Lecture 31

rucksackbulgeAI and Robotics

Dec 1, 2013 (3 years and 8 months ago)

93 views

FORS 8450 • Advanced Forest Planning



Lecture 31


An overview of methods for incorporating

wildfires into forest planning models


Forest planning processes have become increasingly complex over

the last few decades, due to advances in computer technology, advances

in basic scientific knowledge of the functional relationships among

natural resources, and the introduction or modification of land use regulations.


Both spatial and temporal

characteristics of desired

future landscapes may now

also be important to recognize.



In some cases, the potential

impact of wildfires is desired

by decision
-
makers.

Background

Efforts to model the potential effects of forest fires vary from the

manipulation of non
-
spatial vegetation strata, to the spatial recognition

of the spread of fires across a landscape.


This lecture will

describe a suite of

methods people have

used for incorporating

wildfires into forest

planning models.

Background

What this lecture covers:


Attempts to account for wildfire in a forest planning process.



Linear programming


Binary search


Benefit / cost analysis


Simulation


Heuristics



and other methods


What this lecture does not cover:


Specific fire modeling processes

Fire simulation without the integration of forest planning processes

Stand
-
level analyses of fire impacts

Background

Why is the incorporation of fire effects into forest planning important?

1) The potential for fire loss is high in some regions.


2) The potential for fire loss is highly variable.


3) Losses from fire can significantly affect timber supplies.


4) Losses from fire can destabilize local economies.


Kalabokidis, K.D., Gatzojannis, S., and Galatsidas, S. 2002. Introducing wildfire

into forest management planning: Towards a conceptual approach. Forest Ecology

and Management. 158(1
-
3): 41
-
50.


Described here is a conceptual model to quantify wildfire conditions through inventory

methods and landscape analysis (GIS).


Process:

1. Estimate external factors to determine the risk of a forest to wildfire.

2. Estimate internal factors to determine the productivity and resilience of the forest.

3. Combine the two to evaluate wildfire loss potential across a forest.


Problems: (a) some factors are qualitative, (b) factors are measured using non
-
standard

measurement units, and thus (c) incorporating these into a decision process is difficult.


Suggestions: Estimate the relative weight of wildfire factors, and rank these into

"danger classes." Use a benefit / cost analysis to inform decisions related to forest

planning.

Benefit / Cost Analysis

1/18

Reed, W.J., and Errico, D. 1986. Optimal harvest scheduling at the forest level

in the presence of the risk of fire. Canadian Journal of Forest Research.

16: 266
-
278.


A Linear programming, generalized form of a Model II planning problem.

Deterministic (fixed) proportions of areas in each age class are lost to fire.


Process:

1. Find first time period harvest in absence of fire.

2. Observe what would have happened to the forest with fire.

3. Adjust first period harvest, re
-
run problem, examining subsequent time period(s).


This is an "iterative state updating procedure." A close approximation to an optimal

solution for a forest plan can be developed using deterministic wildfire distributions

that closely resemble the stochastic disturbance levels.


Associated work:

Moll and Chinneck (1992); Martell (1994); Boychuk and Martell (1996)


Linear Programming

2/18

Armstrong, G.W. 2004. Sustainability of timber supply considering the risk of

wildfire. Forest Science. 50(5): 626
-
639.


A linear programming / Monte Carlo simulation model.

Burned areas are randomly generated from a lognormal continuous distribution.

This is a "sequential update procedure."


Process:

1. At the beginning of time period 1, determine the allowable cut, ignoring the possibility


of wildfire.

2. Update the inventory.

3. Randomly determine the burned proportion, and reduce the area of each age


class by that proportion. The burned areas go to the youngest age class.

4. Repeat the process for time periods 2
-
n
.


This process was repeated 1,000 times to generate allowable cuts as probability

distributions.


Associated work:

Armstrong et al. (1999)

Simulation

3/18

Gassmann, H.I. 1989. Optimal harvest of a forest in the presence of uncertainty.

Canadian Journal of Forest Research. 19: 1267
-
1274.


Maximized harvest in a finite time horizon using a Model II problem formulation.


In general, some area in each age classes may be harvested, then a random portion


of the remaining area may be destroyed by a disturbance (assumed here to be fire).


Loss rates from fires could be deterministically or randomly chosen.


A problem using a deterministic proportion of disturbance in each age class was first


addressed using linear programming.


