Design and Development of Forest Fire Management System

Signal & Image Processing : An International Journal (SIPIJ) Vol.2, No.4, December 2011
DOI : 10.5121/sipij.2011.2407 73
DESIGN AND DEVELOPMENT OF FOREST FIRE
MANAGEMENT SYSTEM
Dr. S. Sridhar1,
Annam Zulfigar2
and Paramathma Senguttuvan3
1
Associate Professor, Department of Information Science and Technology, Anna
University, Chennai, India
ssridhar@annauniv.edu
2,3
Department of Information Science and Technology, Anna University, Chennai, India
ABSTRACT
Forest fire is one of those natural disasters that have been causing huge destruction in terms of loss of
vegetation, animals and hence affects the economy. Image segmentation techniques have been applied on
satellite images of forest fire to extract fire object and some data mining techniques have been used for
predicting the spread of forest fire. This paper proposes a novel approach to isolation of fire region using
time-sequenced images, classifying fire images from non-fire images, predicting its movement and
estimating the area burnt. Once the images are enhanced, the fire region is segmented out. Feature
extraction provides the necessary inputs for classification of images as fire and non-fire images. Linear
regression is used to predict the movement of forest fire to facilitate better evacuation strategy. Burnt area
is calculated from the difference image. This work is helpful in drafting evacuation strategies quickly by
predicting the movement of forest fire and facilitates the kick-off of rehabilitation activities by identifying
and assessing the burnt area.
KEYWORDS
Forest fire management, Image segmentation, Classification, Forest fire movement prediction, Burnt area
calculation
1. INTRODUCTION
Forest fire poses a huge challenge to the human community by destroying vegetation and animals
on a large scale within a short span of time. Forest fires are generally started by lightning, but also
by human negligence, and can burn thousands of square kilometers. Forest fires are caused by the
drying out of branches and leaves, and therefore become highly flammable. Satellite images
provide sufficient amount of forest fire images in frequent intervals.
With the advancement in remote sensing technologies, satellites are able to facilitate study of fire
dynamics along with capturing of weather data. Meteosat Second Generation satellites observe
the earth continuously and send images including those of forest fires to the ground station [1].
Spatial image mining of these images can help us understand forest fire better and predict its
behavior.
Image Mining is focused on extracting patterns, implicit knowledge, image data relationship or
patterns which are not explicitly found in the images from databases or collections of images.
Some of the methods used to gather knowledge are: image retrieval, data mining, image

74
processing and artificial intelligence [2]. The satellite images help us find out the hotspots caused
by forest fire [3].
1.1. Review of literature
Preprocessing of satellite images prior to image segmentation is essential. Images may have noise
which can be detected and removed during the preprocessing step. Also, by enhancing edges of
the input images, the dynamic range of chosen features is reduced.
Region-based segmentation extracts a specific region from an image based on an initial seed. This
method examines neighboring pixels of initial “seed points” and determines whether the pixel
neighbors should be added to the region. The process is iterated on until the region of interest is
extracted. From the segmented region, features are extracted and used for classification.
The features of the segmented region (Average Intensity of the image, Average pixel range,
Number of white pixels in segmented image, green plane average, Entropy, NDVI – Normalized
Difference Vegetation Index) are extracted and used for classification.
Classification refers to an algorithmic procedure for assigning a given piece of input data into one
of a given number of categories or classes. It involves grouping data into classes based on some
measure of inherent similarity [4]. In spatial classification, the attributes of the neighboring
objects also influence the class membership. Hence, neighborhood factor needs to be included in
our calculations for classification. In classification, the users first define the classes and provide a
training set which includes the input data along with the classes associated with it. Based on the
training set, the classification rules are inferred [5]. These rules are applied on the test dataset. K-
nearest neighbor (KNN) is used in this work.
To evaluate the classification accuracy, a standard method called confusion matrix is used in
remote sensing. It contains information about actual and predicted classifications done by a
classification system. Performance of such systems is commonly evaluated using the data in the
matrix. Each time the user gives a new positive or negative training example, the posterior
probabilities are updated [6].
With the combination of statistics and image processing techniques, it is possible to predict the
direction of the forest fire movement. Linear regression helps to do the prediction. The variables
used for prediction are time, location, spread, intensity, NDVI, wind direction and wind speed.
This prediction will help the fire managers to take necessary actions to prevent the further spread
and loss.
To assess the fire affected area, the pre- and post-disaster images of the same location are needed.
Image differencing is a process of subtracting two different timed images of the same location
pixel by pixel to create the difference image [7]. After finding the fire affected region, it is
possible to find out the total area from it. The assessment of the burnt region gives the
approximate loss. This will help the forest department to do the necessary plan for rehabilitation
work like reseeding the vegetation in the fire affected area.

