A method for the segmentation of very high spatial resolution images of forested landscapes

Abstract

Pixel-by-pixel image analysis methods are not applicable when using VHR image material in multisource forest inventory applications. One possible solution to this problem is to define the units of image analysis by means of image segmentation. The paper presents a two-phase segmentation method (COS) based on segmentation in the feature space and co-occurrence region merging (CRM). The resulting segments were tested in the spectral feature extraction, and the estimation of plot-level total volume and volumes by tree species. The study material consisted of an AISA spectrometer image and 254 relascope field data plots. Two different segmentations were derived and the performance of segment-based feature sets were compared to that of a feature set extracted from the local neighbourhood of the field plots. Cross-validation techniques and a k-nn estimator were applied in the estimation tests. The estimation results which had the smallest rms errors were achieved with segment-based features in the cases of total volume and the volume of deciduous species. In the cases of pine and spruce, the features from the local neighbourhood performed best. The study suggests that COS produces segments which can be used as units for further image analysis.