When probabilities of fire loss were incorporated into the problem, a special computer


program based on the Dantzig
-
Wolfe decomposition principle was employed. This


was still a Model II problem, and termed "stochastic programming."

Stochastic Programming

4/18

Stevens, N.S. 1986. Optimal harvest rates considering the risk of forest fire.

M.Sc. Thesis, Department of Forest Science, University of Alberta, Edmonton. 132 p.


The multi
-
objective problem incorporated the value of timber harvest, and the


a value to describe the smoothness of wood flows from the planned activities.


The state variable used to describe the forest was the even
-
flow level that was


determined by area
-
volume check (a proxy for the age
-
class distribution of the forest


at any point in time).


State
-
transition matrices were generated from a Monte Carlo simulation model that


recognized different harvest levels and fire regimes, which were based on lognormal


distributions.

Stochastic Dynamic Programming

5/18

Montgomery, C.A., Brodie, J.D., and Cleaves, D.A. 1986. Allowable cut effect

and fire
-
damage appraisal. Western Journal of Applied Forestry. 1(4): 100
-
103.


Used the SHRUB model (an enhanced version of HARVEST).

Forest age classes (strata) and growth assumptions are used in conjunction with


harvest rules to assess the impacts of fire on a long
-
term harvest schedule.


Process:

1. Given the objectives and constraints, develop a forest plan using the binary search model.

2. Assume a single fire occurred in time period 1, and remove areas assumed to be burned


from their strata, and place them in the youngest age class (strata).

3. Develop another forest plan using binary search.

4. Compare the present value of the two plans.


A comparison of the with, and without, assumptions of wildfire reveals trade
-
offs

in the growth and harvest of individual age classes (strata) over time. Applying this

approach to different types of forests helps one understand how the impact of wildfire

can perhaps be significantly different.

Binary Search

6/18

Beukema, S.J., and Kurz, W.A. 2000. Vegetation Dynamics Development

Tool users’ guide, Version 4.0, ESSA Technologies Ltd., Vancouver, BC.

(www.essa.com/forestry/vddt).


VDDT (Vegetation Disturbance Dynamics Tool)

Initially developed for the Interior Columbia River Basin Project.

A non
-
spatial state
-
transition model that incorporates disturbance probabilities and


successional pathways to project landscapes into the future.

The spatial version of this model is TELSA.


Pathway diagrams for forest and rangeland vegetation complexes were developed.

Expert opinion is used to understand the successional pathways, disturbance probabilities,


and impacts of disturbances and management actions on stand structure and


composition.



State
-
Transition Models

7/18

Van Wagner, C.E. 1979. The economic impact of individual fires on the whole forest.

The Forestry Chronicle. 55(2): 47
-
50.


A simple model examining fire loss appraisals, using a net present value approach.

The focus is on a regulated forest assigned to a grid of "units" that are all of equal size.


Process:

1. A deterministic selection is made of the units within which fires damage the timber


resources. This occurs at the beginning of the simulation. These stands are sent


to the youngest age class and regeneration is assumed to have occurred.

2. Harvests are scheduled for the entire time horizon using a rule of scheduling oldest


age classes first.


The results show how the impact of fire can influence the net present value of the entire


forest, when discount rates and growth rates vary.

Basically, substitutions for what would have been the next scheduled harvest units


are made, after assuming these were destroyed from fire and subsequently


regenerated.



Spatial Simulation Model

8/18

Van Wagner, C.E. 1983. Simulating the effect of forest fire on long
-
term annual

timber supply. Canadian Journal of Forest Research. 13: 451
-
457.


Again, the focus is on a regulated forest assigned to a grid of "units" that are all of

equal size.


Process:

1. A random selection is made of the units within which fires damage the timber


resources. These stands are sent to the youngest age class and regeneration is


assumed to have occurred. The rate of disturbance is fixed.

2. Harvests are scheduled for other units at a pre
-
defined percentage cut per


year, using a "highest volume" first approach.


Annual harvest levels, given the percent of harvest allowed and the percentage


of area burned, can then be simulated.

Spatial Simulation Model

9/18

Yang, J., He, H.S., and Gustafson, E.J. 2004. A hierarchical fire frequency model

to simulate temporal patterns of fire regimes in LANDIS. Ecological Modelling.

180: 119
-
133.


LANDIS landscape simulation model. A grid
-
based modeling approach.

Size, frequency, and intensity of fires can be simulated using probability distributions.

