the lower limit of usefulness for a classification model. However, the
ROC curve of a perfectly predicting model would lie on the left side and the top of the
unit square graph. This information can be summarized by the area under the ROC
curve (the AUC): the model corresponding to the curve with the largest area is the best
in overall predictive ability. The AUC is equivalent to the non-parametric Wilcoxon/
Mann–Whitney test (Vayssieres et al. 2000).
We used the ROC curve of the model to identify its optimum cutoff level, that is the
probability value above which a fire would be predicted to occur. Using this cutoff
level, we calculated its sensitivity (Se), that is the proportion of fire occurrences that
are correctly predicted by the model, and specificity (Sp), that is the proportion of
non-fire events correctly predicted. The model’s optimum cutoff level was P = 0.525;
with this value, the model was able to correctly predict 79.73% of the fires (i.e.
Se = 0.7973) and 79.86% of the non-fire events (i.e. Sp = 0.7986) over the
1997–1999 fire seasons (table 2).
The predictive ability of the model was assessed by calculating the AUC and its 95%
confidence interval (CI). The ROC of a model with no predictive ability (i.e. a model
no better than chance at discriminating between fire and non-fire events) follows a
diagonal line going from point (0,0) to (1,1), overlying an area of only 0.5 units. By
contrast, a perfect model has a square ROC overlying a unit area (Vayssieres et al.
2000, Osborne et al. 2001). This model has an AUC of 0.8616 (figure 6), significantly
larger than 0.5 (95% CI 0.8435–0.8797).
To validate the model, its predictions for the 2000 fire season (March–May) were
calculated based on the dynamic variables NDVIjan and LSTfeb for January and
February 2000. Using the optimum cutoff level of P = 0.525, the total accuracy
found of 69.3% included 135 fires and 132 non-fires recorded in the period
March–May 2000.
7. Spatial distribution of the probability of fire occurrence

The model obtained was then applied to the proper data layers of the GIS for producing
maps delineating the spatial distribution of the probabilities of forest fire occurrence for
the years 1997, 1998, 1999 and 2000. Pixels with probabilities higher than 0.525 (the
Table 2. Predictive capacity of the model using the optimum cutoff level of P = 0.525.
Observed fire Observed non-fire Total
Predicted fire 641 175 816

