Application of ROC curve to assess pixel-based classification methods on UltraCam-D aerial imagery to discriminate tree crowns in pure stands of Brant`s oak in Zagros forests

Abstract

Sustainable forest management in Zagros Mountains entails accurate information on tree crown density, which could be possibly derived from remote sensing data. Moreover, those remote sensing products need to be objectively evaluated. In this study, the results of three pixel-based classifiers of UltraCam-D aerial imagery were evaluated for classifying Brant`s oak (Quercus brantii Lindl.) crowns in Zagros forests in western Iran. This was carried out by means of receiver operating characteristic (ROC) curve. Therefore, a 30 ha plot was selected in pure Brant`s oak stand, in which the location and crown area of all trees were mapped. The UltraCam-D aerial imagery was classified by maximum likelihood (ML), artificial neural networks (ANNs) and support vector machines (SVMs) classifiers. The classification results were then evaluated by ROC curve and were presented by overall accuracy and Kappa coefficient. Results showed that the ML classified returned the largest area under ROC curve of "tree crowns" (0.894), whereas the lowest rate was found for SVM classifier (0.819). Sensitivity and specificity of "tree crowns" in ML classifier (0.999 and 0.999, respectively) were higher than those in two other classifiers. Although the precision of SVM classifier was the highest in discriminating "tree crowns" (1.000), the achieved accuracy of tree “crown class” was higher for ML classifier (0.999). This study concluded that using ROC curve enables an evaluation accuracy and precision of common pixel-based classifiers of such aerial imagery to discriminate tree crowns.

Keywords

• Tree crown
• Zagros
• pixel-based classifiers
• ROC curve
• UltraCam-D