2. PROPOSED ARCHITECTURE
In this paper, a combination of two systems
Prediction and Assessment System
segmentation and classification system
and classification of images. Region based segmentation
identify whether a particular image has the fire affected area or not
extracted from the fire objects to help in classification. Then,
classified based on the features extracted
movement of forest fire to help quicker evacuation decisions and calculates the burnt area for
facilitating rehabilitation efforts.
Before doing the segmentation, the images should be enhanced to get better segmentation results.
Image enhancement refers to accentuation or sharpening of image features such as edged,
boundaries or contrast to make it more useful for analysis and
increase the information content in the data but only increases the dynamic range of the chosen
features so that they can be detected easily [8]. Image enhancement includes contrast
manipulation, noise reduction, edge sharpe
image enhancement enables easy extraction of fire objects.
K-nearest neighbor algorithm (k-
examples in the feature space. The trainin
Image Processing : An International Journal (SIPIJ) Vol.2, No.4, December 2011
RCHITECTURE
a combination of two systems - Segmentation and Classification system and
Prediction and Assessment System - is proposed. The architecture is presented in figure 1.
lassification system consists of region based segmentation, feature extraction
images. Region based segmentation segments the forest fire objects to
identify whether a particular image has the fire affected area or not. Relevant features are
extracted from the fire objects to help in classification. Then, the fire and non-fire image
based on the features extracted. The prediction and assessment system predicts the
Figure 1. System architecture
ke it more useful for analysis and display. This process does not
manipulation, noise reduction, edge sharpening, filtering, pseudocoloring and so on. In this work,
image enhancement enables easy extraction of fire objects.
-NN) is a method for classifying objects based on closest training
examples in the feature space. The training phase of the algorithm consists only of storing the
75
Segmentation and Classification system and
The architecture is presented in figure 1. The
segmentation, feature extraction
orest fire objects to
. Relevant features are
fire images are
ystem predicts the
display. This process does not
ning, filtering, pseudocoloring and so on. In this work,
NN) is a method for classifying objects based on closest training
g phase of the algorithm consists only of storing the

