usingageographicinformationsystemfor qualitative ......qualitative insights that users need. in...

Using a Geographic Information System forQualitative Spatial Reasoning about Trafficability

AbstractModern geographic information systems (GIS) arepowerful quantitative tools for geospatial reasoning. Weshow how a GIS can be used as a metric diagram to supportqualitative spatial reasoning . We illustrate the techniquewith the problem of reasoning about trafficability -- anentity's capability for movement through some terrain .Many common geospatial reasoning tasks rely heavily uponestimates of trafficability, and in some domains, off-roadtrafficability is of primary concern. Maps rarely provide theappropriate representation of space for such problems . Weshow how GIS technology can be usedas a metric diagramfrom which place vocabularies appropriate to the task canbe computed . We describe the domain theories that enablethe computing of these place vocabularies, and providemodels for qualitative trafficability reasoning based onthese descriptions . The results suggest that the power ofqualitative representations and reasoning can be brought toa variety of geospatial applications, riding on the progressmade by the commercial world in GIS software .

Introduction

Modern geographic information systems (GIS)provide powerful and useful facilities fordigitizing and mapping representations ofgeographic space . They are widely used forperforming complex transformations of this data,and for solving common geospatial problems,(e.g ., placement of structures under constraints) .However, most GIS computations arequantitative, relying on visualization tools andextensive user interactions to provide thequalitative insights that users need. In contrast,most human reasoning about geographic spaceappears to use a qualitative interpretation of that

QR99 Loch Awe, Scotland

CPT James J. Donlon'

Kenneth D. Forbus

Qualitative Reasoning GroupNorthwestern University

1890 Maple AvenueEvanston, IL, 60201, USA

donlonj@csl .carlisle .army.mil,forbus@ils .nwu.edu ;

space. We believe that qualitative spatialrepresentations can be useful for reasoning aboutgeospatial problems. The metric diagram/placevocabulary (MD/PV) model uses a combinationof quantitative and qualitative representations,decomposing space in task-specific ways into thekind of regions that are intuitive to human usersand useful for qualitative reasoning X1,6] . Ourkey observation is that GISs can be used asmetric diagrams . That is, they can directly answerquantitative queries (which is their standard use),and they can be used to construct placevocabularies, i .e ., task-dependent qualitativespatial representations, which is a novel use forsuch systems.

In this paper we use the task of geospatialreasoning for trafficability analysis problems toshow how GISs can be used in qualitative spatialreasoning. We begin with a brief description ofGIS systems. Then we describe how the GISdigital representations of terrain are combinedwith symbolic reasoning to automatically producequalitative descriptions of terrain. The nature oftrafficability problems is discussed next, followedby a description of our domain theory for solvingsuch problems . We outline the successful resultsobtained with Our implemented system, andpropose directions for future work .

i Current address: Knowledge Engineering Group, United States Army WarCollege, Center for Strategic Leadership, 650

Wright Avenue, Carlisle, PA, 17013, USA

62

Brief review: How GIS systems work

GIS systems enable the production and analysisof digital representations for geospatial problems .They accept geographic data from many sourcesand in many formats, and provide two- and three-dimensional visualizations of this data from amultitude of perspectives . They are used in manyproblems, such as mapping, land management,city planning, and vehicle routing. Althoughvariations exist, in general GIS systems encodetheir geospatial data and accompanying attributeinformation in a relational database .

GIS systems organize this information inlavers .Each layer contains geospatial information ofsome particular type with a consistent set ofspatial entities (points, arcs, or polygons) thatcover the region of interest . For example, it isnot unusual for a digital terrain data set to consistof polygon coverages that describe vegetatedareas, water bodies, etc., arc coverages for roads,rivers, sewer lines, etc ., and point coverages thatrepresent facilities, nodes in transportationnetworks, etc .

The properties used to distinguish regions withina layer are either symbolic labels (e.g .,vegetation-cropland vs . vegetation-woodland),numerical values, or numerical ranges . Whatdifferentiates a region in a GIS layer as distinct isthat all of the attribute values within this regionare uniform.