Predicted non-fire 163 694 857
Total 804 869 1673
Correctly predicted proportion (%) 79.73 79.86 79.80
Figure 6. Receiver operating characteristic (ROC) curve for the model of probability of fire
occurrence. This model has an area under the curve of 0.8616, significantly greater than 0.5
(95% CI 0.8435–0.8797).
model’s optimum cutoff level) are those where a fire would be expected; pixels with
lower probabilities are those where a fire would not be expected. The probabilities were
divided into three uniform intervals below the model’s optimum cutoff level and three
above this value yielding six class intervals: lowest (, 0.209), lower (0.209–0.367), low
(0.367–0.525), high (0.525–0.683), higher (0.683–0.841) and highest (. 0.841). Italics are
used to denote these probability intervals. The sequence of maps shows a different
occurrence pattern of wild fires for each season (figure 7).
Figure 7 shows that the pixels with the highest probabilities of fire occurrence form
distinct clusters over the mountain ranges of the study area, at elevations above
1800 m asl, typically dominated by temperate forests and grasslands. The lowest
probabilities of fire are predicted at the plateaus and hills of the study area (elevation
, 2000 m asl), usually covered by agriculture, xerophytic scrubland or urban areas.
However, the diverse levels of probability of fire occurrence showed a different
proportion each year. During the 1998 and 1999 fire seasons, the areas from high
(0.525–0.683) to highest (. 0.841) probability of forest fire occurrence covered close
to 23% of the total; 6% more than in 1997 and 2000 (figure 8). Such an increment is
probably related to ENSO effects; in that season there were also an abnormal number
of fires. However, in the 1998 fire season, the area with highest probability (. 0.841)
covered about 6% of the total. This situation differed considerably from the other
seasons, where that class covered about 3% in 1997 and 1999, and 0.7% in 2000. These
data confirm that in the presence of ENSO the area of highest probability of forest fire
occurrence tends to increase compared to other seasons.
A complementary analysis allowed us to define the proportions of probability of fire
occurrence for each class of vegetation and land use in the 41 800 km2 of the study area
(table 3). Water bodies, such as dams and lakes, cover 0.5% of the surface. For any fire
registered in the water area, the model will predict the lowest probability of forest fire
occurrence every year. Urban areas devoid of vegetation cover 4.3% of the study area.
In this class every season contains a few fires (, 9); therefore, for each year a total
surface . 4.1% is predicted as a lowest probability, and a small proportion (0.2%) is
registered as lower and low probabilities. The agricultural land comprises 50.2% of the
region; each year more than 34% of that is predicted as a lowest probability; the rest is
ForestSAT 2007 5005
Figure 7. Probability of forest fire occurrence estimated from the logistic model during the
1997–2000 fire seasons, overlaid to records of actual forest fires (hotspots) detected on NOAA-
AVHRR 14 images (black squares). The 1997 map includes five test areas (numbers enclosed in
a circle) that were used to evaluate the spatial and temporal variability of the probability of fire
occurrence in the temperate forest from 1997 to 2000. The map for the 2000 fire season was
elaborated for model validation.
Figure 8. Percentage area of the study region with different levels of fire risk (as predicted by
the logistic model) during the 1997–2000 season.
registered from a lower to a higher category of fire occurrence. However, during 1998,
the higher probability was duplicated (0.7%) with respect to the previous year. The dry
tropical forest covers 9.5% of the study area. During 1997, 8.2% of this class is predicted
as a lowest probability. Its remaining surface (1.3%) is predicted as lower and low
intervals. These proportions were modified in the following two years: the lowest
probability increased to 5.5% in 1998 and to 4.3% in 1999; as a result, their comple-
mentary surface was predicted in a wider interval, which included high probabilities.
During 2000, the pattern was similar to 1997, although the lowest probability reached
6.8%. The xerophytic scrubland covers 1.6% of the region. Its maximum proportion
(0.8%–1.2%) of probability can be identified as a lowest or a lower category; the small
remaining proportion (, 0.3) is predicted as low or high levels. Grasslands cover 12% of
the study area; in this class a high fire probability can be expected. A large percentage of
grasslands are predicted with the lowest probability as they are usually located close to
agricultural areas; the remaining grasslands areas show probabilities of fire risk ranging
from lower to highest. During 1998 the highest probability of grasslands achieved 0.15%,
which is not common to other years. Temperate forest covers 21.9 % of the study area;
this is the region that sustains the largest number of fires and thus the whole spectrum of
probabilities is observed. The 2000 fire season anomalously showed a larger proportion
of the high probabilities. However, during 1998, the highest probability of fire occur-
rence for temperate forest reached 5.6%, which represents double the amount with
respect to 1997 or 1999. This information corroborates that in the presence of ENSO,
the highest probability tends to increase.
As we have shown, the percentage coverage of probability of fire occurrence for
various land use classes changed from 1997 to 2000. We also analysed the spatial and
temporal variability of fire occurrence in the temperate forest, on five test areas of 5 · 5
pixels, which were selected within the forest area; their location is shown in figure 7. To
illustrate this, two test areas out of the five analysed are shown in figure 9. In the first
period, from 1997 to 1998, the test areas presented a wide range of variability from 4%
to 60%, with an average of 7.2 and an SD of 5.8 pixels. In the next period, from 1998 to
1999, the test areas also showed a broad variation from 24% to 80%, with an average of
15.4 and a SD of 5.4 pixels. In the last period, from 1999 to 2000, the test areas presented
Table 3. Percentage coverage of seven land use classes in the study area with different levels of fire risk, as predicted by the logistic model
in the 1997–2000 period.
Predicted Urban/
probability of Temperate Dry tropical Xerophytic devoid of Total
Year fire occurrence Agriculture forest forest scrubland Grasslands Water vegetation (%)