feature vectors and class labels of the training samples.
defined constant denoting the number of clusters present in the data, and an unlabelled sample is
classified by assigning the label which is most frequent among the
Euclidean distance is used as the distance met
appropriate clusters, the cluster centers are recalculated. Instance classification and cluster center
computation are done in iterations until the cluster centers stabilize.
class of unknown sample is predicted based on the nearest training instance [10].
affected areas are typically ash colored and a sample image (Figure2) depicts the affected area.
2.1. Preprocessing of forest fire im
The images used in this work are simulated and captured in real time manner. The study area of
the forest is manually fired and the images were shot in ten seconds interval. Hundred images of
640x480 resolution have been captured and used in this work
Median filter is used to remove the outliers (noise) while maintaining the sharpness of the image.
Image adjustment is done on the noise
various steps involved are:
i. Read the satellite image
ii. Convert into grayscale images
iii. Check for speckle noises
iv. Apply two dimensional median filter
v. Check whether the satellite images lack contrast
vi. Apply contrast stretching technique to enhance the image
Figure 3 depicts the images at each stage of preprocessing along
feature vectors and class labels of the training samples. In the classification phase,
classified by assigning the label which is most frequent among the k training samples nearest to it.
Euclidean distance is used as the distance metric [9]. Once the instances are placed in the
computation are done in iterations until the cluster centers stabilize. In instance based training, the
ss of unknown sample is predicted based on the nearest training instance [10].
Figure 2. Fire affected area
of forest fire images
640x480 resolution have been captured and used in this work.
Image adjustment is done on the noise-removed image to increase the contrast of the image
Read the satellite images
rt into grayscale images
Check for speckle noises
wo dimensional median filter
Check whether the satellite images lack contrast
Apply contrast stretching technique to enhance the image
Figure 3 depicts the images at each stage of preprocessing along with their intensity profiles.
76
In the classification phase, k is a user-
training samples nearest to it.
ric [9]. Once the instances are placed in the
In instance based training, the
ss of unknown sample is predicted based on the nearest training instance [10]. Forest fire
removed image to increase the contrast of the image. The
with their intensity profiles.

77
Figure 3. Top row - Filtered image and its intensity profile, Bottom row – Contrast enhanced
image and its intensity profile
2.2. Image Segmentation
The forest fire region is extracted out of the images using region growing algorithm. The
basic formulation of the region-growing algorithm is:
1) The addition of all sub-regions produces the entire image region
2) Every sub-region must be connected with other region, which means there is no isolated
region
3) The intersection of two different regions always provides a NULL value, which means
there regions must be disjoint
4) The predicate of all the sub-regions must be true
5) The predicate of two regions are different
Region growing algorithm works as follows: It starts with selecting a seed. Seed value can be a
specific gray level or color information. In this work, gray level (255) is used as seed. Fire pixels
are depicted as white pixels and hence this seed value is selected. After identifying the seed, the
neighboring pixels are examined to look for same characteristics. Such pixels are grouped into a
region. The growth continues iteratively until there is no more pixels to be grouped into the
region. A sample segmented image highlighting the fire object is shown in figure 4.
0 100 200 300 400 500 600 700 800
0
50
100
150
200
250
Distance along profile
0 100 200 300 400 500 600 700 800
0
50
100
150
200
250
300
Distance along profile

78
Figure 4. Segmented image
2.3. Feature Extraction
In order to proceed with classification, the features need to be extracted from the images. The
selected features are stored in an array for further processing. The features should uniquely
identify the image and should have lesser dimensionality to reduce the computational time of the
subsequent steps.
The following features are extracted from the set of fire and non-fire images:
i. Average Intensity of the image (avg_i)
ii. Average pixel range (avg_pix_range)
iii. Number of white pixels in segmented image (white_pix)
iv. The green plane average (green_plane_avg) from original image
v. Entropy
vi. NDVI – Normalized Difference Vegetation Index
Table 1 shows extracted feature set for forest fire images.

