Modelling Historic Fire Boundaries Using Inventory Data

2004 Western Forest Mensurationists Conference

Kah-Nee-Ta Resort, Warm Springs, ORJune 20-22, 2004

Rueben SchulzDr. Peter Marshall

Project and Study Area

• “Comparing Stand Origin Ages with Forest Inventory Ages on a Boreal Mixedwood Landscape”– Historic fire mapping in Saskatchewan– 90 000 ha (220 000 acre) study area– Area only recently logged

Project and Study Area

We are here

Project Context

• Emulate natural disturbance regimes for management – Ecosystems will continue to

function if we keep disturbances like they were historically

– Wildfire is the primary disturbance on this landscape

• Time-Since-Fire dataset • Use existing inventory data

– derive fire information – guide sampling

Image: NASA and Canadian Forest Service

Datasets: Time-Since-Fire

• Time-since-fire (TSF) data – Records the year of the last wildfire disturbance – Expensive to collect

• TSF does not record site and species differences

• Does not include human-made features• Ages recorded to the nearest year• Location of ground plots different from inventory

– Near fire edges

• ~ 900 fire polygons

Datasets: TSF

Datasets: Forest Inventory

• New inventory follows the Saskatchewan Forest Vegetation Inventory (SVI) standard

• Inventory records current condition

• Includes human-made features

• Up to 3 tree layers

• Ages in 10 year classes

• ~10 000 inventory polygons

What was done

1. Sampling to collect time-since-fire data

2. Analysis:1. Regression modelling

2. Clustering fire events

Sampling

• Data over the full study area was collected in 2002– Aerial photos – fire

boundaries and location of ground plots

– Ground plots – tree ages, fire scars and release information

• Combined to produce a time-since-fire dataset

Single-Aged Regression Models

Forest Inventory

Multi-Aged Regression Models

Fill in missing values from neighbours

Predicted Time-Since-

Fire

Cluster Predicted Time-Since-Fire

Predicted Fire Events

Analysis Overview

Regression Modelling

• Predict time-since-fire from forest inventory inputs– Inventory stand used as unit of observation– Time-since-fire is the dependent variable – Continuous predicted variable

• not continuous in space

• Linear model forms

• Used GLM with categorical inputs as dummy variables

Modelling: Single-Aged Polygons

• Significant variables– Inventory stand age– Modification value– Leading and Secondary species– Average stand age within 400m

• R2 ~ 0.4

Modelling: Multi-Aged Polygons

• More age values to deal with

• Significant variables– Inventory stand age (average)– Modification value– Average stand age (max) within 400m

• R2 ~ 0.25

Modelling: No-Aged Polygons

• Stands with no inventory information – Roads, gas lines, clearcuts

• Large, road and gas polygons split up

• Assigned neighbour values to fill holes

Modelling: Output

Modelling: Fit

Clustering

• Models only predict time-since-fire for an individual forest inventory stand

• Want location and extent of fire events

• Grouped stands with similar predicted time-since-fire using hierarchical clustering

• Was spatially constrained– Penalty distance matrix added to clustering

Clustering: Output

Clustering: Output

Clustering: Fit

Challenges and Future

• Forest inventory ages are not a direct substitution for time-since-fire

• More work to do– R2 for models is poor

• Future work– Examine autocorrelation in regression models– Improve fire boundary detection– Inclusion of additional data: DEM, Landsat

imagery

Acknowledgements

• NSERC• Forest Development Fund of the Saskatchewan

Forest Centre (SFC)• Sustainable Forest Management (SFM) Network• Forest Ecosystems Branch of the Saskatchewan

Environment and Resource Management (SERM)

• Mistik Management Ltd. • Phil Loseth –draft inventory manual, last years

conference