1997 Lowest 41.4 1.4 8.2 1.2 6.6 0.5 4.2 63.4
Lower 5.0 2.6 1.1 0.3 2.8 0.0 0.1 11.9
Low 2.3 3.6 0.2 0.1 1.4 0.0 0.0 7.6
High 1.3 5.0 0.0 0.0 0.8 0.0 0.0 7.1
Higher 0.3 6.8 0.0 0.0 0.4 0.0 0.0 7.4
Highest 0.0 2.6 0.0 0.0 0.0 0.0 0.0 2.7
Total (%) 50.2 21.9 9.5 1.6 12.0 0.5 4.3 100.0
1998 Lowest 36.2 1.0 5.5 0.9 5.0 0.5 4.1 53.2
Lower 7.2 1.4 2.7 0.6 2.8 0.0 0.1 14.8
Low 3.7 2.3 1.0 0.1 2.1 0.0 0.1 9.2
High 2.4 3.9 0.3 0.0 1.2 0.0 0.0 7.8
Higher 0.7 7.8 0.0 0.0 0.8 0.0 0.0 9.3
Highest 0.0 5.6 0.0 0.0 0.2 0.0 0.0 5.7
Total (%) 50.2 21.9 9.5 1.6 12.0 0.5 4.3 100.0
ForestSAT 2007
1999 Lowest 34.0 0.4 4.3 0.4 3.6 0.5 4.1 47.3
Lower 9.1 1.4 3.5 0.9 3.4 0.0 0.1 18.4
Low 4.6 2.7 1.4 0.3 2.6 0.0 0.1 11.6
High 2.1 5.7 0.3 0.1 1.7 0.0 0.0 9.8
Higher 0.3 9.3 0.0 0.0 0.6 0.0 0.0 10.3
Highest 0.0 2.5 0.0 0.0 0.0 0.0 0.0 2.5
Total (%) 50.2 21.9 9.5 1.6 12.0 0.5 4.3 100.0
2000 Lowest 37.8 0.7 6.8 0.5 5.1 0.5 4.2 55.7
Lower 8.0 2.5 2.4 0.8 3.8 0.0 0.1 17.6
Low 3.4 4.5 0.4 0.2 1.9 0.0 0.0 10.5
High 0.8 7.1 0.0 0.1 0.9 0.0 0.0 8.9
Higher 0.1 6.3 0.0 0.0 0.3 0.0 0.0 6.6
Highest 0.0 0.8 0.0 0.0 0.0 0.0 0.0 0.8
Total (%) 50.2 21.9 9.5 1.6 12.0 0.5 4.3 100.0
5007
Figure 9. Test areas to analyse the variability of the probability of fire occurrence in the
temperate forest from 1997 to 2000.
a 20–40% variation, with an average of 8.4 and an SD of 5.6 pixels. The remaining three
areas show similar variations. These data confirm that the probability of forest fire risk
within the temperate forest undergoes significant changes between years. The large
variability of probabilities of fire risk between the first two periods corroborates the
effects of both ENSO and La Niña. However, the smaller variations shown in the final
period suggest that, in the absence of anomalous meteorological phenomena, the
probability of fire risk would return to a normal state, but corroborating this assertion
will require a considerably longer period of observations.
8. Conclusions
The model constructed here adequately describes the occurrence pattern of wildfires
recorded in the 1997, 1998 and 1999 seasons, based on hotspots detected from
NOAA-AVHRR imagery and predictive variables that can be easily measured or
derived from available maps or satellite imagery (such as climate, topography, vege-
tation type, NDVI and LST). The model is also able to accurately predict the
occurrence of fires during the following year (2000).
Hotspots detected from NOAA-AVHRR imagery offered a solid basis for model-
ling the probability of forest fire occurrence in the test area of central Mexico. The
model’s accuracy was similar to that of models based on historical records of forest
fires (Chou et al. 1993, Vega-Garcı́a et al. 1999). This suggests that similar fire risk
maps constructed from satellite images could help to avoid potentially catastrophic
events in other regions or countries where the lack of forest fire reports restricts the
predictability and understanding of fire risk, when proper parameters are incorpo-
rated into the model.
ForestSAT 2007 5009
The use of dynamic variables such as NDVI and LST, also derived from NOAA-
AVHRR imagery, allows the model to predict, within a spatial context, the prob-
ability of forest fire occurrence just prior to each fire season. Our results confirm that
these variables are important indicators of forest fire risk (Vidal et al. 1994, González-
Alonso et al. 1997, Sebastián-López et al. 2002, Maselli et al. 2003). Although the LST
has not been used to the same extent as the NDVI, its potential application to other
types of fire risk research should be further explored.
The model includes precipitation and vegetation as two static variables, which were
derived from actual thematic maps. Because of their dynamic character, these variables
have to be updated periodically. However, where fast processes of deforestation and
change of land use occur, the inclusion of current vegetation maps is necessary. The
results in the test area yielded acceptable results, and we consider that as the period of
observations is extended, the accuracy of the predictions will tend to improve.
The modest input requirements of the model (six variables readily derived from
satellite imagery or thematic maps) allow forest fire risk predictions to be easily
calculated, for example during the weeks previous to the onset of the forest fire
season. The probability method described here can be readily applied to any region
where the static and dynamic variables can be similarly defined. Furthermore, the
model of fire occurrence can be expanded to include additional variables that may be
relevant to fire behaviour and management in other regions.
Acknowledgements
We thank Jorge A. Meave for thoroughly reviewing an earlier version of the
manuscript. L.M.-D. was supported by Consejo Nacional de Ciencia y
Tecnologı́a (CONACYT). We thank the GIS and Remote Sensing Laboratory of
the Instituto de Geografı́a, Universidad Nacional Autónoma de Mexico for pro-
viding material support. Special thanks to Ann Grant for suggestions and correc-
tions to the English text.
References
ÁLVAREZ, R., GÓMEZ, G., MACÍAS, L. and SALMERÓN, O., 1998, The AVHRR receiving station at
the Institute of Geography, UNAM, in Mexico. In Proceedings of the 7th Annual
International TeraScan Users Conference, National Center for Atmospheric Research,
Boulder, CO (San Diego, CA: SeaSpace Corp.).
BAKER, R.J., 1987, GLIM 3.77 Reference Manual, 2nd edn (Oxford: Numerical Algorithms
Group).
BECKER, F. and LI, Z.L., 1990, Towards a local split window method over land surface.
International Journal of Remote Sensing, 11, pp. 369–393.
BOLES, S.H. and VERBYLA, D.L., 2000, Comparison of three AVHRR-based fire detection
algorithms for interior Alaska. Remote Sensing of Environment, 72, pp. 1–16.
BOND, W.J. and VAN WILEN, B.W., 1996, Fire and Plants (London: Chapman & Hall).
BURGAN, R., CLAVER, R.W. and CLAVER, J.M., 1998, Fuel models and fire potential from
satellite and surface observation. International Journal of Wildland Fire, 8, pp. 159–170.
CABRERA, A.L., GÓMEZ-RODRÍGUEZ, G. and ÁLVAREZ, R., 1998, Method for monitoring forest
fires through AVHRR image processing and digital elevation models in the central
region of the Mexican Republic (Método empleado en el monitoreo de incendios
forestales mediante el procesamiento de imágenes AVHRR y modelos digitales de
terreno en la zona centro de la República Mexicana). In Abstracts IX National
Symposium on Remote Sensing and Spatial Information Systems SELPER-Mexico,
Zacatecas, October 1998 (Resúmenes de la Novena Reunión Nacional Selper-Mexico.