GIS as metric diagram

In the MD/PV model, the metric diagram playsthe role that diagrams do for human spatialreasoning : they provide concrete descriptionsfrom which we can easily answer spatialquestions . Humans routinely decomposecontinuous space into discrete regions accordingto criteria relevant to the desired reasoning .Everyday geographic reasoning provides manyexamples of this, e.g ., zones for urban


development, districts for government . Placevocabularies provide this kind of qualitativerepresentation. The computation of placevocabularies from metric diagrams models thehuman use of visual processing to extract/imposequalitative spatial distinctions on diagrams orscenes, and the ability to use combinations ofpurely qualitative and metric information inspatial reasoning .

The digital data and geometric processingprovided by GIS technology means they can beused directly as a metric diagram for geospatialreasoning . The mixture of symbolic and numericinformation in a GIS is similar in kind to therepresentations found in earlier metric diagrams(e.g ., [4,6]), but optimized for geospatialproblems . The library of geometric and databaseroutines provided with high-end systems are morethan adequate for implementing routines tohandle concrete, quantitative spatial questions,thus satisfying the first constraint on metricdiagrams .

The second service of metric diagrams, thecreation of place vocabularies, is partiallyprovided by the GIS ability to produce new layersthat quantize space into polygons and arcs . Thequestions that must be solved to apply a GIS to aqualitative geospatial domain areI . What place vocabulary(ies) are needed for the

task?2. How can they be computed from the available

data?The rest of this paper shows how we haveanswered these questions for an important classof geospatial reasoning, trafcability problems .

The problem : Trafficability

Trafficability is a measure of the capability forvehicular movement through some region . It is arelationship between some entity (capable ofmovement) and the area through which it moves.Whether an area is trafficable for the entity, andthe measure of how trafficable the area is for theentity, depend on the interaction between the

63

entity's mobility characteristics and the relevantterrain attributes . A simple example is planning acar trip, where one plots routes through anetwork of roads and reasons about the relativemerits of different routes (e .g ., total travel time,possible effects of bad weather, etc.) . A morecomplex example is plotting a route for a multi-vehicle safari through open country. While ourmethods are applicable to either on-road or off-road movements, we are particularly concernedwith trafficability reasoning for the more difficultcase of cross-country movement .

Some simple on-road trafficability problems arebest solved using known quantitative methods.For example, route finding and travel timeproblems in a fully specified road network can beeasily solved with built-in GIS algorithms for pathfinding and flow optimization . However, manyimportant trafficability problems do not lendthemselves to such solutions . Taking intoaccount weather-related factors or partialknowledge about the terrain make even on-roadtrafcability problems difficult using standardsolution methods. Off-road, cross-countrytrafficability problems are far more complex, andthere are no off-the-shelf tools that automaticallysolve them. Figuring out if an area can bereached with a particular vehicle and exploringmigration patterns are two examples of tasksrequiring off-road trafficability reasoning .

The particular focus of our work is to militarytrafficability problems -- specifically, thetrafficability of military vehicles for cross-countrymovement . Military estimates rely heavily on anunderstanding of the trafficability of terrain. Inthe military domain, qualitative trafficabilitydeterminations about regions of space arecombined to inform a variety of estimates, such asidentifying potential areas through which largeforces might move, or determining where a forcemight perform specific types of operations .

The basis for most military terrain analysis is mapdata and terrain analysis surveys. Use of maps asthe sole basis for trafficability estimates is


problematic because of the amount and variationof representations presented on a singletopographic map. It is also very difficult andlabor-intensive to integrate data gathered in fieldsurveys with the topographic map's representationof the terrain. To manage this informationoverload, the military's terrain analysis processesrely heavily on diagrams that identify places thathave uniform significance or characteristicsrelevant to the desired analysis . These diagrams

transparent sheets, registered to the underlyingmap. The resulting overlays are then usedextensively for producing further estimates. Theyalso provide a representation for communicatingabout the area of operations .

Two particular products of military terrainanalysis inspired the place vocabularies wecomputed. These are the complex factor overlay(CFO)18 1, and the combined obstacle overlay(COO)[9] .