79
Table 1. Sample features extracted from forest fire images
Image Avg_i Avg_pi
x_range
White
_pix
Green_p
lane_avg
entropy NDVI Class
Img1 142.3986 90 2282 80.247 6.329 0.2822 Fire
Img2 110.139 80 5183 86.2305 6.1932 0.271 Fire
Img3 171.8275 125 2803 115.7552 6.0759 0.314 Fire
Img4 114.6921 75 3157 65.1263 6.3928 0.3012 Fire
Img5 82.0199 110 3418 68.9521 6.9123 0.2621 Fire
Img6 122.4886 100 5355 71.9144 6.7136 0.281 Fire
Img7 110.1312 90 5185 86.2286 6.1936 0.31 Fire
Img8 86.1123 80 4128 69.3686 6.6611 0.308 Fire
Img9 110.1294 85 5178 86.228 6.1937 0.252 Fire
Img10 119.3577 75 2703 72.2676 6.1597 0.2802 Fire
Img11 84.8903 82 0 86.088 6.2579 0 Nonfire
Img12 82.5026 80 0 88.8006 6.1247 0 Nonfire
Img13 79.9604 40 280 94.2923 7.1543 0.11 Nonfire
Img14 102.7115 60 1517 114.8943 7.4993 0.1772 Nonfire
Img15 34.8521 30 227 49.0992 5.2054 0.023 Nonfire
Img16 115.2215 60 14 128.1115 7.2118 0.0012 Nonfire
Img17 94.9123 90 0 101.9464 5.7189 0 Nonfire
Img18 121.1446 105 4227 124.6954 6.7622 0.198 Nonfire
Img19 55.7109 20 639 59.757 6.9749 0.16 Nonfire
Img20 84.5778 80 4016 96.8306 6.8563 0.167 Nonfire
2.4. k-Nearest Neighbor Classification
Once the feature set is extracted, the classifier is trained with a labelled feature set in which the
class of every feature set is provided. The feature set is d-dimensional meaning that the features
are in a d-dimensional space where d is the number of attributes. Here fire and non-fire are the
two classes used.

80
Once the system is trained, it can identify a new test data. Based on the number of nearest
neighbors, the test data is classified. Ties between classes are arbitrarily broken.
The steps for classification are as given below:
i. Training: The feature set that has all the samples is given to the system
ii. Choose a value for k
iii. Give the number of classes
iv. Input the test feature set
v. Find the distance between test set with the classes
vi. Classify the test set based on the k nearest neighbors
Here k=3. Hence based on 3 nearest neighbors, the class of the test sample is decided. Table 2
shows the results of KNN method.
Table 2. K-nearest neighbor classification
Row Predicted
class
Actual
class
Prob.
For fire
Actual # of
nearest
neighbors
1 Fire Fire 1 3
2 Fire Fire 1 3
3 Fire Fire 1 3
4 Fire Fire 1 3
5 Fire Fire 1 3
6 Fire Fire 1 3
7 Fire Fire 1 3
8 Fire Fire 1 3
9 Fire Fire 1 3
10 Fire Fire 1 3
11 Nonfire Nonfire 1 3

81
Row Predicted
class
Actual
class
Prob.
For fire
Actual # of
nearest
neighbors
14 Fire Nonfire 1 3
18 Fire Nonfire 1 3
K value can be varied to get better results. Using k values as 1 or 2 will lead to misclassification.
It was observed that above k=5, the number of errors stays at 6.
2.5. Fire movement prediction
First the high intensity location of fire (x,y) is identified. If there is dense vegetation, then it is
more likely that fire will move in that direction. The spread of fire in x and y directions along
with the intensity is also identified. Normalized Difference Vegetation Index (NDVI) is
calculated. Wind direction and wind speed are also noted. Table 3 shows a sample list of the
variable values extracted for regression.
Table 3. Extracted values for regression
X Y SIG X SIG Y I WD WS NDVI
280 465 230 453 242 170 7.0 0.28
204 448 226 435 248 165 7.3 0.31
197 433 224 423 236 172 7.5 0.27
201 419 226 409 247 175 7.4 0.29
198 405 214 395 242 171 7.3 0.26
203 385 219 378 246 178 7.1 0.29
221 375 235 369 235 173 7.2 0.28
227 368 241 363 249 169 7.4 0.31
232 361 243 354 251 170 7.6 0.31