Zacatecas).
CHOU, Y.H., MINNICH, R.A. and CHASE, R.A., 1993, Mapping probability of fire occurrence in
San Jacinto Mountains, California, USA. Environmental Management, 17, pp.
129–140.
CHRISTENSEN, N.L., 1993, Fire regimes and ecosystem dynamics. In Fire in the Environment: The
Ecological, Atmospheric, and Climatic Importance of Vegetation Fires, J.P. Crutzen and
J.D. Goldammer (Eds), pp. 233–244 (New York: John Wiley & Sons).
CHUVIECO, E., AGUADO, I., COCERO, D. and RIAÑO, D., 2003. Design of an empirical index to
estimate fuel moisture content from NOAA-AVHRR images in forest fire danger
studies. International Journal of Remote Sensing, 24, pp. 1621–1637.
CHUVIECO, E. and SALAS, J., 1996, Mapping the spatial distribution of forest fire danger using
GIS. International Journal of Geographical Information Systems, 10, pp. 333–345.
CHUVIECO, E., SALAS, F.J., CARVACHO, L. and RODRÍGUEZ-SILVA, F., 1999, Integrated fire risk
mapping. In Remote Sensing of Large Wildfires in the European Mediterranean Basin, E.
Chuvieco (Ed.), pp. 61–100 (Berlin: Springer-Verlag).
COCHRANE, M.A., ALENCAR, A., SCHULZE, M.D., SOUZA, C.M., Jr., LEFEBVRE, P., NEPSTAD, D.C.
and DAVIDSON, E.A., 1999, Positive feedbacks in the fire dynamic of closed canopy
tropical forest. Science, 248, pp. 1832–1835.
COLL, C. and CASELLES, V., 1997, A split-window algorithm for land surface temperature from
Advanced Very High Resolution Radiometer data: validation and algorithm compar-
ison. Journal of Geophysical Research, 102, pp. 16697–16713.
CONABIO, 1999, Monitoring hot spots (fires) (Monitoreo de puntos de calor (incendios)).
National Commission for the Knowledge and use of Biodiversity. Available online at:
www.conabio.gob.mx.
COX, D.R. and SNELL, E.J., 1989, The Analysis of Binary Data, 2nd edn (London: Chapman & Hall).
CRAWLEY, M., 1993, GLIM for Ecologists (Oxford: Blackwell Scientific Publications).
DÍAZ-AVALOS, C., PETERSON, D.L., ALVARADO, C.E., FERGUSON, S.A. and BESAG, J.E., 2001,
Space–time modeling of lightning-caused forest fires in the Blue Mountains, Oregon.
Canadian Journal of Forest Research, 31, pp. 1579–1593.
EASTMAN, J.A., 2000, Idrisi for Windows. User’s Guide. Version 132.02 (Worcester, MA: Clark
Labs, Clark University).
ESRI INC., 1996, Unix Workstation Arc/Info, Version 7.0.1 for Sun Solaris (Redland, CA:
Environment Systems Research Institute).
EVA, H. and FLASSE, S., 1996, Contextual and multi-threshold algorithms for regional active fire
detection with AVHRR data. Remote Sensing Reviews, 14, pp. 333–351.
FLASSE, S.P. and CECCATO, P., 1996, A contextual algorithm for AVHRR fire detection.
GALINDO, I., LÓPEZ-PÉREZ, P. and EVANGELISTA-SALAZAR, M., 2003, Real-time AVHRR forest fire
detection in Mexico (1998–2000). International Journal of Remote Sensing, 24, pp. 9–22.
GARCÍA, E., 1981, Notes on Climatology (Apuntes de Climatologı́a) (México, DF: Universidad
Nacional Autónoma de México).
GARCÍA, E., 1996, Charts (Isohyets) and Mean Annual Temperature (Isotherms), Scale
1:1 000 000 (Cartas de precipitación total anual (Isoyetas) y temperatura media anual
(Isotermas), escala 1:1 000 000) (México, DF: CONABIO).
GOLDAMMER, J.G., ROADS, J., HEILMAN, W.E., RIBEAU, A., ENGLEFIELD, P., DOTZAUER, H. and
STOCKS, B.J., 2003, Early warning of wildland fire. In Second International Conference
on Early Warning, Bonn, Germany. Available online at: www.fire.uni-freiburg.de/
course/meeting/Blueprint%20Wildland%20Fire-Final.pdf.
GONZÁLEZ-ALONSO, F., CUEVAS, J.M., CASANOVA, J.L., CALLE, A. and ILLERA, P., 1997, A forest
fire risk assessment using NOAA-AVHRR images in the Valencia area, Eastern Spain.
ForestSAT 2007 5011
HERNÁNDEZ-LEAL, P.A., GONZÁLEZ-CALVO, A. and ARBELO, M., 2004, Use of NOAA-AVHRR