Vegetatior

Soil factors

Slone

ComplexFactornverlav

Figure l. Complex Factor Overlay

A CFO is the cumulative partitioning of spaceaccording to criteria from several terrain analysiscategories . It is derived by creating overlays foreach category (i .e ., slope, soil factors, vegetation,surface materials, and hydrology), and thencombining them to identify the areas withhomogenous characteristics across all overlays .The result is a partitioning of space that is as

6 4

(overlays, in the military vernacular) areproduced by outlining the regions that areuniform for the relevant characteristics on

detailed as the available terrain data permits(Figure 1) . This overlay allows the analyst tocompute estimates of vehicle speeds for eachregion, based on vehicle capabilities and formulaethat relate those characteristics to the terrain inthe region .

In contrast, the COO is a characterization of theterrain's effects on vehicular movement,characterized in a much more general way.Terrain in each region is said to be unrestricted(U), restricted (R), or severely restricted (SR) formovement in the relevant context, e.g ., movementof a brigade equipped with tracked and wheeledvehicles . (Water regions are identifiedseparately.) Thus, the production of a COOresults in fewer individuated places, as adjacentU, R, and SR regions are aggregated (Figure 2) .

E~b em

~~rrl~'/lIY-/f

1 -41I%4!I

Hydrography

Vegetation

Slope

Etc. . .

CombinedObstacleOverlay

Figure 2. Combined Obstacle Overlay

The COO is primarily used for determining wheremilitary units are likely to be moving andconducting operations . The echelon of the unitunder consideration helps to determine what sizeareas of R and SR terrain are significant enoughto be included on the overlay .

Both the CFO and the COO are examples ofplace vocabularies made explicit in humanpractice . They are qualitative decompositions ofspace used to help solve particular classes oftasks. The qualitative reasoning challenge is toautomate the production of these descriptions .


Using a GIS to construct place vocabularies

Places are individuated by where some importantproperty or combination of properties is uniform.Each polygon in a GIS layer is defined to beuniform with respect to whatever property (orproperties) is associated with that layer. In asense the GIS already implements a placevocabulary for any digital terrain datarepresentation, albeit of an extreme form .

The key idea in using a GIS for qualitative spatialreasoning is to (a) figure out how to compute theproperties needed for a task from the availabledata and (b) derive a new layer that decomposesthe terrain in terms of uniformity in thoseproperties . The new layer constitutes a placevocabulary for that characterization of the terrain .For example, in the combined obstacle placevocabulary, the characterization is whether apiece of terrain is U, R, or SR, and those regionswhere this property is uniform correspond exactlyto the regions found in a Combined ObstacleOverlay .

To compute the properties required, wedeveloped a domain theory for trafficability,expressed in CML [l] augmented with KIF [10] .To create the new GIS layers corresponding tothe COO and CFO, we developed algorithms thatused the geometric processing library ofARC/INFO 115] to do the necessarycomputations . We first describe the domaintheories, and then we describe the GISprocedures .

Trafficability domain theoriesReasoning about vehicle trafficability requiresknowledge of terrain and its effects onmovement, vehicles and their capabilities to moveon and off road, and how terrain affects themovements of these vehicles over the terrain. Wedeveloped domain theories to address each ofthese three areas, as well as domain theories thatrepresent knowledge about how the digitalrepresentation of the terrain is encoded. We will

6 5

briefly describe each of these theories, andprovide some representative definitions andaxioms as illustration . The size of these domaintheories is indicated in Table l .

Domain #Theory PredicatesTrafficability 87Terrain 504Military 285VehiclesGISTotals :Table l : Size of domain theories

(defRelation trafficable (?P ?M) :_(and (movement-episode ?M)

(path ?P)(not (or (weather-denied ?P ?M)(exists (?S ?T ?V)

(and (segment-of ?S ?P)(terrain-of ?S ?T)(vehicle-of ?M ?V)(or (insuff-clearance ?P ?V)

(untraff-surface ?T ?V)

TrafficabilityTrafficability is a binary relationship between theterrain and the vehicle . This relationship can beanalyzed and represented either qualitatively orquantitatively, depending on the task .

