APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND
BASAL AREA MAPPING USING SATELLITE IMAGERY AND FOREST
INVENTORY DATA

George Xian, Zhiliang Zhu
Raytheon, USGS EROS Data Center
47914 252nd Street, Sioux Falls, SD 57198
xian@usgs.gov, zhu@usgs.gov

Michael Hoppus
US Forest Service
Northeastern Research Station
5 Radnor Coporate Center, 100 Matsonford Rd – Suite 200, Radnor, PA 19087

Michael Fleming
Raytheon, EROS Data Center
Alaska Field Office
fleming@usgs.gov


ABSTRACT

Accurate, current, and cost-effective fire fuel data are required by management and fire science communities for
use in reducing wildland fire hazards over large areas. In this paper we present results of applying decision-tree
techniques to mapping vegetation parameters (such as vegetation types and canopy structure classification) required
for fire fuel characterization. Specifically, we present preliminary results of mapping forest types and average basal
area by different forest types at 30-meter resolution. Input data into the decision tree model included Landsat-7
ETM+ spring, summer and fall greenness, brightness and wetness of the tasseled cap transformation, topographic
data layers such as slope and elevation, and forest variables measured on inventory plots in the Mid -Atlantic region.
Using decision-tree models, eight forest types were successfully identified in training cases and mapped for the
entire mapping area. Forest basal area per unit area (conifer and deciduous) was estimated as well using regression-
tree models. Cross-validation conducted for both forest types and basal area showed that discrete forest type
estimation error was 35% and continuous basal area relative errors were between 58 and 72%. Accuracy was higher
in homogeneous forested lands and lower in areas with fragmented forest cover. The study demonstrated that
decision tree and regression tree methods are efficient for large-area vegetation mapping if sufficient large-amount
of reference data are available.


INTRODUCTION

Fire fuel data are important for wildland fire suppression planning and hazard assessment. Research in complex
fire behavior and fire effects has the need for accurate fire fuel classification and estimation over large area.
Mapping fire fuels requires estimation of vegetation structure. Remote sensing data have been used to estimate
forest structure for almost two decades. Successful works include modeling canopy optical-geometric properties (Li
and Strahler 1985; Li and Strahler 1992). Many researchers have used the modeling method with remote sensing
data and classification algorithms to estimate vegetation cover. These include using Landsat TM data and Li and
Strahler’s Geometric-Optical model to estimate crown cover projection, canopy size, and tree density in Southeast
Queensland Australia, with the overall mapping accuracy of 67.4% (Scarth and Phinn 2000); the California multi-
attribute vegetation mapping for forest service lands in California using TM data and digital elevation model
(Franklin et al. 2000). However, canopy models are computationally intensive, species specific, and often are not
inversable. To effectively classify land cover from remotely sensed data, many researches have adopted tree-based
techniques such as decision tree and regression tree. Successful effects include Friedl and Brodley (1997), Borak
and Strahler (1999), and Friedl et al. (1999). Therefore there is great interest in testing the tree-based techniques for
APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

mapping certain vegetation cover and structure variables for use in large -area fire fuels characterization. In this
study, we tested decision tree technique and its derived method – regression tree to map spatial distribution of forest
types and forest basal areas. This research was conducted in a mapping area located in the mid -Atlantic region that
is also USGS mapping zone 60. Seasonal vegetation data obtained from Landsat 7 images and topographic
information were used in the research.


MAPPING FOREST TYPES USING TREE-CLASSIFICATION

Using a commercial tree-classification package, C5.0 (www.rulequest.com), we applied decision tree classifier
(Quinlan 1993; Freund and Schapire 1996; Qu inlan 1996; Friedl et. al. 1999) to mapping forest types defined by the
Society of American Foresters (Eyre 1980). As part of this research effort, scientists traveled to Forest Service to
extract relevant forests inventory data, merged the data with sampled Landsat 7 tassel cap transformation bands,
then removed all geographic coordinates to build up a training data set. A recent study (Huang et. al. 2002) showed
that tassel cap transformation from Landsat 7 at satellite reflectance was a useful tool for compressing spectral data
into a few bands associated with physical scene characteristics. In our research, we applied the tassel cap spring,
summer and fall brightness, greenness, wetness transformed from Landsat-7 ETM+ reflectance band 1-5 and 7.
Topographic attributes included slope and digital elevation model. Eight different forest types were classified. The




















