Individual tree-based species classification in high spatial resolution aerial images of forests using fuzzy sets

Abstract

This paper presents an application of fuzzy set theory for classification of individual tree crowns into species groups, in high spatial resolution colour infrared aerial photographs. In this type of digital image, the trees are visible as individual objects. The number of individuals to classify might be very large in the acquired set of photographs, but the applied grade of membership (GoM) model, which this paper focuses on, is suitable for dealing with large datasets. The extent of each tree crown in the image is defined using a previously published procedure. Based on colour information (hue), an optimal fuzzy thresholding technique divides the tree crown universal set into a dominant set and its minor complement. Nine different features of each image object are estimated, and transformed using principal component analysis (PCA). The first three or four PCs are subsequently used in the GoM model. Furthermore, the concept of fuzzy relation is applied to one of the descriptors: to predict a centroid of the star-shaped pattern of Norway spruce. The GoM model needs initial membership values, which are estimated using an unsupervised fuzzy clustering approach of small subareas (branches in the tree crowns) and their corresponding digital numbers in each colour band (RGB-images). The complete classification system comprises three independent components: decisions on coniferous/deciduous, Scots pine/Norway spruce, and Birch/Aspen. The accuracies (ground patches excluded), using the supervised GoM model with crossvalidation, are 87%, 76%, and 79%, respectively. The accuracy for the compounded system is 67%.