Qualitative expressions of trafficability areexpressions of whether a vehicle is capable ofeffectively traveling through a region, and thequalitative effects the terrain will have on thatmovement . For example, the determination that aregion is trafficable for a movement of some kind(represented here as a movement episode [2J)relies on determining that no part of the path isuntrafficable .) The path of the movement mustprovide sufficient clearance for the physicalextent of the movement, and a suitable surface forthe movement of the vehicle :

I On the "negative polarity" of considering a segmenttrafficable : It is more natural and convenient to specifywhich conditions prevent movement, and assume theirnon-existence unless there is evidence to the contrary . Forexample, in the absence of proof of weather-denied, weshould assume it is false .


A region's surface is considered unsuitable if theslope, vegetation, soil conditions, or surfacematerials of the region are prohibitive, or if theregion is a body of open water?

(defRelation untrafficable-surface (?T ?V)._ (or (too-steep ?T ?V)

(too-vegetated ?T ?V)(weak-soil ?T)(obstructed-surface ?T)(dvalue (obstacle-category ?t) W)))

Further expanding the illustration, a region is toovegetated for the cross-country movement if thevegetation is both too closely spaced for thevehicle to drive around and too thick for thevehicle to override :

(defRelation too-vegetated (?T ?V)._ (and (vegetated ?T)

(and (< (stem-spacing ?T)(min-turning-radius ?V))

(> (stem-diameter ?T)(override-diameter ?V)))))

These axioms can be very useful for identifyingcritical relationships that constrain possiblemovements .

We also use compositional modeling techniques[3,5J to express qualitatively the effect the terrainin a region will have on a movement through it,such as the effects of vegetation :

(defmodelfragmentmovement-rate-vegetation-effects:participants ((M :type movement-episode)

(P :type path)):conditions ((possible-trajectory-of M P)):consequences((qprop+ (movement-rate M P) (SS P))(qprop- (movement-rate M P) (SD P))))

Where SS represents the distance between stemsor trunks in a vegetated area, and SD representsthe average diameter of those stems . This axiomgives us the basis for making qualitativejudgements such as, all other things being equal, a

We use the relationdvalue to specify the value of adiscrete variable . Similarly, we usenvalue to specifythe value of a continuous parameter . This allows us todisentangle the semantics of these relations from=,equal, eql, etc . for reasoning .

6 6

#Axioms164

#facts19

731 16388 0

151 251434 60

vehicle can travel faster through scrub than it canthrough dense woods. This is just the kind ofcommon-sense judgement that human analystsroutinely apply when interpreting descriptions ofterrain and making predictions about movementsover that terrain .

The effects of weather also play a significant rolein trafficability analysis . We can use similarmodels to account for these effects, enabling thedomain theory to take advantage of reasoningabout weather information, if available :

(defmodelfragmenttrafficability-weather-effects:participants ((M :type movement-episode)

(P :type path)):conditions ((possible-trajectory-of M:consequences((qprop- (movement-rate M P)

(visibility P))(qprop- (movement-rate M P)

(wetness P))(qprop- (movement-rate M P)

(accumulation-of rain(qprop- (movement-rate M P)

(accumulation-of snow

(deffunction ccm-speed-dry(?veh ?geo-area) :_

(* (F1/2 ?veh ?geo-area)(F3 ?veh ?geo-area)(F4D ?veh ?geo-area)(F5 ?veh ?geo-area)))


For problems that demand more quantitativetrafficability determinations, we draw from the U.S. Army§ cross-country movement analysistechnique for estimating vehicle speed overterrain regions[8] In practice, these places arederived from the CFO overlay described above.Vehicle speed is calculated for a given vehicle foreach type of region based on a series ofmathematical transformations. Cross-countrymovement speed (CCM) is calculated based onthe vehicle§ top speed, unconstrained, whichthen is degraded based on the cumulative effectsof slope (FI) and surface configuration (F2),vegetation (F3), soil effects (F4), and surfacematerials (P5) :

The variable ?geo-area is bound to a GIS polygon.Determining each of the above factors involves anumber of steps. For example, determining the

vegetation factor (F3) involves the combinationor comparison of six other factors .

