The FEDAP arson pattern recognition strategy shown in Figures i and. 2 uses adaptive decision rules based upon both labeled and new pattern samples.
Application of Pattern Recognition in Arson Investigation DAVID J. ICOVE
University of Maryland H E U B E R T J. CRISMAN
Prince George's County Fire Department The authors discuss the application of pattern recognition methods to the development of techniques for aiding the arson investigator in determining trends in incendiary and suspicious fires. E C E N T study 1revealed that destructive fires in 1972 in the United A RStates took the lives of 11,900 and caused $2.93 billion in property damage. The number of incendiary and suspicious fires alone totaled 84,200 and resulted in nearly $286 million in property loss, representing increases of 16.8 and 22.6 percent respectively over the previous year. The detection and investigation of arson-related fires remains within the jurisdiction of fire prevention bureaus in most cities. In large metropolitan areas, such as Prince George's County Maryland where this study originated, heavy caseloads prevent the detailed analysis of all investigations. In an effort to provide the comprehensive engineering analysis of these investigations, a set of algorithmic computer programs have been developed for the fire service using pattern recognition methods. The emergence of FEDAP, 2 the Fire Engineering Data Analysis P;ogram, paralleled the implementation by Prince George's County Fire Department of MIRS, ~ the Modular Information Reporting System. Designed by the National Bureau of Standards, MIRS is intended to become a nationwide standard for the computerized encoding of emergency incidents. PATTERN
RECOGNITION
APPROACH
Successful implementation of a supervised arson pattern recognition (APR) system for the prediction and prevention of incendiary crimes4 requires a set of valid decision rules. Using labeled pattern samples of known arson trends, highly reliable decision rules can be derived. The FEDAP arson pattern recognition strategy shown in Figures i and 2 uses adaptive decision rules based upon both labeled and new pattern samples. The arson investigator serves as an integral part of the learning 35
36
Fire T e c h n o l o g y
Figure 1. A n arson inuestigator controlling the arson recognition system at a remote computer terminal.
system, and the performance of the system should improve with time.
PATTERN SAMPLES Each pattern sample consists of an individual arson-related fire numerically encoded by M I R S and defined by F E D A P as a partitioned row matrix. The i'th fire is represented by a seven-dimensional pattern space
such that the features describing the i'th fire include: G~I = Date of incident, G~2 = Location code, G~ -- Type o f incident,
_.~ NEW G. pATTERN SAMPLES
~
G,
LABELED
PATTERN sAMPLES
ARSON ][NVFST[GATOR
RULE ~ PmTTERN MODIFIQATION CLASSIFICATION 1 G, Figure 2. The F E D A P supervised arson pattern recognition system flou.,ehart.
Pattern Recognition
37
G~4 = Act or omission, G~5 = Time alarm received, G~6 = D a y of the week, and G~ = N u m b e r of alarms. A collection of m pattern samples or incidents is represented b y a partitioned matrix G = (G1G2...G~). P a t t e r n samples used in this study consist of 361 incendiary and suspicious fires collected from M I R S data for February through September 1973o 5 No labeled pattern samples based upon completed investigations were initially made available; however, several trends were later verified as valid labeled patterns.
PATTERN CLASSIFICATION The process of deriving decision rules for pattern spaces is called pattern classification. 6 Decision boundaries separate pattern classes through a specified decision rule. Since it is difficult to visualize patterns greater than three dimensions, pattern classifications are frequently handled in three, two or one dimension. One or more arson related fires occurring within a single location code is the decision rule for defining a data "cluster" C j, where the subscript is the M I R S location code of the cluster. Based upon a two-dimensional pattern space displaying the total number of clusters within a prescribed area (X2) versus the total number of incidents within an individual cluster (X~), several pattern classes become evident. For the 361 new pattern samples, four decision boundaries are selected as indicated in Figure 3. These decision boundaries define five pattern classes S , , S~., . ° . $ 5 as indicated with characteristic symbols. These new pattern samples represent the first training set used by F E D A P in the development of a valid set of decision rules. These characteristic pattern class symbols are used during feature selection. FEATURE SV.LECTION Feature selection is an acceptable technique for reducing the dimensionality of the G - - pattern space arson data. Often the process is multistage after several pattern classifications made by the arson investigator as illustrated previously in Figure 2. The Prince George's County Fire Department uses a map consisting of 185 major map grids, each containing 24 major grids. Using feature selection, cluster classification and the F E D A P algorithms,"- eleven adjacent major map grids have been isolated that contain 37 percent of the 361 total arson related fires. A two-dimensional display of the geographical location of these clusters is shown in Figure 4 using the I N M A P algorithm computer map. Notice that all of the five pattern classes are represented. Continuing the feature selection process, a better visual representation of the higher pattern classes is displayed in Figure 5 using the algorithm
Fire Technology
38 X2
185
I I
•
•
0
•
•
•
@
Q
,4 T 0 T A L
I 40
30 C L U S T E R S
20
10
0
+ I I I + I I I + I I I + I !
•
•
•
g
0
•
•
0
•CLASS.
