Towards Operative Forest Inventory by Extraction of Tree Level Information

The work related to this paper is part of an on-going study called NewForest - Renewal of Forest Resource Mapping. In this study the methodologies developed for individual tree crown (ITC) recognition and crown width estimation will be combined with forest variable estimates that are produced using features calculated from segmented VHR satellite image. A field visit to Karttula, Eastern Finland, was conducted to collect the class and geo-location information for 1164 ground objects (900 trees and 264 non-tree objects). These data were used for the classifier model and feature selection and for species classification accuracy assessment. For testing the classifier ability to predict tree species proportions, an independent set of 178 forest field inventory plots was used. Seven classes were defined: pine, spruce, deciduous, shadow, open area, bare ground, green vegetation. A modified Local maximum (LM) filtering technique was used for individual tree crown (ITC) detection. The spectral signatures of an ITC were sampled with a radius of r=1.5 m around the ITC brightest pixel (feature set A). Also a set of 9 contextual features were extracted from circular neighbourhood (r=7.25 m) around the ITC (feature set B). A classifier model and feature selection was performed. A 5NN classifier provided the best overall performance in tree species classification in terms of classification accuracy and generalization. The overall classification accuracy for the seven classes was 73.8% with feature set A using 5NN classifier. With feature sets A and B combined the accuracy was 74.1%. The average RMS errors in species proportion prediction were 2.6% with feature set A and 2.5% with feature sets A and B combined.