One of thesefactors

is

Vl a,

a

factor

that

accounts

for

avehicle§ ability to drive around the trees in avegetated area :

(deffunction Vla (?veh ?geo-area) :=(cond ((<= (Vla-calc ?veh ?geo-area) 0) 0)

((<= (Vla-calc ?veh ?geo-area) 1) 1)(t (Vla-calc ?veh ?geo-area))))

(deffunction Vla-calc (?veh ?geo-area) :_(/ (- (SS ?geo-area)

(SD ?geo-area)(width ?veh))(min-turning-radius ?veh)(* 4 (width ?veh)))))

Although the reasoning involved in this estimateis largely quantitative, qualitative descriptions ofspace allow the reasoner to perform thesecalculations when the terrain data is missing orincomplete . Normally this calculation is notpossible if a value is missing for any one of theproperties that are involved in calculating thefactors . For example, while our GIS data setcontained the structure for the full range ofmilitary terrain attributes needed to determinethese factors (e.g ., stem spacing (SS) and stemdiameter (SD), above) the coverage rarelycontained values for these attributes . To fill inmissing information, our reasoner uses the namedcharacterization of the region, e.g . EB020(scrub/brush), to conduct default reasoning aboutthe missing quantities .

Finally, we formalized the characterization ofterrain as used in the combined obstacle overlay.Recall that in the COO terrain areas are identifiedas U, R, SR. The following are some of theaxioms that define severely restricted terrain:

(defrelation severely-restricted-terrain(?t ?v)

._ (and (dvalue (obstacle-category ?t) SR)(transportation-device--vehicle

?v)))

(forall (?t)(=> (and (landform ?t)

(or (>= (percent-slope ?t) 45)(and (< (SS ?t)

(meters 6))(>= (SD ?t)(meters .15))

(dvalue (obstacle-category ?t)

6 7

(forall (?t)(_> (and (waterbody ?t)

(or (> (water-current ?t)(feet-per-second 5))(water-depth ?t)(feet 4))(military-gap-width ?t)

(forall (?x)(=> (weak-soil ?x)(dvalue (obstacle-category ?t) SR)))

TerrainThe trafficability domain theory relies on adetailed representation of terrain, including :

Physiography - Representations of landformations and their presentation (e.g ., hills,valleys, and dunes) . Characteristics used todescribe landforms include : Percent of slope,slope intercept frequency, elevation, major typesof landforms .

Hydrography -- Water and drainage associatedfeatures (e .g ., river, stream, lake, sabkhat) .Characteristics include : Water depth, watercurrent, banks

Vegetation - Presence of plant life (i .e ., forest,swamp, grassland), with characteristics including :vegetative roughness, stem spacing, stemdiameter, vegetation type, canopy closure, tree-height

Surface Materials/Soils - Soil type andclassification, and a classification of surfaceroughness types, including : Area soil type,primary soil partition, secondary soilcharacteristic, rating cone index, surfaceroughness

We have represented terrain in sufficient detail toconduct the kind of trafficability reasoningdescribed above . This involves a generalcharacterization of terrain features :

OR99 Loch Awe, Scotland

(defrelation vegetated (?x)(and (geographic-area ?x)

(<= (VR ?x) 1)(>_ (SS ?x) 0)(>_ (SD ?x) 0)))

(sparsely-vegetated :self)(default-dvalue

(vegetation-type :self)B2)

(vr-scrub :self)))

it involves default reasoning for missinginformation, e.g .,

(defrelation VR-scrub (?x):=> (default-nvalue (VR ?x) .80))

and relevance to military trafficability analysis :

(defrelationseverely-restricted-vegetation (?x)

._ (and (<= (SS ?x) 6)(>= (SD ?x) .15)))

Military vehiclesThe vehicle theory contains an ontology ofmilitary vehicles, with an emphasis on theirquantities and attributes relevant to cross-countrymobility . Vehicle characteristics provided by thedomain theory include maximum , road speed,width, length, height, weight, on-road gradability,off road gradability, fuel capacity, fuelconsumption (idle), fuel consumption(secondary), fuel consumption (cross-country),override diameter, vehicle cone index (1 pass),vehicle cone index (50 passes), minimum turningradius, load class, maximum gap to traverse,ground clearance, maximum step, maximum tilt,maximum gradient, and maximum straddle .

