Forest classification - problem formulation and initial experiments

The focus of the research presented in this report is the challenge of computer-based forest classification of digitized infrared aerial photography. Interpretation of such demanding imagery has until now exclusively been performed manually. the current research is a first step in the direction of the goal of semi-automatic interpretation of forest imagery. The first objective of the work presented has been to identify and define the problems of forest mapping in an as concise may as possible and in terms of a remote sensing approach. The second objective has been to do initial data exploration and testes of standard image analysis methods. It has been determined that the geometrical distortions in an aerial photography will heavily influence the performance of automatic classification if the effects are not modelled and corrected. The discrimination experiments show that both texture and spectral features to some extent may be used to discriminate tree species and cutting classes. In data with low geometrical distortion, some GLCM textural features showed good discrimination ability. Based on the distribution of a set of ten derived spectral classes, we can discriminate properly between (i) spruce, cutting class 1; (ii) spruce and pine, cutting class 3; (iii) spruce, cutting classes 4-5; and (iv) deciduous forest.