CLASS $I
•
S 2
CLASS
CLASS•CLASS. Sa • $4
•
•
•
•
•
®
•
o
•
•
+
•
•
•
o
•
•
+.....
+. . . . .
+.....
4-. . . . .
• +.....
+ X1
3
4
5
6
7
8
•
•
+ .....
+. . . . .
l
2 TOTAL
55
•
•
•
I ~". . . . .
•
0
INCIDENTS
IN
CLUSTER
Figure 3. Determination of decision rules for p a t t e r n classes.
D E N M A P , a c o m p u t e r i z e d c o n t o u r d e n s i t y map. E a c h class p r e v i o u s l y d e f i n e d i n F i g u r e 3 a s a c h a r a c t e r i s t i c c l a s s s y m b o l is d i s p l a y e d i n a fivelevel grey scaled c o n t o u r map. T h e t w o h i g h e s t a d j a c e n t clusers, C33~ 4 a n d C41el, a r e c h o s e n f o r f u r t h e r f e a t u r e s e l e c t i o n s . W h e n d i s p l a y i n g t h e c l u s t e r s i n G - - s p a c e b y t w o dim e n s i o n s , s e v e r a l t r e n d s a p p e a r . F i g u r e 6 is t h e p a t t e r n s p a c e f o r m e d b y
I
1
[ I I
I
1 2 ~ 1
1
I
I I I ! I
T I I I I
1 2 1
2
]
7 1 ]
]. 3 1 3 2 a !
1 1 1
! T i [ ",.
1 4 4 1 2
1 a
]
5
1 5
2 1 ]
1
2 2
? 1 !
3 5
2 1
1
2
1 1 2 l I
T I Figure 4. T h e I N M A P algorithm c o m p u t e r map T d i s p l a y i n g datc~ clusters. I T
1
i
2 2 1
i
!
3 1 ]
i 1 4 1
5
T
! I
Pattern Recognition
39
ii!iii iiiiil!iiii!ii!iiii!iiili!ii!iiiiiii!ii!ilii!iiii!i!!!!!!!!!!!!!i ii!i i iiii!lili!i!i!!iiii!i!!!i!!ii!ii!!ii!i!!i!!iiliiiiii ii
ii~iiiiii!i~i:~ii~iig!~ii~iii~ff~ii~:~i~:ii~ii~i5~i!~ig~ii2:~i~!i.iii! !ii!i i iilii ! iliiiii!i!i~i!iii~ii!iiiiiiiii!i~iii!ii!iiiiiiiii i!!!i!ililiiiiii!iiiiiiiii!!!ill!!iiii!iiii!ii!iiiiiiiiiiiii!iii .o~ ~A~.r:.~:o:!i~7 s..'i'ff~ii.:-~!f'i~ff'" : :~ !?is!E~i ii!!'~;~J~s~?iis?Fsii~i~5~i!!!v!~iEii~!?z-??'i?ii-ii~ii~ iiiiii!!!!iii::ii!i!i!:.!::i::i2:ii!::iiii~!::.ii::i~iiiiiiiiiiiiiii~~!:~i!i ~ 441f! !!!f!f;ili~ ii~iiiiiiiii~i!~iiii~!i~iiiiiSi~i~:iiiiiiiiiiiii~ii~iiii~i~ii!i~iiiiiiii~ ~ ......................... :::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
::::::::::::::::::::::::::::::::::::: ,:;;::::::::: :;:: - ~ . ::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
~..,::::;::::::::::'....:55E!i!5!E
!ii~i!51iEEE!!!~iiiii!!!!~i!~i!i!i!iV~ii~iEi~;!~F~i::::::::::::::::::::::::::::::::::::::::::::::::::::
i!?!~iiiii!~i!~i!
",~ii i !ii!i i i !iiiiii!ii!~i~i!iii~:ii~i!~iiii~i!iii~i~?i~iiiiii~!i~!~ii!i~i!i~i~!~ii~i i iiiiiii!i!iiiiiiiiiiiiiiiiii!iii .~i~iiii~i!~i~iiii~i~ii!!ii!~i~!i~ii~!~ii!i~ii~ii~iiiii~i!i~ii~ii~i~iiiii~ii!ii~!!ii!~iiii~i~!~iiii~!i~!~ii!iiiiii~iii~i!i~iiiii~i~i!~iill)::!~!~i~iiiiiiiiiii~iiiiiiiiii!!iiiiiii!!~!
~ii~i~i.:!........... ?~!!~ii!i~~-::iii~ii!~i,=~.0~ i!i/~:.~;~;~;~!!!!!~!!!!!!~i l;i~!!i!1~ii~i~i~~:,i!i~i~i;i-;.................................... i!i~i;}!!!ii~!~!~!}~}~!!i?i}~i!}?~ii!~ii!~!ii~)~)iiii; ................................... ":!E[!i~!E~Ei?! !i~ ! ! i [ ' ~ ' ! . : .:!L"! i'!i!i~
i:~2;
ii!i:::::.:::::~Ei!E!.5::=:fi.~,.~
;
::
:: :::::::::: ::: ::::::::;:: ::
.: " ~ e
:
======================================
!:, i~i~iEi!i!E~i~::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::~..:.::.:.F
~
~