We account for a wide range of military vehicletypes, which allows us to reason about a varietyof trafficability scenarios . The domain theoryranges from representations of general classes ofvehicle :

6 8

(meters 17))))(dvalue (obstacle-category ?t) SR))) as well as descriptions of terrain regions that have

(forall (?x) some uniformity of attributes :(_> (and (terrain-feature ?x)

(or (dvalue (vegetation-type ?x) K) (defentity scrub(dvalue (vegetation-type ?x) J) :subclass-of (vegetated-terrain-feature)(dvalue (vegetation-type ?x) I))) :consequences ((uncultivated :self)(dvalue (obstacle-category ?t) SR))) (sapling-growth :self)

(defentity vehicle-wheeled:subclass-of(transportation-device--vehicle)

:quantities((number-of-wheels)):consequences((default-dvalue (drive-type :self)

wheeled)))

to specific vehicle types:

(defentity M-2; ; Bradley Infantry Fighting vehicle (BFV):subclass-of(InfantryFightingvehicleMilitaryvehicle-TrackedMilVeh-amphibious)

:consequences((default-nvalue(max-road-speed :self) 66)(default-nvalue(fuel-Capacity :self) 175)

(default-nvalue(fuel-Consumption-idle :self) 6 .4

. . . Etc .

:documentation "The standard IFV of U .S . Armymechanized infantry units .Equipped with . . .")

GIS Topology and Feature CodingA theory of GIS topology is necessary to providea way to describe and reason about therepresentation of the terrain. This gives us thevocabulary and definitions needed to produce theplace vocabulary layers in the GIS .For example, the domain theory containsknowledge about the way GIS manages data :

(defEntity gis-coverage-polyGIs coverage with polygon toplogy

:subclass-of (gis-coverage))

It includes the geometric representations used toindividuate regions of space :

(defEntity gis-polygon:subclass-of (gis-feature):quantities ((gis-poly-area)

(gis-poly-perimeter)))

And it includes axioms that define the relationsbetween these regions, which allow us to expressthe qualitative spatial descriptions that make upthe place vocabulary :

(forall (?polyl ?poly2 ?arc)(_> (and (gis-arc ?arc)

(gis-polygon ?polyl)(gis-polygon ?poly2)


(gis-right-of-arc ?arc ?polyl)(gis-left-of-arc ?arc ?poly2))

(and (gis-adjacent-polygons?polyl ?poly2)

(gis-adjacent-polygons?poly2 ?polyl))))

Another GIS domain theory module provides atranslation out of the format peculiar to the digitalterrain data set, and into the common predicatevocabulary provided by the terrain andtrafficability domain theories . It is a body ofrules that translates feature ID codes such asEB020 to the appropriate predicate vocabulary,(scrub GIS-POLY-021) : (

(forall (?poly)(=> (gis-poly-vt ?poly EB020)

(scrub ?poly)))

Because this translation is handled in aknowledge-based way, GIS data coded to someother format could be used just as effectively byproviding the appropriate feature coding domaintheory .

GIS Place Vocabulary AlgorithmsPlace vocabularies are computed in response to arequest from a reasoning system . For the sake ofcontext, it will be helpful to understand how weutilized a client-server relationship between thereasoning system and the GIS to enable this . Theprocess begins when the reasoner has recognizedthe need for a place vocabulary (i .e ., a, COO orCFO). The request is sent to the GIS via remoteprocedure call . The GIS server executes scriptsthat perform the requested operation, producesthe resulting place vocabulary in a format that thereasoning system can read, and signals thecompletion of the task to the reasoner . Thereasoner can then access the place vocabularythrough direct communication with the GIS (itselfa DBMS) or by directly reading uncompresseddata files. The reasoner thus has access to all ofthe information it needs .

( Our translation is for terrain data that is coded to theNIMA Vector Smart Map data Level I format (VMAP-1).[I31

6 9

We next outline the algorithms used forcomputing CFOs and COOS. The actual GIScode is written in Arc Macro Language (AML)for ARC/INFO version 7.x [151 ; We summarizethem in terms of the following operations :" Clip creates a new coverage by copying the

subset of an input coverage that overlaps agiven rectangle . Polygons that intersect thesides of the rectangle are clipped to match theboundaries .

