Mapping Wildfire Disturbances in Southern California Using Machine Learning Algorithms John Rogan [email protected] Geography Department Clark University.

Mapping Wildfire Disturbances in Southern California Using Machine Learning Algorithms

John Rogan

[email protected]

Geography DepartmentClark University

Research Context

• Land cover/use change – Mapping and Monitoring

• Growing interest in Rapid Response Information Systems

• Empirical information about wildfire cause and behavior can inform wildfire risk analysis

• Paucity of historic burn severity information

Wildfire Monitoring Programs

Monitoring Fire Effects

• The physical environment and its response to fire

• Factors affecting fire behavior

• Ecosystem/watershed damage assessment

• Evaluating success of a management ignited fire

• Appraising the potential for future treatments

Existing Monitoring Efforts (BAER)Watershed Scale

Source: California Department of Forestry and Fire ProtectionSource: California Department of Forestry and Fire Protection

Existing Monitoring EffortsEcoregion Scale

Source: California Department Source: California Department of Forestry and Fire Protectionof Forestry and Fire Protection

A = Detected from satelliteB = Conventional methods

Existing Monitoring EffortsNational Scale

Source: Canadian Center for Remote SensingSource: Canadian Center for Remote Sensing

Existing Monitoring EffortsGlobal Scale

Source: University of Maryland Global Fire ProductSource: University of Maryland Global Fire Product

Research Objectives

• Test a new methodology to map fire severity in San Diego County (1985-2000)

• Employ machine learning to map severity, while integrating environmental variables with spectral variables (categorical and continuous)

• Examine the contribution of ancillary variables to burn map accuracy

Fire/Burn Severity

• Severity - A descriptive term that integrates various phenomenological characteristics of a fire-altered landscape

• Physical and biological manifestation of combustion on vegetation and soil

• Direct Effects Influenced by– Fuel consumption - Topography– Crown scorch - Disturbance history– Soil heating– Bole Charring

Previous Research

• Focus on mapping burn scars (coarse resolution)

• Recent emphasis on burn severity/mortality/damage levels (fine-medium resolution) for impact assessment

• Retrospective burn area mapping at medium resolution (e.g., Hudak and Brockett 2004—IJRS)

• BUT, challenges remain…………

Scene Model

30 meters

0o Slope

60o Slope

30 meters

SOIL GV NPV SHADE CHAR/ASH

Sensor Source

Post-Fire IKONOS-2 Image

California Wildfire Threat

Source: California Department of Forestry and Fire ProtectionSource: California Department of Forestry and Fire Protection

San Diego San Diego CountyCounty

Study Area Significance

• Impacts of natural disturbance processes are increasing in severity

• Public lands began burning more frequently than private lands in the mid-1970s. This trend is increasing

• Population increase and peri-urban spread into fire-prone areas (WUI)

Landsat TM and ETM+ Data

May August

Signal

Date MonthTime SinceLast Fire

1985 July 3

1988 August 2

1990 June 1

1992 June 0

1996 July 0

1998 July 1

2000 August 0

Environmental Variables

Ground Reference Data

SITE-Dominance

• Grassland (6)

• Chaparral (10)

• Conifer-Hardwood (5)

• Mixed (12)

Composite Burn Index Key and Benson (2000)

Wildfire Effects (After Key and Benson)

Methods (Data Processing Flow)

TMPost-Fire

Reducing the Effects of Spatially -Varying

HazeSpatially -Varying

Haze

•• aspect

• slope

• elevation

ClassifierAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data•• aspect

• slope

• elevation

• burn signatures (SMA)

• aspect

• slope

• elevation

ClassifierAssess

ClassificationAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data

NotCorrect

Refine Training Data

Accuracy Assessment

Accuracy Assessment

Error MatrixError MatrixError MatrixError Matrix

TMPost-Fire

Reducing the Effects of Spatially -Varying

HazeSpatially -Varying

Haze

Reducing the Effects of Spatially -Varying

HazeSpatially -Varying

Haze

Reducing the Effects of Spatially -Varying

HazeSpatially -Varying

Haze

•• aspect

• slope

• elevation

ClassifierAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data•• aspect

• slope

• elevation

• burn signatures (SMA)

• aspect

• slope

• elevation

ClassifierAssess

ClassificationAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data

NotCorrect

Refine Training Data

Accuracy Assessment

Accuracy Assessment

Error MatrixError MatrixError MatrixError Matrix

•• aspect

• slope

• elevation

ClassifierAssess

ClassificationAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data

NotCorrect

Refine Training Data•• aspect

• slope

• elevation

• burn signatures (SMA)