data in fire risk maps (Utilización de datos NOAA-AVHRR en mapas de riesgo de
incendio). In XI Latin America Symposium on Remote Sensing and Spatial Information
Systems-SELPER (Memorias del XI Simposio Latinoamericano de Percepción
Remota), Santiago, Chile.
HOLBEN, B.N., 1986, Characteristics of maximum-value composite images from temporal
AVHRR date. International Journal of Remote Sensing, 7, pp. 1417–1434.
ILLERA, P., FERNÁNDEZ, A. and DELGADO, J.A., 1996, Temporal evolution of the NDVI as an
indicator of forest fire danger. International Journal of Remote Sensing, 17, pp.
1093–1105.
INEGI, 2000, Digital Terrain Model Scale 1:250 000 (Modelo Digital del Terreno. Escala 1:250
000). Charts: Ciudad de México, Morelia, Cuernavaca y Chilpancingo (Aguascalientes:
Instituto Nacional de Estadı́stica, Geografı́a e Informática, INEGI).
KENNEDY, P.J., BELWARD, A.D. and GREGOIRE, J.-M., 1994, An improved approach to fire
monitoring in West Africa using AVHRR data. International Journal of Remote
Sensing, 15, pp. 2235–2255.
KLEINBAUM, D.G. and KLEIN, M., 2002, Logistic Regression: A Self-learning Text (New York:
Springer-Verlag).
KOHLER, U. and KREAUTER, F., 2005, Data Analysis using STATA (College Station, TX: Stata
Press).
LAURITSON, L., NELSON, G. and PORTO, F.W., 1979, Data extraction and calibration of TIROS-
N/NOAA radiometers. NOAA Technical Memorandum NESS-107 (Washington, DC:
NOAA).
LEBLON, B., 2001, Forest wildfire hazard monitoring using remote sensing: a review. Remote
Sensing Reviews, 20, pp. 1–57.
LEE, B.S., ALEXANDER, M.E., HAWKES, B.C., LYNHAM, T.J., STOCKS, B.J. and ENGLEFIELD, P.,
2002, Information systems in support of wildland fire management decision making in
Canada. Computers and Electronics in Agriculture, 37, pp. 185–198.
LEIGHTON, M. and WIRAWAN, N., 1986, Catastrophic drought and fire in Borneo tropical rain
forest associated with the 1982–83 El Nino Southern Oscillations event. In Tropical
Rain Forest and the World Atmosphere, G. Ponce (Ed.), pp. 75–102 (Washington, DC:
Association for the Advancement of Science).
LÓPEZ, S., GONZÁLEZ, F., LLOP, R. and CUEVAS, J.M., 1991, An evaluation of the utility of
NOAA AVHRR images for monitoring forest fire risk in Spain. International Journal of
Remote Sensing, 12, pp. 1841–1851.
MANZO-DELGADO, L., AGUIRRE-GÓMEZ, R. and ÁLVAREZ, R., 2004, Multitemporal analysis of land
surface temperature using NOAA-AVHRR: preliminary relationships between climatic
anomalies and forest fires. International Journal of Remote Sensing, 25, pp. 4417–4423.
MANZO-DELGADO, L. and MEAVE, J.A., 2003, Vegetation seen from space: foliar phenology
through remote sensing (La vegetación vista desde el espacio: la fenologı́a foliar a través
de la percepción remota). Ciencia, Journal of the Mexican Academy of Science (Ciencia,
Revista de la Academia Mexicana de Ciencia), 54, pp. 18–28.
MASELLI, F., ROMANELLI, S., BOTTAI, L. and ZIPOLY, G., 2003, Use of NOAA-AVHRR NDVI
images for the estimation of dynamic fire risk in Mediterranean areas. Remote Sensing
of Environment, 86, pp. 187–197.
MCPHADEN, M.J., 1999, Genesis and evolution of the 1997–98 El Niño. Science, 283, pp. 950–954.
O’BRIEN, L., 1992, Introducing Quantitative Geography. Measurement, Methods and Generalized
Linear Models (London: Routledge).
OSBORNE, P.E., ALONSO, J.C. and BRYANT, R.G., 2001, Modelling landscape-scale habitat use
using GIS and remote sensing: a case study with great bustards. Journal of Applied
Ecology, 38, pp. 458–471.
PALACIO-PRIETO, J.L., LUNA-GONZÁLEZ, L. and MACÍAS-MORALES, L., 1999, Fire detection in
Mexico using AVHRR imagery (1998 season) (Detección de incendios en México
utilizando imágenes AVHRR (temporada 1998)). Investigaciones Geográficas, Boletı́n