" Union creates a new coverage with polygonswhose boundaries are the union of thepolygons in the input coverages .

" Dissolve merges adjacent polygons within acoverage that have identical values for aproperty provided as input.

Algorithm for generating a Complex FactorOverlaySince the CFO consists of the finest-graineddistinctions that can be made based on the inputdata, computing it is relatively straightforward .Given a set of input coverages Ci, representingthe factors and a rectangle representing the areaof interest (AI),

l . Let CFO = a new empty coverage2. For each input coverageCi,2.1 Let C = Clip(Ci,AI)2.2 CFO = Union(CFO,C)

Given data that has already been pre-processed tomeet military standards, this algorithm correctlycomputes a CFO by definition .

Algorithm for generating a CombinedObstacle OverlavComputing the COO is slightly more complexbecause we must compute the appropriate levelof abstraction. In addition to a set of inputcoverages, Cis, and an area of interest AI, anumber corresponding to the echelon forwhomthe analysis is being done (e.g ., brigade, division,corps, etc.) must be provided :

1 . Let COO = a new empty coverage2. For each input coverage Ci,

OR99 Loch Awe, Scotland

2 .1 Let Ci = Clip(Ci,AI)2.2 For each polygon P in Ci, assign obstacle-category W, U, R, or SR according to theconstraints of the domain theory .2.3 Let Ci = Dissolve(Ci, obstacle-category)

3 . For

each

input

coverage Ci,

let

COO

=Union(COO, Ci)

4.For each polygon P in COO, let obstacle-category(P) be the most restrictive of thevalues for obstacle-category of thecorresponding polygons in the inputcoverages.

5 . Let COO = Dissolve(COO, obstacle-category)6. Remove all polygons in COO that are too small

for the echelon under consideration

Essentially, this algorithm find the maximalregions that are severely restricted, restricted, andwater, and then prunes out regions that are toosmall for the given echelon.

Empirical Results

We used these techniques to generate CFOs andCOOs to answer terrain analysis and trafficabilityquestions relevant to planning and conductingmilitary operations . The GIS data set we usedrepresented the terrain in the Straits of Hormuzregion . The relevant coverages in this terraindata included vegetation, hydrology, slope, androad network. This array of coverages, and the(predominantly desert) terrain in the areaprovided an opportunity to test these techniques,and the early results are promising. Inspection bymilitary personnel of the CFOs and COOsproduced was used to judge correctness, andplausibility of trafficability results judged in termsof producing results consistent with U.S . Armypractice . In all areas tested, correct CFOs andCOOS were created, and trafficability questions(e .g ., maximum speed in particular regions)produced correct answers .

Producing CFOs and COOS gave us theopportunity to produce place vocabularies thatcorrespond to authentic qualitative descriptionsproduced by Army analysts, and to model the

70

well-established terrain analysis practices that usethose descriptions . The payoff is the automatonof this practice, authentic results conductingterrain analysis and trafficability tasks, andcompelling explanations grounded in qualitativereasoning. While military analysts currentlyspend hours or days producing the same overlaysfor an area of operations, this technique allowsthe same descriptions to be producedimmediately, on demand. They can be producedin response to a human analyst, as well as inresponse to an automated reasoning process.

Discussion

We have shown that a GIS can be used toproduce representations for qualitative spatialreasoning . The geometric processing facilities ofthe GIS provide the capabilities in a metricdiagram . Layers of a GIS computed with respectto task-relevant properties constitute placevocabularies . Thus the power of qualitativerepresentations and reasoning can be brought togeospatial applications, riding on the progressmade by the commercial world in GIS software .