X Y SIG X SIG Y
234 353 247 346
The movement of forest fire can be done by applying linear regression analysis on these variables
as given below:
xt+4 ~ α0 + α1 xt+ α2 σx,t+ σ3It+ σ4
yt+4 ~ β0 + β 1 xt+ β 2 σx,t+ β 3It+ β
where α and β are calculated from the data, I is the intensity,
2.6. Burnt area calculation
Burnt region is identified by image differencing technique which uses
taken before the fire and after the fire.
region. The identified region gives an idea about the loss of vegetation and animals that liv
that area. After finding the burnt region, the total loss can be calculated.
done by Poly area function provided by MATLAB. The burnt area is represented as a set of
vertices stored in vectors X and Y. Passing these to
burnt area is returned. Burnt area is calculated in terms of number of pixels.
Figures 5 and 6 depict the Pre-disaster and Post
pre-disaster image does not have any fire in
area and fire indications.
SIG Y I WD WS NDVI
346 244 173 7.3 0.28
WDt + σ5 WSt+ σ6 NDVI
4 WDt + β 5 WSt+ β 6 NDVI
are calculated from the data, I is the intensity, σ is the spread in x and y directions.
image differencing technique which uses the difference of images
taken before the fire and after the fire. The result of image differencing gives the burnt forest
region. The identified region gives an idea about the loss of vegetation and animals that liv
that area. After finding the burnt region, the total loss can be calculated. Burnt area calculation is
function provided by MATLAB. The burnt area is represented as a set of
vertices stored in vectors X and Y. Passing these to Polyarea(X,Y) function of MATLAB, the
Burnt area is calculated in terms of number of pixels.
disaster and Post-disaster images respectively. It is clear that the
disaster image does not have any fire indications while the post disaster image has the burnt
Figure 5. Pre-disaster image
82
is the spread in x and y directions.
the difference of images
The result of image differencing gives the burnt forest
region. The identified region gives an idea about the loss of vegetation and animals that lived in
Burnt area calculation is
function provided by MATLAB. The burnt area is represented as a set of
TLAB, the
It is clear that the
dications while the post disaster image has the burnt

83
Figure 6. Post-disaster image
The identified burnt region by using image differencing is shown in figure 7.
Figure 7. Identified burnt region

84
3. RESULTS AND DISCUSSIONS
3.1 Dataset used
A set of 100 (50 fire and 50 non-fire) images is considered for classification purpose. The images
are JPEG images with a standard resolution of 640x480. The images and the features are stored in
MySQL database. Extracted features include average intensity of the image, mean of all pixels,
number of white pixels, average value of green plane, entropy and NDVI.
A set of 20 time-sequenced images is considered for prediction. The variables considered for
prediction include location, spread, intensity, wind speed, wind direction and NDVI.
3.2 K-nearest neighbor classification
K value is set as 3. The extracted feature set is used for training. The classes in the
training set are shown in table 4.
Table 4. Classes in training data
No. of Class 2
Class 1 Fire
Class 2 Non-fire
The probability of success is calculated for every tuple. Here, 3 nearest neighbors are
considered. The confusion matrix for KNN classification is shown in table 5.
Table 5. KNN Classification – Confusion matrix
Classification confusion matrix
Predicted Class
Actual class Fire Non-fire
Fire 50 0
Non-fire 2 48
Out of the 100 images taken, all the fire samples were classified correctly but 2 non-fire images
were misclassified as fire images. This was due to cloud interpretation of non-fire image. % error
was calculated and table 6 lists this data.
Table 6. Error report of k-nearest neighbor classification
Error Report
Class No. of cases No. of Errors %Error
Fire 50 0 0.00

85
Non-fire 50 2 4.00
Overall 100 2 2.00
Accuracy calculation for KNN is calculated by identifying the number of misclassifications. The
table 7 shows the accuracy calculations for KNN.
Table 7. KNN Classification – Accuracy calculation
Total no. of images taken: 100
Class True classification False classification
Fire 50 (TP) 0 (FP)
Non-fire 48 (TN) 2 (FN)
Accuracy = (TP+TN) = 98/100 = 0.98 = 98%
(TP+FP+TN+FN)
3.3 Forest fire movement prediction
Prediction is done by regression algorithm. Initially, the regression algorithm finds the coefficient
values. Then, by using the regression equation, the fire movement direction is predicted.
Table 8 shows the predicted direction. The current value in the table shows the actual location of
the forest fire. The predicted value shows the next fire movement location in a certain time
interval. Based on the predicted direction of forest fire movement, the fire fighters can move their
resources.
Table 8. Predicted forest fire movement
Row x-axis y-axis
Predicted
value
Current
value
Residual Predicted
value
Current
value
Residual
1 223.7829021 280 56.21709789 461.7380501 465 3.261949936
2 207.4202112 204 -3.42021122 445.3256581 448 2.674341901
3 210.6747962 197 -13.6747962 433.1966163 433 -0.19661625
4 215.4495014 201 -14.4495014 417.6810101 419 1.318989909
5 211.7962699 198 -13.7962699 40.7229581 405 -0.72295808
6 213.9468891 203 -10.9468891 383.4197075 385 1.580292531