Figure 1 The model forest type classification on the mapping zone


boosting technique (Quinlan 1996; Friedl et. al. 1999) for improving classification accuracy was adopted in the
process. To evaluate predictive accuracy of the classifier, we used f-fold cross validation for training cases. All cases
were divided into f blocks of the same size and class distribution. A classifier in a block was constructed from the
cases in the remaining blocks and tested on the cases in the hold -out block. The error rate of a classifier produced
from all the cases was estimated as the ratio of the total number of errors on the hold -out cases from all folds to the
total number of cases. This gives an indication of how a decision tree trained from this data would be expected to
perform on independent data or the robustness of the model. Huang et. al. (2001) has shown good agreement
between cross validation accuracy and independent test data accuracy. The cross-validation result showed that the
estimation accuracy was 65.2%, and the standard error was 2.4%. Based on the training cases results, a decision tre e
model was formulated from a number of decision rules and used to estimate forest type distribution in this region.
Figure 1 is the forest type classification on the mapping zone from the decision tree model estimation. The figure
shows that Loblolly and Shortleaf, Oak/Pine zone and Oak/Hickory groups cover most area in this region. Other
types of tree can be seen on the mapping zone as scattered patches.

APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

FOREST BASAL AREA ESTIMATION USING REGRESSION TREE TECHNIQUE

Forest inventory data include several forest type basal area values (square feet per acre) for the mapping zone.
We display three dimensional plots of coniferous and deciduous basal area values in this region in Figure 2 and
Figure 3, respectively. The values of coniferous basal area (CBA) are high in both northeast and southwest sides.
Deciduous basal area (DBA) has two apparent large value peaks in both central south and north regions.


309.70
232.27
154.85
77.42
0.00

Figure 2 3-D surface plot of Forest Inventory database: Conifer basal area


228
171
114
57
0

Figure 3 3-D surface plot of Forest Service Inventory database: Deciduous bas al area

To predict continuous forestry basal area using remote sensing data, we chose a modified regression-tree-based
predictive model, Cubist (www.rulequest.com), that is a sister system of decision-tree classification model. The
modified regression-tree-based model estimates a case's target value in terms of its attribute values by building a
model containing one or more rules, where each rule is a conjunction of conditions associated with a linear
expression. A rule indicates that, whenever a case satisfies all the conditions, the linear model is appropriate for
predicting the value of the target attribute. The rule -based predictive model thus resembles a piecewise linear model
(except that the rules can overlap) based on training data.
APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

In this study, we set out to predict values of deciduous and conifer basal areas. Training data included onsite
deciduous and conifer basal area measurements from Forest Service inventory database; Landsat 7 tasseled cap
greenness, wetness and brightness; slope and digital elevation model data. To improve the model prediction
accuracy, booting was also adopted for constructing more accurate regression model.

Deciduous tree basal area prediction

The regression tree rules were developed for target values based on attributes from training data set. A total of
twenty-seven rules were generated. These rules included almost all attributes listed in the training data. After rules
of the target value were built for all conditions, the linear model was applied to predicting values of target attributes
based on each rule conditions and Landsat 7 ETM data associated with geographic slope and elevation over the
whole mapping region. All target values were estimated for 30 meter resolution pixels in the region. The model
prediction of deciduous forest basal area over the region was plotted in Figure 4. Most large values of basal area
were located around the central and north parts of the mapping zone where the land slope and altitude value are
relatively high.




















Figure 4 Model estimation of Deciduous Basal Area

Coniferous basal area prediction

From the Forest Service inventory data plot in Figure 2 we know that coniferous forest basal area has relative
large values in the northeast and southwest parts of the mapping zone. To estimate the magnitude of CBA over the
region, we used the same regression tree model and data set used for deciduous trees. The only change in the model
was that the target attribute became CBA. Running from training data set, a rule-based predictive model was built up
with twenty-six rules for the target value. Then, the model was applied to the entire region using prepared Landsat 7
ETM+ images. The model prediction of CBA over the region was displayed (Figure 5). Large values in both
southwest and northeast regions were very apparent. Generally, the linear regression model prediction showed that
the distribution of modeled basal area values were very close to Forest Service Inventory data shown in Figure 2.
The lack of coniferous forests in northwest region of the mapping zone is apparent. In the southern part of the
mapping zone, coniferous and deciduous forests mix with relative large basal areas for both trees.




APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

Figure 5 Model estimation of Coniferous Basal Area

Validation of model prediction


The accuracy of the model prediction needs to be assessed. The prediction of the regression tree model on the
training cases from which it was constructed was evaluated by calculating the regression tree model average error,
relative error, which is the ration of the average error to the error that would result from always predicting the
meaning value, and R2. These estimations were obtained from training examples. We needed to validate the
predictive continuous variables of the model on new cases. Similar with the decision-tree classification cross
validation to assess predictive accuracy, we used f-fold cross-validation here to obtain a reliable estimate of
predictive accuracy. The existing training cases were divided into f blocks of nearly the same size and target value
distribution. The accuracy of a model produced from all the cases was estimated by averaging results on the hold -out
cases.
The accuracies of regression tree model calculation and prediction were shown in Table 1.

Good model performance requires the relative error value to be less than 1. From our training cases, the model
relative error was 0.40 and 0.50 for conifer and deciduous forest basal areas, respectively. The R2 reached 0.76 for
conifer and 0.67 for deciduous. The result suggested that the model had relative higher consistency for conifer than
for deciduous over the whole mapping region. The cross validation results showed that R2 equals to 0.46 for conifer
forest and 0.38 for deciduous forest.

Table 1. Accuracies of regression-tree model prediction
Items
Tree model
Tree model R2
Cross validation
Cross validation
relative error
relative error
R2
Conifer Basal
0.40
0.76
0.58
0.46
Area
Deciduous Basal
0.50
0.67
0.72
0.38
Area

Comparison between K-NN and regression-tree model


The k-nearst neighbor (K-NN) classifier is a simple classification method that has been found effective for
mapping forest structure variables (Michie, 1994; Tomppo, 1996). Generally, K-NN method estimates a field
parameter at a particular location by a weighted average of k nearest neighbor (known) pixels. The weighted
function is distance based. The number of nearest neighbors, k, is selected such that the k nearest distances relative
to the pixel location are within this chosen range. One of authors had applied this technique to a small area in the
APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

central east region of the mapping zone. The forest basal areas were estimated using K-NN and Forest Inventory
database. Here, we compared the values of model predictions for both conifer and deciduous forest basal areas from
regression tree and K-NN models with Forest Inventory field data. A linear model fit generated the linear residual
standard error and R2 values from two technique predictions associated with real values.

Table 2. Comparison between two model prediction and Forest Inventory data
Residual Standard Error
R2
Estimation
K-NN
Tree-Model
K-NN
Tree-Model
Conifer Basal Area
25.29
24.23
0.2622
0.4969
Deciduous Basal Area
23.97
18.63
0.6124
0.7659






Table 2 presents comparison of results between regression tree and KNN model predictions and observed
values using linear regression fit. The regression tree model had smaller residual standard error and large R2 values
than K-NN for the same forest type. The regression tree model predictions had less standard errors than KNN.
Figure 6 and 7 are linear regression fit plots for conifer and deciduous forests predictions, respectively. We also
noticed that deciduous forest basal area predictions in KNN and regression tree models had standard errors of
0.0984 and 0.0846, respectively. The conifer forest basal area predictions had standard errors of 0.1233 and 0.1181
from KNN and tree model, respectively. The regression tree model had larger R2 values and less standard errors than
K-NN model. Deciduous basal area prediction was closer to the observed measurements in this area. This result was
different from the cross examination we discussed in section 3.3 where deciduous prediction had less accuracy than
conifer. This difference might be caused by the difference in the locations where the models were applied. The
comparison data were from a small region where deciduous forest was majority vegetation cover. The model
prediction might be in favor to it in this small area. But for the entire mapping region, the average result could be
different.



150
120

80
120

60
100
80
80
40
Dtree.BA.CON
KNN.BA.CON
Dtree.BA.DEC
KNN.BA.DEC
50

40
20
40


0
0

0
0
10
60
110
160
10
60
110
160

BA.CON.FIA
BA.DEC.FIA
Dtree.BA.DEC
Dtree.BA.CON
KNN.BA.DEC

KNN.BA.CON


Figure 6 Comparison of conifer basal area Figure 7 Comparison of deciduous forest
prediction from Regression-Tree and K-NN basal area prediction from Regression-Tree
models to the Forest Inventory data and K-NN models to the Forest Inventory data




APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

CONCLUSIONS

Forest types and average basal area, required for fire fuels characterization, were mapped in USGS mapping
zone 60. The decision tree approach was used in this study. The training data included Landsat 7 data and Forest
Inventory database. Eight different forest types were classified in the region. The Loblolly and Shortleaf Group
forests were apparent in the east and southeast part of the region. Oak/Hickory Group covers most of north and
northwest region. The predictive accuracy of decision tree model from training data set reached 65.2%. By using the
same Landsat imagery and Forest Service Inventory database basal area on site measurement data, we used the
regression tree model to estimate the deciduous and conifer forests basal area values over the whole mapping region.
The regression tree model prediction showed that most of large basal area values (>100) of deciduous forest were in
the northwest and northeast of the region. The result was consistent with the model classification conclusion and
Forest Service Inventory data. The conifer forest basal area estimation showed that most of large values were in
southern part of the region. In general, the conifer forest basal area prediction was more accurate than the deciduous
forest results.
We also compared regression tree model results with results obtained from K-NN model in a small area in
central-east part of the mapping region. Two model predictions were compared with Fores t Service inventory data
using linear regression fit. The regression tree model results were closer to the Forest Service Inventory data.