Simulated disturbances remove trees from age classes.


Process:

1. Select sites to be disturbed.

2. Calculate the disturbance intensity.

3. Remove tree cohorts from age classes in the areas disturbed.

4. If an ignition becomes an initiated fire, it is grown (spread) up to a pre
-
determined


size drawn from a size distribution, depending on fuel type.


Associated work:

Shifley et al. (2000)


Spatial Simulation Model

10/18

Peter, B., and Nelson, J. 2005. Estimating harvest schedules and profitability

under the risk of fire disturbance. Canadian Journal of Forest Research. 35: 1378
-
1388.


FPS
-
ATLAS harvest scheduling simulator.

Vector polygons represent the stands.

Size and frequency of fires are simulated using probability distributions.

When a disturbance is initiated, a polygon is randomly selected, then a patch size


target is selected. Adjacent polygons are added to the patch until a size target is met or


there are no more eligible polygons. Patches are allowed to spread in any direction.


Process:

1. Schedule harvests for the first time period, using a modified "oldest
-
first" rule,


and account for adjacency and green
-
up.

2. Call a routine that determines the areas that are burned.

3. Determine whether to salvage or naturally regenerate the burned areas.

4. Repeat the process for the remaining time periods.

5. Assess whether a sustainable harvest level has been achieved, perhaps re
-
run the model.

Spatial Simulation Model

11/18

Keane, R.E., Holsinger, L.M., and Pratt, S.D. 2006. Simulating historical landscape

dynamics using the landscape fire succession model LANDSUM version 4.0.

USDA Forest Service, Rocky Mountain Research Station, Ft. Collins, CO. General

Technical Report RMRS
-
GTR
-
171CD. 73 p.


LANDSUM (LANDscape SUccession Model, version 4.0).

A spatial, state
-
transition model that operates at the stand level.

Units are stored spatially as polygons.

Includes a spatially
-
explicit fire spread model, which operates at the pixel level


within stands, and fires are spread using directional vectors of wind and slope.

Disturbance initiations are modeled stochastically, and effects are based on current


vegetation conditions.

Maximum fire size, ellipses, and cellular automata approaches to spreading fire.

If stands are partially burned, the burned and unburned portions are assigned


different successional classes and stand ages.

Can work in conjunction with VDDT to determine successional pathways.

Human
-
induced disturbances (harvests) can also be modeled stochastically.

Spatial Simulation Model

12/18

Kurz, W.A., Beukema, S.J., Klenner, W., Greenough, J.A., Robinson, D.C.E.,

Sharpe, A.D., and Webb, T.M. 2000. TELSA: The Tool for Exploratory Landscape

Scenario Analyses. Computers and Electronics in Agriculture. 27: 227
-
242.


TELSA (Tool for Exploratory Landscape Scenario Analyses)
-

a spatially
-
explicit model.

Requirements: Stand polygons, streams, wetlands, lakes, transition probs (from VDDT).

Disturbances are modeled by polygon. Adjacency and activity limits are controlled.


Process:

1. Simulate natural succession in the first time period.

2. Simulate natural disturbances in polygons, up to a user
-
defined limit per year and


size, thus disturbances can be spread to eligible neighboring polygons.

3. Simulate salvage logging.

4. Assign management activities to stands deterministically from a sorted list of stands.

5. Repeat for all other time periods.


Associated work:

Provencher et al. (2007), Strand et al. (2009)

Heuristic Model

13/18

Campbell, K.A., and S.M. Dewhurst. 2007. A hierarchical simulation
-
through
-
optimization

approach to forest disturbance modelling. Ecological Modelling. 202: 281
-
296.


Modeled the temporal pattern and process of fire disturbance through traditional


harvest scheduling methods, using a simulated annealing heuristic.

This is a goal programming problem, where the objective was to minimize deviations from


desired landscape conditions.

Requires vector data (stand polygons, etc.).


As with harvests, disturbances are modeled by polygon, are randomly located


regardless of stand age, and are based on transition probabilities.

Thus the temporal pattern of disturbance is modeled as a harvest, from historical evidence


of the frequency of fires.

The process does not model directly the disturbances, however targets regarding the


size and configuration of fires were noted as future directions for the work.

Heuristic Model

14/18

Johnson, K.N., Sessions, J., Franklin, J., Gabriel, J. 1998. Integrating wildfire into

strategic planning for Sierra Nevada forests. Journal of Forestry. 96(1): 42
-
49.


