AAIR Project No PL 94-2392

Intergroup Discussion
 Model Testing

Discussion Group - David, Erik, Uva, Chris, Barry, James, Roger, Andrew

n.b. this information probably needs checking by someone. I wasn't around for this, so I don't really understand the notes.

Aspects of Model Testing:

Data: Resolution (Fire 30m)

Model: Mechanism/parameter

Geography: Mechanistic and Logistic Models
    Test Finnish model in UK     Test Swedish model in Finland

Operation: Use by Forest Managers (including communication of model meaning

Presence/Abundance:Cell X Cell, Unit X Unit, time - binary characteristic profile

Measure of success: better than random

Method: Expert opinion, statistical...

Perception/human: Tolerance of risk

Statistical: residuals; confusion matrix; cross-validation; spatial sampling; jack-knife test

Participant Models and Approaches

CNIG
1. Rothermel: Use fire risk maps to test assumptions
      Input data: Some from models (linear mixture model, rms error for map)
      LAI: 30 trees

2. Pixel-homo (30 or 90m) Mixture Model
      Aggregations - sensitivity to number/content/type (3X3 mean)

3. LAI - depends on variability

NRS
1. Mechanistic Model (Barry)
     Internal consistency
     - bending moment with respect to obs
     - spacing and bending moment with respect to obs
     Intra-model calibration (Treesnap V HWIND)
Finnish data <-> UK model
Species V Soils (which combinations exist in different countries and for which do data exist)

Spruce HWIND Logistic Model Treesnap
Pine HWIND Logistic Model Treesnap
Birch HWIND Logistic Model Treesnap(?)

Assumption: Tree sizes -x% agree
           species
           soil type

2. Validation/Verification Operationality (Chris)
Sample/Stratification
           - predicted V Actual
           - Space: cell by cell / area by area

  Actual
Predicted
  no yes
no 237 17
yes 12 23

- Time: windspeed/cell

Umeå
Q: What is a good model?
A: Model better than random

Select variables:
- pre-selection and
- statistical

Significance of variables - statistically

Outputs: geography - spatial correlation
Cross-validation:
- presence in plot
- selection of threshold w.r.t. damage

Residuals:

  Predicted
Actual Damage
0    1
  62%
53%
56%		   1099
  1   52%
4.2%
4.3%		 88

Private

| Administration | Data | FTP site (MLURI) |

Open

| Description | Objective | Participants | Photo Gallery | WWW Resources | PROJECT home page |

Marianne Broadgate - m.broadgate@macaulay.ac.uk

Last modified: Tue Aug 13 11:54:08 BST 1996