ACKNOWLEDGMENT

The authors would like to thank US Forest Service for supplying Forest Service Inventory database. This study
was made possible in part by the Raytheon Corporation under US Geological Survey contract 1434-CR-97-CN-
40274.


REFERENCES

Eyre, F.H. (Ed) (1980). Forest Cover Types of the United States and Canada. Society of American Foresters.
Franklin, J., C.E. Woodcock, and R. Warbington (2000). Multi_attribute Vegetation Maps of Forest Service Lands
in California Supporting Resource Management Decisions. Photogrammetric Engineering & Remote
Sensing 66(10): 1209-1217.
Freund, Y. and R.E. Schapire (1996). A decision-theortic generalization of on-line learning and an application to
boosting. 2nd Euro. Conf. EuroCOLT96
Friedl, M. A., C.E. Brodley, A. H. Strahler (1999). Maximizing Land Cover Classification Accuracies Produced by
Decision Trees at Continental to Global Scales. IEEE Transactions on Geoscience Remote Sensing 37(2):
969-977.
Huang, C., L. Yang, C. Homer, M. Coan, R. Rykus, Z. Zhang, B. Wylie, K. Hegge, Z. Zhu, A. Lister, M. Hoppus, R.
Tymcio, L. DeBlander, W. Cooke, R. McRoberts, D. Wendt, and D. Weyerman (2001). Synergistic Use of
Data and Landsat 7 ETM+ Images For Large Area Forest Mapping. Thirty Annual Midwest
Mensurationists Meeting and Third Annual Forest Inventory and Analysis Symposium. Traverse City, MI.,
USA.
Huang, C., B. Wylie, L. Yang, C. Homer, G. Zylstra (2002). Derivation of a tasselled cap transformation based on
Landsat 7 at satellite reflectance. International Journal of Remote Sensing
Li, X., and A. H. Strahler (1985). Geometric-optical modeling of a conifer forest canopy. IEEE Transactions on
Geoscience Remote Sensing. GE-23, 705-721.
Li, X., and A. H. Strahler (1992). Geometrical-optical bidirectional reflectance modeling of the discrete-crown
vegetation canopy: Effect of crown shape and mutual shadowing. IEEE Transactions on Geoscience
Remote Sensing GE-30: 276-292.
Quinlan, J. R. (1993). C4.5: Programs for machine learning. Morgan Kaufmann Publishers.
Quinlan, J. R. (1996). Bagging, boosting, and c4.5. 13th National Conference of Artificial Intellegence, Portland,
OR.
Michie, D., D.J. Spiegelhalter, and C.C. Taylor (editors) (1994). Machine Learning, Neural and Statistical
Classification. Ellis Horwood, Hertfordshire, UK, 289p.
APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

Scarth, P., S. Phinn (2000). Determining Forest Structural Attributes Using an Inverted Geometric-Optical Model in
Mixed Eucalypt Forests, Southeast Queensland, Australia. Remote Sensing Environment. 71: 141-157.
Tomppo, E. (1996). Multi-source National Forest Inventory of Finland: New Thrusts in Forest Inventory.
Proceedings to the New Zealand Preharvest Inventory. Scandinavia Journal of Forest Research, 14, 182-
192.




>168

APPLICATION OF DECISION-TREE TECHNIQUES TO FOREST GROUP AND BASAL AREA
MAPPING USING SATELLITE IMAGERY AND FOREST INVENTORY DATA
Pecora 15/Land Satellite Information IV/ISPRS Commission I/FIEOS 2002 Conference Proceedings

---

Document Outline

• Return to Main Menu
• =================
• Search Titles
• Search Authors
• Search Sessions
• ================
• Next Page
• Previous Page
• =================
• Pecora 15 Papers
  • Table of Contents
  • Author Index
• ISPRS Papers
  • Table of Contents
  • Author Index
• FIEOS Papers
  • Table of Contents
  • Author Index
• =================
• Search CD-ROM
• Search Results
• Print
• =================
• Help
• Exit CD

---