"Safe Forests"
-

A gradient search heuristic used to schedule timber harvests


and fuel treatments with the hope of reducing the potential for high severity wildfires.

Elevation, slope, aspect, custom fuel models, and BEHAVE were used to estimate


the potential damage for each vegetation class based on flame length.


Process:

1. Given the objectives and constraints, develop a forest plan using the heuristic model.

2. Generate wildfires for all time periods from a probability distribution (amount, size,


location), up to an average amount of fire over a given period of time.

3. Burn polygons using their risk factor and the likely wind direction, and spread the


fire up to the desired size.

4. Decide whether to salvage stands immediately or to post
-
pone harvests.


Associated work:

Sessions et al. (1999)
-

used FARSITE


Heuristic Model

15/18

Kim, Y.
-
H., P. Bettinger, and M. Finney. 2009.
Spatial optimization of
the pattern

of
fuel
management

activities and subsequent effects on simulated wildfires.

European Journal of Operational Research. 197(1): 253
-
265.


Methods for scheduling forest management activities in a spatial pattern (dispersed,


clumped, random, and regular) are presented, with the intent to examine the effects


placement of activities have on resulting simulated wildfire behavior.

Both operational and fuel reduction management prescriptions are examined, and a


heuristic (great deluge algorithm) was employed to schedule the activities.


Process:

1. Develop a forest plan using the heuristic model, scheduling activities in patterns


across the landscape.

2. Generate wildfires using FARSITE code. Fifteen ignition points were randomly


selected for fires in the first decade. Each ignition point was then used to model


fires when different patterns of activities were assumed.

3. Assess impact on fire behavior and scheduled timber harvest volume.


Heuristic Model

16/18

Bettinger, P. 2009. A prototype method for integrating spatially
-
referenced wildfires

into a tactical forest planning model. Research Journal of Forestry. 3(1): 8
-
22.


A heuristic process was developed to illustrate the potential to incorporate a


feed
-
back mechanism for stochastic processes into the development of a forest plan.

The heuristic was 1
-
opt and 2
-
opt tabu search.

Fire was simulated using FARSITE code. Non
-
complete mortality was assumed.


Process:

1. Optimal stand
-
level management prescriptions are assigned to timber stands to


generate an initial plan of action that leads to the highest, and most even scheduled


timber volume.

2. The spatial location of wildfires are then simulated period
-
by
-
period using FARSITE code.


The number of fires per period was fixed, yet the location of ignition was random.

3. Timber stands that are affected by wildfire are re
-
assigned management prescriptions


based on their resulting vegetative condition, using a burn severity rating.

4. The scheduled wood volume is then re
-
evaluated from the period in question forward


through the time horizon.


Heuristic Model

17/18

Zuuring, H.R., Troutwine, J.M., Jones, G., and Sullivan, J. 2005. Decision support

models for economically efficient integrated forest management. 2005 ESRI

International User Conference Proceedings. 25 p.


MAGIS (Multi
-
resource Analysis and GIS)


A spatial harvest scheduling model that develops a treatment schedule based on the


hazard ratings from SIMPPLLE, TOM, MTT, and other socio
-
economic goals.

Linear programming, mixed integer programming, or heuristic methods are available.

Tactical, long
-
range planning is facilitated, and treatments account for disturbance


processes such as fire.


Using SIMPPLLE and MAGIS in an iterative manner can allow one to plan the optimal


size and location of forest treatments that are most effective in reducing the size of


wildfires.


Optimization / Heuristic

18/18

1. Do stand
-
level goals prevent the attainment of landscape
-
level goals?


2. Do landscape
-
level goals prevent the attainment of stand
-
level goals?


3. Can the threat (or impact) of fire be reduced through alternative landscape


management policies?


4. Can forest management activities planned at the stand
-
level affect the


behavior of large fires?


5. Does the arrangement (spatial and temporal) of management activities


affect large fire behavior?

Key questions recent literature has attempted to answer

1. How fires are accounted for in models


Strata reductions


Pixel
-
based growth and assessment


Vector
-
based polygon impacts



2. Whether fire impacts are randomly or deterministically estimated


When random, whether a statistical distribution is used, or whether a model is


employed to let fires burn


3. When fires are assumed to occur


Before harvests are scheduled


After harvests are scheduled

Differences in models

Do you

have any

burning

questions

or comments?