EARTHQUAKE DAMAGE FROM AERIAL IMAGES BY PROBABILISTIC METHOD Shota Izaka Hitoshi Saji Shizuoka University Introduction Backgrounds After largescale earthquake Urban areas are seriously damaged ID: 255200
Download Presentation The PPT/PDF document "ESTIMATION OF" is the property of its rightful owner. Permission is granted to download and print the materials on this web site for personal, non-commercial use only, and to display it on your personal computer provided you do not modify the materials and that you retain all copyright notices contained in the materials. By downloading content from our website, you accept the terms of this agreement.
Slide1
ESTIMATION OFEARTHQUAKE DAMAGEFROM AERIAL IMAGESBY PROBABILISTIC METHOD
Shota
Izaka
, Hitoshi
Saji
(Shizuoka University)Slide2
IntroductionSlide3
BackgroundsAfter large-scale earthquakeUrban areas are seriously damagedMany people require rescuing and aidFor effective rescue and victim supportRapid action is neededA wide range of information is important
Aerial images are suitable
for disaster observationSlide4
Conventional methodMatching analysisComparing pre-disaster and post-disaster imagesDifficulty of matching analysisDifficult to obtain pre-disaster imagesAffected by shooting conditions and timeChanges of shadowsConstruction and destruction of buildingsSlide5
Our goalsRapid analysis of damageUse only post-disaster aerial imagesNot using the training dataAssisting various rescue and victim support activities
Providing information available for various purposes
Assisting human decisionsSlide6
Ways of assisting human decisionsRemaining undetermined regionsWe don’t force to classify all regionsThe final decision is left to the people in the fieldShowing the likelihood of damagesThe result available for various purposesTarget area estimation of rescue activity
Determination of the road passable for emergency vehiclesSlide7
MethodSlide8
OverviewAerial Image
Segmentation
Feature
extraction
Result
for buildings
Digital
map
Region
classification
Result
for roads
Road mask creationSlide9
Road mask creationCreating road mask from digital mapRoads change little over time Our method is not affected by the time when the map is created
Digital map
Road maskSlide10
SegmentationInitial SegmentationSegment into small basic regionsUnification of similar regionsConsidering color and texturesAvoiding to unify roads and buildings
Before segmentation
After segmentationSlide11
Feature extractionCollapsed buildingsSegmented into small regionsHaving short random edges Extracting short edges as a feature of damages
Collapsed buildings
Segmented regions
EdgesSlide12
Feature extractionUndamaged buildingsMaintaining their shapesHaving a large area Extracting building regions
as a feature of undamaged
Undamaged buildings
Segmented regions
EdgesSlide13
Region classificationUsing the probabilistic relaxation methodLabeling method using the probability We use the method to classify each region by damage probabilitySlide14
Defining initial probabilityConsidering extracted featuresThe proportion of short edgesThe area of regionBuilding region or not
Large
area
Building
High short edge rate
Probability definitionsSlide15
Probability updateUpdate using similarityConsidering the region similar to damaged region as damaged region
Probability update model
Low
High
High
High
High
High
High
High
High
High
High
High
High
High
High
HighSlide16
Extracting undamaged regionsRegions are converged high or low probabilityExtracting low probability regions as undamaged regionsConsidering regions not converged as undetermined regions
High probability
Result of extraction
Low probability
UndeterminedSlide17
Extracting damaged regionsExtracting damaged regions from high probability regions
High
probability
Damaged regions extraction model
Low
probability
Undetermined
Damaged
UndeterminedSlide18
Redefining initial probabilityRedefining probability by randomness of edgesUsing variance of edge angles
Edge model of
undamaged buildings
Edge model of
collapsed buildingsSlide19
Result of classification■:Undamaged regions■:Undetermined regions 1Low risk of damage■:Undetermined regions 2
High risk of damage
■
:Damaged regions
Result of classification
Undetermined
Damaged
Undetermined
UndamagedSlide20
Image divisionDividing a result image into buildings and roadsResult of buildingsEstimation of building damagesResult of roadsDetermination of road passableSlide21
ExperimentSlide22
DataAerial imagesGreat Hanshin EarthquakeCaptured on January 18, 1995Provided by PASCO Corp.Digital mapA topographic map of Kobe cityProvided by Kobe City Urban Planning BureauSlide23
Result of classification for buildings
Input image
Result image
■
:Undamaged regions
■
:Undetermined regions 1
■
:Undetermined regions 2
■
:Damaged regionsSlide24
Result of classification for roads
Input image
Result image
■
:Undamaged regions
■
:Undetermined regions 1
■
:Undetermined regions 2
■
:Damaged regionsSlide25
Evaluation of accuracyCreating answer imagesUsing visual judgmentComparing with results
Result of classification
Undetermined
Answer
Damaged
Undamaged
Damaged
Undamaged
UndeterminedSlide26
Detection rateEvaluating pixels in same category
Result of classification
Answer
Damaged
Undamaged
Damaged
Undamaged
Damaged
Undamaged
Damaged
UndamagedSlide27
Detection ratewith human decisionsEstimating rate after human decisionsAdding undetermined regions
Result
Damaged
Undamaged
Answer
Damaged
Undamaged
Damaged
Undamaged
Damaged
UndamagedSlide28
False detection rateEvaluating pixels in wrong categoryVisual judgment
Considered undamaged regions
Damaged
Undamaged
Considered damaged regions
Result of classification
Damaged
UndamagedSlide29
Answer for buildingsResult image
Answer image
■
:Undamaged regions
■
:Undetermined regions 1
■
:Undetermined regions 2
■
:Damaged regionsSlide30
Answer for roadsAnswer image
Result image
■
:Undamaged regions
■
:Undetermined regions 1
■
:Undetermined regions 2
■
:Damaged regionsSlide31
Result of accuracy evaluation in buildingsUndamaged regionsDetection rate:77.2%With human decisions:93.1%False detection rate:10.1%Damaged regionsDetection rate:74.0%With human decisions:87.0%
False detection rate:17.7%Slide32
Result of accuracy evaluation in roadsUndamaged regionsDetection rate:85.5%With human decisions:93.4%False detection rate:19.0%Damaged regionsDetection rate:65.3%With human decisions:79.6%
False detection rate:14.6%Slide33
Review of resultsObtained high detection ratesExcept for damaged regions in roadsFeatures of damage on roads are unclearMany regions classified into “Undetermined” Requiring human decisions
Road image
Result of classificationSlide34
Review of resultsObtained low false detection ratesRoads have more errors than buildingsCaused by objects on roadsCars, roofs, shadows of buildings
Roof and car
Error
Shadow
and car
ErrorSlide35
ConclusionOur results can be used for various rescue and victim support activityEstimation of building damagesDetermination of road passableOur future directionsImproving building detectionDetecting objects on roadsSlide36
EndSlide37
The Sendai earthquakeMost of the damage was caused by the TsunamiMost of the buildings are flooded outOur method aim to detect collapsed buildingsHuge area of damageNot possible to capture by aerial images
Applying to the earthquake is future works