del Instituto de Geografı́a, 38, pp. 7–14.
PICKETT, S.T.A., KOLASA, J., ARMESTON, J.J. and COLLINS, S.L., 1989, The ecological concept of
disturbance and its expression at various hierarchical levels. Oikos, 54, pp. 129–136.
PIERCE, J. and FERRIER, S., 2000, An evaluation of alternative algorithms for fitting species
distribution models using logistic regression. Ecological Modelling, 128, pp. 127–147.
PRICE, J.C., 1984, Land surface temperature measurement from the split window channels of
the NOAA 7 Advanced Very High Resolution Radiometer. Journal of Geophysical
Research, 89, pp. 7231–7237.
REED, B.C., BROWN, J.F., VANDERZEE, D., LOVELAND, T.R., MERCHANT, J.W. and OHLEN, D.O.,
1994, Measuring phenological variability from satellite imagery. Journal of Vegetation
Science, 5, pp. 703–714.
ROBINSON, J.M., 1991, Fire from space: global fire evaluation using infrared remote sensing.
RODRÍGUEZ-TREJO, D.A., 1996, Forest Fires (Incendios Forestales) (México, DF: Universidad
Autónoma Chapingo – Mundi Prensa).
RODRÍGUEZ-TREJO, D.A. and FULÉ, P.Z., 2003, Fire ecology of Mexican pines and a fire
management proposal. International Journal of Wildland Fire, 12, pp. 23–37.

