- A
bid
S. *
PEE NE ONE
i.
Reference
classified | forest | Non-forest | > points | % correct
forest 547 44 591 93
non-forest | 2 403 405 100
> points | 549 | 447 996
% correct | 99 90 95
Table 1: Error matrix: forest / non-forest
Flat Terrain Mountainous Terrain
Forest Types
Deciduous Forests > 90% 85 - 90%
Spruce > 90% 85 - 90%
Pine 70 - 80% to be investigated
Spruce/Pine 60 - 70% to be investigated
Mixed Coniferous / Deciduous 80 - 90% 80 - 90%
Larch and Larch / Spruce to be investigated | 85-90 %
Swamp Forest 80 - 90% Does not occur
Natural Age Classes
Coniferous: 3 classes 80 - 90% 80 %
Deciduous: 3 classess 75 - 80% to be investigated
Forest density Classes
Coniferous 5 classes (20% steps) | > 85% 75 - 80%
Coniferous 3 classes (30% steps) | > 90% >85 %
Schardt, Mathias
Table 2: Producer Accuracy of the classification of different forest types
2 POTENTIAL OF SATELLITE REMOTE SENSING FOR FOREST INVENTORY
The feasibility of satellite remote sensing for the classification of forestry parameters has been proven in the past by several
studies. The following examples selected from the above listed projects are based on Thematic Mapper and SPOT and
shall demonstrate the potential of available remote sensing data at the assessment of different inventory relevant forest
parameters on a scale of 1 : 50.000 to 1 : 100.000. These studies were related to the classification of the parameters
forest / non-forest, tree species types, age classes and forest density. For the classification of forest types the maximum
likelihood method and for the stratification of forest and non-forest areas the threshold-level procedures was applied. The
verification of the classification result was performed by means of independently selected verification areas.
Forest/Non-forest
Investigations in alpine regions on the stratification of Forest and Non-Forest Areas on the base of SPOT / PAN data
has shown that an accuracy of more that 95% can be obtained. The nomenclature definition of forests in this case
defines forest as an area with more than 10% crown cover. The accuracy assessment was performed by comparing the
forest mask with 1054 sample points. These sample points (in a regular grid of 100m by 100m steps ) were visually
categorized into forest / non-forest / forest-border by the use of CIR-ortho- photos. Table 1 displays the error matrix for
the forest / non-forest points.
Comparable results have been achieved in flat and mountainous test sites have been achieved by investigations carried out
by (M. Schardt, 1990) and (Keil et al., 1990)
Tree species types, age classes and density classes
Investigations carried out by (M. Schardt, 1990, M. Schardt, 1995) in four German test sites as well as the experiences
made with the forest mapping in Alpine regions carried out within the EU-projects ALPMON and SEMEFOR (Schardt
and Schmitt, 1996) proved that the classification of the major tree species groups is possible with a satisfactory accuracy.
The separable forest types and the average approximate accuracy values achieved in these projects are summarized in
table 2:
Shortcomings
According to the positive experiences in the field of satellite based forest classifications quoted above it can be stated that
satellite remote sensing is an appropriate tool to inventor and monitor forest parameters at a small scale with reasonable
accuracy levels. However, in consideration of intensive large scale inventories on stand and district level some significant
shortcomings were identified when using Thematic Mapper and SPOT data:
International Archives of Photogrammetry and Remote Sensing. Vol. XXXIII, Part B7. Amsterdam 2000. 1317