86
7 221.5452448 221 -0.54524475 375.4020998 375 -0.42809977
8 221.6208616 227 5.379138354 371.6275453 368 -3.62754529
9 222.0492577 232 9.950742324 363.5926829 361 -2.59268286
10 237.0876243 234 -3.08762427 354.071155 353 -1.071155
3.4 Burnt Area Assessment
The burnt area is identified by image differencing. This gives the burnt region of the forest as
resultant. Then, the total lost area can be calculated. In this work, the burnt area is calculated as
29583.8777 pixels.
4. CONCLUSIONS
In this work, a novel approach to find forest fire and classify fire/non-fire images, predict fire
movement and assess burnt area. This work can help fire fighters to take necessary action to
control the spread of the fire as well as trigger off evacuation from the region. This also supports
assessment of the burnt area after the disaster so that rehabilitation activities can be initiated
immediately.
The overview of the experiments conducted and the summary of the results obtained are
highlighted below:
• A set of 100 (50 fire and 50 non-fire) images is considered for classification purpose.
Features used for classification include average intensity of the image, mean of all pixels,
number of white pixels, average value of green plane, entropy and NDVI.
• Classification of these images produced 98% accurate results. The misclassification was
due to non-fire images wrongly classified as fire images.
• A set of 20 time-sequenced images is considered for prediction. The variables considered
for prediction include location, spread, intensity, wind speed, wind direction and NDVI.
• Prediction results show a misprediction of 1%-7% in most of the cases with a maximum
of 20% in one of the cases.
As an extension to this work, this system can be made available on the net for fire rescue teams.
This can be developed as an expert system to find fire indications earlier to warn fire managers in
advance. Other areas of research include applying better segmentation techniques to extract the
forest fire object and improved prediction methods to get reduced misprediction rates.
REFERENCES
[1] Rajasekar Umamaheshwaran, Wietske Bijker, and Alfred Stein, “Image Mining for Modeling of
Forest Fires from Meteosat Images”, IEEE Transactions on Geoscience and Remote Sensing, Vol. 45,
No.1, pp. 246-253, January 2007.
[2] P.Stanchev, “Using image mining for image retrieval”, IASTED Conference Computer Science and
Technology, Cancum, Mexico, pp. 21-218, 2003.