There are several directions that should beexplored next . First, further testing is needed,especially using terrain data that differssubstantially from the arid Hormuz area . OurHormuz data also lacked information about soils,surface configuration, and surface roughness, sowe have not validated our techniques with thoseaspects of terrain data . Second, closely relatedproblems for which trafficability plays animportant role (e.g ., analyzing the suitability of anarea for an operation, accessibility, and pathfinding) should be tackled, to provide a moreencompassing set of tests. Third, while weexperimented with GIS scripts that would call ourreasoning system while they were running, thiswas both inefficient and impractical, given theconcurrent development of the domain theories,reasoning system, and GIS proceduresConsequently, we hand-coded the knowledgefrom the domain theory into the GIS scripts. Thisis a sensible solution, given that doctrine changes


very slowly, and the same scripts can thus bewidely used . However, either on-line integratedreasoning or automatic production of GIS scriptsmay be worth exploring for other GIS tasks.Fourth, the domain theories and placevocabularies should be expanded to include otherspatial representations that are related totrafficability that would be a great value to terrainanalysts . These would demand additionalgeometric processing, and different models inorder to derive them. The avenue of approachoverlay is an example of such a task, and a naturalextension of the work we have done here .

There are many other geospatial tasks whereusing a GIS for qualitative spatial reasoning islikely to provide important benefits . Examplesinclude anthropology (e.g . the migrations ofcultures) and agriculture (e.g . crop selection) .Regardless of the application, harnessing thegeospatial computing power of commercial GISshould contribute significantly to creatingqualitative spatial reasoners .

Acknowledgements

This research was supported by the DefenseAdvanced Research Projects Agency, under theHigh Performance Knowledge Bases program,and by the Air Force Office of SponsoredResearch . We thank Tom Mostek, Rob Rasch,and Jeff Usher for their help and usefuldiscussions.

7 1

References

1 . B . Falkenhainer, A . Farquhar, D. Bobrow, R.Fikes, K. Forbus, T. Gruber, Y . Iwasaki, & B .Kuipers . CML: A Compositional ModelingLanguage . Knowledge Systems Laboratory,KSL-94-16, September 1994.

2 . Cohn, A., & Randall, 2.(1992) . An IntervalLogic for Space Based on (Connection'.Proceedings of the 10th European Conferenceon Artificial Intelligence . 394- 398.

3 . Falkenhainer,

B.

and

Forbus,

K."Compositional Modeling : Finding the RightModel for the Job"; Artificial Intelligence,51(1-3), October, 1991 .

4 . Forbus, K . "Spatial and Qualitative Aspects ofReasoning about Motion" Proceedings of thefirst annual conference of the AmericanAssociation for Artificial Intelligence, August1980 .

5 . Forbus,

K.

Qualitative

Process

Theory.Artificial Intelligence, (24), 1984, pp 85-168

6. Forbus,

K.,

Nielsen,

P.

and

Faltings,

B.'Qualitative Spatial Reasoning : The CLOCKProject', Artificial Intelligence, 51 (1-3),October, 1991 .

7 .Forbus, K. 1995 . Qualitative SpatialReasoning : Framework and Frontiers . InGlasgow, J., Narayanan, N., andChandrasekaran, B. DiagrammaticReasoning : Cognitive and ComputationalPerspectives. MIT Press, pp . 183-202 .

8 . FM 5-33 Terrain Analysis, Headquarters,Department of the Army, July 1990 .

9 . FM 34-130 Intelligence Preparation of theBattlefield, Headquarters, Department of theArmy, July 8,1994 .

10 . M. R. Genesereth & R. E. Fikes . KnowledgeInterchange Format, Version 3 .0 ReferenceManual . Knowledge Systems Laboratory,KSL-92-86, June 1992 .

11 . Glasgow, J., Karan, B ., & Narayanan, N .(Eds.) . (1995) . Diagrammatic Reasoning .Cambridge Mass. : AAAI Press/The MITPress .


12 . Hayes,

P.

(1979).

The

naive

physicsmanifesto. In D. Michie (Ed.), ExpertSystems in the Micro-Electronic AaeEdinburgh University Press .

13 . Military

specification

MIL-PER-89040,Vector Product Interim Terrain Data (VITD)May

8,

1995,http :// 164.214.2 .59/publications/specs/printed/VITD/vitd.html .

14. Student Handout,

"Tactical

Commander'sDevelopment Course Battle Book", U.S .Army Command and General Staff College,Fort Leavenworth, KS

15 . ARC/INFO Geographic Information Systemfor Windows NT, version 7.2,©Environmental Systems Research Institute,Inc .

7 2

usingageographicinformationsystemfor qualitative ......qualitative insights that users need. in...

Documents