• aspect

• slope

• elevation

ClassifierAssess

ClassificationAssess

Classification

Final Change Mapwith Canopy Classes

Correct

NotCorrect

Refine Training Data

NotCorrect

Refine Training Data

Accuracy Assessment

Accuracy Assessment

Error MatrixError MatrixError MatrixError MatrixError MatrixError MatrixError MatrixError MatrixError MatrixError MatrixError Matrix

(a) (b)

Band 1Band 2

Band 3Bands 4,3,2

Haze Correction (Pre-)

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

44

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

444 Haze4 Haze

2 Haze2 Haze

3 Haze3 Haze

1 Haze1 Haze

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

44

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

444 Haze4 Haze

2 Haze2 Haze

3 Haze3 Haze

1 Haze1 Haze

4 Haze4 Haze2 Haze2 Haze

3 Haze3 Haze

1 Haze1 Haze

(a) (b)

Band 1Band 2

Bands 4,3,2Band 3

Haze Correction (Post-)

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

44

Band XBand X

Ban

d Y

Ban

d Y

11

22

33

44

Spectral Mixture Analysis

• Decomposition of mixedpixel spectral response

• Production of fractionalrepresentation of sub-pixel proportions

• Biophysically-meaningfulestimates of land covercomponents

Classification Tree Analysis

• A type of MLA used to predict membership of cases of a categorical dependent variable from their measurements on one or more predictor variables

• Hierarchical, non-linear recursive partitioning

• Structurally explicit

Desired Map Accuracy

0

10

20

30

40

50

60

70

80

90

100

1 3 4 5

Number of Burn Severity Classes

Me

an

Ov

era

ll A

cc

ura

cy

(%

)

Source: Rogan and Franklin (2001)Source: Rogan and Franklin (2001)

Case Study (Pre-Fire)

Case Study – (Post-Fire)

Case Study Results - SMA

0 %

FractionalAbundance

100 %

Shade GV

BV Soil

RMS Vegetation MapHardwood

Grassland

Chaparral

Conifer

Scrub

Fire perimeter

Case Study Results - Variable Selection

Map Class Variables Selected

No Burn - Vegetation

GV, Soil, Veg

No Burn - Water Shade, Slope

No Burn - Soil Soil, Veg

Severe Burn BV, Veg, Slope, Shade

Moderate Burn BV, GV, Veg

Light Burn GV, Soil, BV

Case Study Results – Burn Severity

No burn – Vegetation 90%No burn – Water 100%Severe burn 87%Moderate burn 60%Light burn 74%

No burn – Bare Soil 84%

CLASSCLASS ACCURACY (kappa)ACCURACY (kappa)

82.5%

County-Wide Results

• Variable Selection by Site(s)

– Grass Burn, Soil– CHP Burn, GV, Soil, Veg, SlopeVeg, Slope– CON/HDW Burn, GV, Soil, Veg, SlopeVeg, Slope, Shade– Mixed Burn, GV, Slope, VegSlope, Veg, Shade, Slope, AspectSlope, Aspect

• Mean Burn Map Accuracy by Site(s)

– Grass 87% (SD = 11%)– CHP 81% (SD = 10%)– CON/HDW 84% (SD = 7%)– Mixed 70% (SD = 16%)

County-Wide Results

• Time since fire (TSF)– Most problematic for grasslands, where TSF >

3 months– Least problematic for CHP, CON/HDW

• Map accuracy – Lowest for complex classes (e.g., mixed)– Highest for simple classes (e.g., grassland)

• Variable Selection – Many for complex classes (e.g., mixed)– Few for simple classes (e.g., grassland)

Implications

• Map accuracy – Range – 70-80%, depending on landscape type

and TSF– Subtle burn classes are least accurate

• Variable Selection – Varied by landscape type (all used for complex

areas)– Implication for fire risk mapping?

– The larger the fire, the greater the potential for confusion caused by landscape heterogeneity

Wildland Fire Mapping Triangle

Burn Severity Map

Machine Learning

Pre

dictive V

egeta

tion

Modelin

g

Imag

e Pro

cess

ing

and E

nhan

cem

ent

“….search for standard methods for mapping fuels and fire regimes at high (spatial) resolutions over broad areas”.

Rollins et al. (2004, p. 86)

Acknowledgements

• NASA Land Cover Land Use Change Program

• US Forest Service and CDF

• SDSU – Janet Franklin and Doug Stow

• UCSB – Dar Roberts and Alexandria Digital Library

• U of Arizona – Steve Yool