RODRÍGUEZ-TREJO, D.A. and PYNE, S.J., 1998, Mexican fires of 1998. International Forest Fires
News, 20, pp. 61–63.
ROWELL, A. and MOORE, P.F., 2000, Global Review of Forest Fires (Gland, Switzerland: WWF/
IUCN).
RUBIO, E., CASELLES, V. and BADENAS, C., 1997, Emissivity measurements of several soils and
vegetation types in the 8–14 mm wave band: analysis of two field methods. Remote
Sensing of Environment, 59, pp. 490–521.
RZEDOWSKI, J., 1986, Vegetation of Mexico (Vegetación de México) (Mexico, DF: Limusa).
SAUNDERS, R.W. and KRIEBEL, K.T., 1988, An improved method for detecting clear sky and
cloudy radiances from AVHRR data. International Journal of Remote Sensing, 9, pp.
123–150.
SEBASTIÁN-LÓPEZ, A., SAN-MIGUEL-AYANZ, J. and BURGAN, R.E., 2002, Integration of satellite
sensor data, fuel type maps and meteorological observations for evaluation of forest fire
risk at the pan-European scale. International Journal of Remote Sensing, 23, pp.
2713–2719.
SEMARNAP, 1998, Forest Fires in Mexico (Los incendios Forestales en México: 1998).
(Mexico, DF: Secretarı́a de Medio Ambiente, Recursos Naturales y Pesca,
SEMARNAP).
SEMARNAT, 2001, National Forest Inventory 2000, scale 1:250 000. (Inventario Forestal
Nacional 2000, escala 1:250 000). Charts: Ciudad de México, Morelia, Cuernavaca y
Chilpancingo (Mexico, DF: Secretarı́a de Medio Ambiente y Recursos Naturales,
SEMARNAT).
SEMARNAT, 2002, Report on the Environment’s Status in Mexico. Summary of Environmental
Statistics (Informe de la Situación del Medio Ambiente en México. Compendio de
estadı́sticas ambientales) (Mexico, DF: Secretarı́a de Medio Ambiente y Recursos
Naturales, SEMARNAT).
STOCKS, B.J., LAWSON, B.D., ALEXANDER, M.E., VAN WAGNER, C.E., MCALPINE, R.S., LYNHAM,
T.J. and DUBÉ, D.E., 1989, The Canadian Forest Fire Danger Rating System: an
overview. The Forestry Chronicle, 65, pp. 450–457.
STROPPIANA, D., PINNOCK, D. and GREGOIRE, J.-M., 2000, The global fire product: daily fire
occurrence from April 1992 to December 1993 derived from NOAA AVHRR data.
TIMMERMANN, A., OBERHUBER, J., BACHER, A., ESCH, M., LATIF, M. and ROECKNER, E., 1999,
Increased El Niño frequency in a climate model forced by future greenhouse warming.
Nature, 398, pp. 694–697.
ForestSAT 2007 5013
VAYSSIERES, M.P., PLANT, R.E. and ALLEN-DIAZ, B.H., 2000, Classification trees: an alternative
non-parametric approach for predicting species distributions. Journal of Vegetation
Science, 11, pp. 679–694.
VEGA-GARCÍA, C., WOODARD, P.M., TITUS, S.J., ADAMOWICZ, W.L. and LEE, B.S., 1999, Two
models for the prediction of forest fires in Whitecourse forest, Canada (Dos modelos
para la predicción de incendios forestales en Whitecourt forest, Canada). Investigaciones
Agrarias: Sistemas y Recursos Forestales, 8, pp. 5–24.
VIDAL, A., PINGLO, F., DURAND, H., DEVAUX-ROX, C. and MAILLET, A., 1994, Evaluation of a
temporal fire risk index in Mediterranean forests from NOAA thermal IR. Remote
Sensing of Environment, 49, pp. 296–303.
VIEGAS, X., BOVIO, G., FERREIRA, A., NOSENZO, A. and SOL, B., 2000, Comparative study of
various methods of fire danger evaluation in southern Europe. International Journal of
Wildland Fire, 9, pp. 235–246.
WOLTER, K. and TIMLIN, M.S., 1998, Measuring the strength of ENSO event: how does 1997/98
rank? Weather, 53, pp. 315–324.