[3] Stefan Voigt, Thomas Kemper, Torsten Reidlinger, Ralph Kiefl, Klaas Scholte and Harald Mehl,
“Satellite Image Analysis for Disaster and Crisis
Geoscience and Remote Sensing, Vol. 45, No.6, pp. 1520
[4] V. Karasova, “Spatial data mining as a tool for improving geographical models”, Masters thesis,
Dept. Surveying, Helsinki Univ. Tech, France, pp. 1
[5] Richard J.Roiger and Michael W. Geatz, “Data Mining: A tutorial based primer”, Pearson Education,
First Indian reprint, 2005.
[6] Herbert Daschiel and Mihai Datcu, “Information mining in remote sensing image
evaluation”, IEEE Transaction on Geoscience and Remote Sensing, vol. 43, no.1, pp. 188
[7] S. Anderson, M.Taylor, P. Sutton, M.Steinberg, “Assessment of urban land use cover changes in
Cuidad Victoria, Tamaulipas, Mexico”, www.
[8] Anil. K. Jain, “Fundamentals of Digital Image Processing”, Pearson Education, Second Indian reprint,
2004.
[9] Margaret H. Dunham and S.Sridhar, “Data Mining: Introductory and advanced topics”, Pearson
Education, Second Edition, 2007.
[10] Ian H. Witten & Eibe Frank, “Data Mining: Practical machine learning tools and techniques”,
Elsevier, Second Edition, 2006.
Authors
[1] Dr. S.Sridhar is an Associate Professor in Department of
Information Science and Technology, Anna University,
Chennai. His research interest is in the areas of Image
Processing, Medical Imaging and Data Mining algorithms.
[2] Annam Zulfigar is pursuing her Masters of Engineering in
Multimedia Technology at College of Engineering Guindy,
Anna University, Chennai. Her research interest is in Image
Processing.
[3] Paramathma Senguttuvan holds a Masters degree in
Multimedia Technology and his interest areas are Image
Processing.
Stefan Voigt, Thomas Kemper, Torsten Reidlinger, Ralph Kiefl, Klaas Scholte and Harald Mehl,
for Disaster and Crisis-Management Support”, IEEE Transactions on
nd Remote Sensing, Vol. 45, No.6, pp. 1520-1528, June 2007.
V. Karasova, “Spatial data mining as a tool for improving geographical models”, Masters thesis,
Helsinki Univ. Tech, France, pp. 1-65, 2005.
Richard J.Roiger and Michael W. Geatz, “Data Mining: A tutorial based primer”, Pearson Education,
Herbert Daschiel and Mihai Datcu, “Information mining in remote sensing image archives: System
evaluation”, IEEE Transaction on Geoscience and Remote Sensing, vol. 43, no.1, pp. 188
S. Anderson, M.Taylor, P. Sutton, M.Steinberg, “Assessment of urban land use cover changes in
Cuidad Victoria, Tamaulipas, Mexico”, www.isprs.org.
Anil. K. Jain, “Fundamentals of Digital Image Processing”, Pearson Education, Second Indian reprint,
Margaret H. Dunham and S.Sridhar, “Data Mining: Introductory and advanced topics”, Pearson
Education, Second Edition, 2007.
Ian H. Witten & Eibe Frank, “Data Mining: Practical machine learning tools and techniques”,
Elsevier, Second Edition, 2006.
Dr. S.Sridhar is an Associate Professor in Department of
Information Science and Technology, Anna University,
Chennai. His research interest is in the areas of Image
Processing, Medical Imaging and Data Mining algorithms.
Annam Zulfigar is pursuing her Masters of Engineering in
Multimedia Technology at College of Engineering Guindy,
ai. Her research interest is in Image
Paramathma Senguttuvan holds a Masters degree in
Multimedia Technology and his interest areas are Image
87
Stefan Voigt, Thomas Kemper, Torsten Reidlinger, Ralph Kiefl, Klaas Scholte and Harald Mehl,
Management Support”, IEEE Transactions on
V. Karasova, “Spatial data mining as a tool for improving geographical models”, Masters thesis,
Richard J.Roiger and Michael W. Geatz, “Data Mining: A tutorial based primer”, Pearson Education,
archives: System
evaluation”, IEEE Transaction on Geoscience and Remote Sensing, vol. 43, no.1, pp. 188-199, 2005.
S. Anderson, M.Taylor, P. Sutton, M.Steinberg, “Assessment of urban land use cover changes in
Anil. K. Jain, “Fundamentals of Digital Image Processing”, Pearson Education, Second Indian reprint,
Margaret H. Dunham and S.Sridhar, “Data Mining: Introductory and advanced topics”, Pearson
Ian H. Witten & Eibe Frank, “Data Mining: Practical machine learning tools and techniques”,

Design and Development of Forest Fire Management System

More Related Content

What's hot (18)

Similar to Design and Development of Forest Fire Management System (20)

Recently uploaded (20)

Design and Development of Forest Fire Management System