Academia.eduAcademia.edu

Figure 1 – uploaded by Jón Atli Benediktsson

See full PDFdownloadDownload figure
Colorado data—information classes and samples Table 1 ments. It is interesting to note that variable four is the most important variable, followed by variable two and five. That is, elevation is the only topographic variable of importance. The variable importance for each individual class can be seen on Fig. 2. Some interesting conclusions can be drawn from this figure. For example, the classification of water re- lies mostly on the fourth spectral channel, and elevation variable is important for the classification of all the aspen classes. Also of interest, the slope variable is the second most important variable for the classification of Colorado Blue Spruce.

Table 1 Colorado data—information classes and samples Table 1 ments. It is interesting to note that variable four is the most important variable, followed by variable two and five. That is, elevation is the only topographic variable of importance. The variable importance for each individual class can be seen on Fig. 2. Some interesting conclusions can be drawn from this figure. For example, the classification of water re- lies mostly on the fourth spectral channel, and elevation variable is important for the classification of all the aspen classes. Also of interest, the slope variable is the second most important variable for the classification of Colorado Blue Spruce.

Related Figures (8)

Fig. 1. Colorado training data—(a) variable importance and (b) outlier analysis.
Fig. 1. Colorado training data—(a) variable importance and (b) outlier analysis.
Fig. 2. Colorado training data—variable importance for each of the 10 classes in a Random Forest.
Fig. 2. Colorado training data—variable importance for each of the 10 classes in a Random Forest.
Fig. 4. Colorado training data: outlier analysis for the individual classes. In each case, the x-axis (index) gives the number of a training sample and t! y-axis the outlier measure. Fig. 3. Colorado training data—variable importance for each of the 10 classes for a single CART tree.
Fig. 4. Colorado training data: outlier analysis for the individual classes. In each case, the x-axis (index) gives the number of a training sample and t! y-axis the outlier measure. Fig. 3. Colorado training data—variable importance for each of the 10 classes for a single CART tree.
Colorado data—confusion matrix for test data in Random Forest classification (using 500 trees and testing 3 variables at each node) Table 4
Colorado data—confusion matrix for test data in Random Forest classification (using 500 trees and testing 3 variables at each node) Table 4
Solorado data—confusion matrix for training data in Random Forest classification (using 500 trees and testing 3 variables at each node) Table 3
Solorado data—confusion matrix for training data in Random Forest classification (using 500 trees and testing 3 variables at each node) Table 3
Colorado data—classification accuracies for the Random Forest classifier Table 2 Table 5
Colorado data—classification accuracies for the Random Forest classifier Table 2 Table 5
Comparison of Random Forests to boosting and bagging with various base classifiers in terms of overall test set accuracies seen when Figs. 2 and 3 were compared. In terms of accu- racies, the Random Forest also outperformed bagging based on the 1R algorithm and boosting of the j4.8 deci- sion tree. On the other hand, the accuracies for the Ran- dom Forest algorithm, bagging with j4.8, and boosting of a decision table were similar, i.e., they were within a 1% difference of each other. The best overall accuracy was obtained by boosting the IR base classifier, resulting in more than 2% increase in accuracy as compared to the accuracy obtained by the Random Forest. Although the Random Forest does not in all cases achieve the accuracies of the boosting and bagging, the Random Forest algorithm has the advantage over boosting that it is much faster (boosting is iterative). Furthermore, Random Forest is in general considerably faster than bagging, depending on the base classifier used in bagging.
Comparison of Random Forests to boosting and bagging with various base classifiers in terms of overall test set accuracies seen when Figs. 2 and 3 were compared. In terms of accu- racies, the Random Forest also outperformed bagging based on the 1R algorithm and boosting of the j4.8 deci- sion tree. On the other hand, the accuracies for the Ran- dom Forest algorithm, bagging with j4.8, and boosting of a decision table were similar, i.e., they were within a 1% difference of each other. The best overall accuracy was obtained by boosting the IR base classifier, resulting in more than 2% increase in accuracy as compared to the accuracy obtained by the Random Forest. Although the Random Forest does not in all cases achieve the accuracies of the boosting and bagging, the Random Forest algorithm has the advantage over boosting that it is much faster (boosting is iterative). Furthermore, Random Forest is in general considerably faster than bagging, depending on the base classifier used in bagging.
Related topics:
Cognitive ScienceRemote SensingRandom ForestsLand cover classificationClassification and Regression TreeTHERMAL INFRARED REMOTE SENSING DATAElectrical and Electronic Engineering
580 California St., Suite 400
San Francisco, CA, 94104
© 2025 Academia. All rights reserved