5i 2009 Manzo Delgado Etal IJRS

Uploaded by

Document Information

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

5i 2009 Manzo Delgado Etal IJRS

Uploaded by

Copyright:

Available Formats

This article was downloaded by: [Inst De Geografia]

On: 13 November 2009

International Journal of Remote Sensing

Assessment of seasonal forest fire risk using NOAA-AVHRR: a case study

PLEASE SCROLL DOWN FOR ARTICLE

Full terms and conditions of use: http://www.informaworld.com/terms-and-conditions-of-access.pdf

Assessment of seasonal forest fire risk using NOAA-AVHRR: a case

L. MANZO-DELGADO*†, S. SÁNCHEZ-COLÓN‡ and R. ÁLVAREZ§

*Corresponding author. Email: llmanzo@igg.unam.mx

1.1 Current estimates

1.2 Logistic model

1.3 Remote sensing data

at weather stations (e.g. temperature and precipitation), can be used as dynamic

3. Extracting variables from NOAA-AVHRR 14 imagery

3.1 Response variables

3.2 Dynamic predictive variables

NDVImin/NDVImax. The three remaining variables were NDVIdec, NDVIjan and

Figure 5. Monthly composites from November to February of land surface temperatures

4. Static predictive variables

(SEMARNAT 2001). The numerous vegetation/land-use categories in the National

5. Construction of the GIS

6. Modelling the probability of fire occurrence

eb0 þb1 x1 þb2 x2 þþbk xk

Table 1. Estimates of the model’s parameters.

Constant -8.072000 1.652000

logitP ¼ b0 þðb1 · NDVIjan Þ þ ðb2 · LSTfeb Þ

7. Spatial distribution of the probability of fire occurrence

Observed fire Observed non-fire Total

Predicted fire 641 175 816

the logistic model) during the 1997–2000 season.

Zacatecas, October 1998 (Resúmenes de la Novena Reunión Nacional Selper-Mexico.

HERNÁNDEZ-LEAL, P.A., GONZÁLEZ-CALVO, A. and ARBELO, M., 2004, Use of NOAA-AVHRR

utilizando imágenes AVHRR (temporada 1998)). Investigaciones Geográficas, Boletı́n

management proposal. International Journal of Wildland Fire, 12, pp. 23–37.

You might also like

eb0 þb1 x1 þb2 x2 þþbk xk