The result in a three-band file with meadow, grass heath and bare rock in RGB (higher relative values are brighter). The regression tree result is on the left and linear regression on right. In the center of the picture is the highest mountain in the study area, where we expect to see high percentages of these cover types. 

Context in source publication

Context 1

... separate set of evaluation data consisted of 100 randomly selected plots within the study area. RMSE and bias was calculated for each individual class and as a total for both regression tree and linear regression results. The results are given in Tab. 3. When the resulting files from regression tree and linear regression are “hardened” (i.e., where the dominant class fraction determines the class label), regression trees have a higher percent correct in all classes compared to regression. The overall number correctly classified in this case with regression trees was 65% as compared to 61% from linear regression. An example of a three-band combination for a smaller area is given in Fig. 3. The visual result from regression appears to be “smoother” with graduated changes between pixels whereas the regression tree result has more discrete units of classes. The results from the accuracy assessment show that regression trees resulted in a slightly lower RMSE (20.1%) than with linear regression (20.6%), both overall and on a per class basis. The regression trees may have performed slightly better due to the inclusion of the ancillary data from the elevation model. However, linear regression shows lower per class bias. The trends for class accuracy are similar between the two methods. The class “other heath” has the highest RMSE, perhaps due to the fact that this class typically had the widest range of values in the training data (Tab. 1). As seen by the low R 2 value from the linear regression (Tab. 2), and by noting that the residuals were not distributed normally, it is likely that the combination of the two types of heath (dry and mesic) are too different to combine together into one type. Likewise with wetland, where a low R 2 value from linear regression and non-normally distributed residuals show the effect of combining three different wetland classes into one. It is also important to note that a 1:100 000 scale wetland mask was used to help in the Landsat classification and therefore, where wetland is present, it may be somewhat over-classified in the Landsat classification, and therefore in the input data. The negative bias for wetland may be a result of this. Within-class variability can affect the result negatively and should be minimized [24]. Vegetation types that typically had lower fractions represented within a pixel, such as grass heath (see Tab. 1), had larger errors in quantifying the higher fractions. This leads to the idea that a better representation of the class variability may be needed for a better result. Non- major classes which would not be well-represented in the training data should perhaps not be included in this classification. Large errors from minor classes can affect the overall accuracy of the other classes. Other studies [8, 12] show that the result from linear regression is influenced by the a priori information of the training sets. The class of water was not well-predicted considering its unique signature, and was perhaps under-represented in the training data. The narrow and low range of water’s DN in the visible and N-IR bands may also explain why it is not being distinguished in mixed pixels. Similar problems with water occurred in other studies [8, 15]. In regression trees, use of the MTCI band was found beneficial in three classes, specifically other heath, meadow, and mountain birch. These classes have a signature with a steep slope at the red-edge region. The regression tree worked better than a previously created hard classification for this area, where dominant classes tended to be over-classified and resulted in a 58% overall accuracy [25]. In comparison to others’ results when classifying multiple land cover types from regression trees (e.g., [15] with 16.43% overall RMSE from regression trees), the accuracies were quite comparable. This study has used seven classes, including a wetland class which is often a highly variable class that is difficult to classify. This study has not looked into the effect of “distant and proximate” training data because the study area was relatively similar. However, if working with the entire mountain chain, this issue would be relevant and should be taken into consideration. The results of the regression tree and linear regression were promising. Further investigation into the regression tree will be done, as this study was a preliminary test. Additional ancillary data can be investigated and incorporated into the training data. Improving the quality and supplementing the input training data may give a better result, although previous studies show the methods should be robust to training data errors [13]. Stratification of the study area may be beneficial. In addition, using a combination of classification methods, such as regression trees for mixed pixels and another method for the more extreme fractions (i.e., absent and pure), should be tried in producing an end product. The objective of this study was to investigate the use of MERIS full resolution data in mapping fractions of basic mountain vegetation classes using soft classification methods. The two methods which were tested and compared were regression trees and linear regression. The seven classes of bare rock, grass heath, other heath, meadow, wetland, mountain birch and water were classified. Regression trees resulted in slightly lower overall RMSE (20.1%) than linear regression (20.6%), although per-class bias was slightly higher for regression trees. These results are preliminary, and further investigation into variations in the regression tree method as well as improved input data, appears to offer a promising technique for fraction ...