Stratification of a Forest Area for Multisource Forest Inventory by Means of Aerial Photographs and Image Segmentation

Abstract

In Finland, the data for forest management planning has been gathered by stand-level visual field inventories. From the viewpoint of forest inventory and monitoring, this kind of a method has several drawbacks: the delineation of the stands is subjective and the stand borders tend to change during the planning period. As an alternative, a two-phase plot sampling has been suggested, where first phase sample plots are generated and stratified on the basis of remote sensing images and other auxiliary data sources. The second phase sample is allocated to these strata and measured in the field. The gathered information is generalised to the first phase sample plots using suitable estimators. The method has been applied with digitalized aerial photographs by extracting spectral features for first phase sample plots from square windows surrounding them. This method is unable to fully exploit the spatial resolution of aerial orthophotos. As an alternative we suggest image segment-based feature extraction and stratification. The main hypothesis is that the spectral strata produced by segments are more homogenous in relation to the actual stand characteristics than strata derived by the sample plot-based method. The hypothesis was tested using plot-and segment-based approaches as the basis of the stratification. Spectral average and standard deviation features were extracted from square windows surrounding the plots and from image segments. Two different segmentations were tested. The results from two study areas (S 1 and S2) were contradictory. The segment-based stratification produces spectrally more homogeneous segments in both study areas, but it did not result in more homogeneous strata in forest characteristics. In S1 the plot-based approach gave better results than the segment-based approaches. In S2, segment-based approach performed better if only spectral average features were employed in the stratification. The introduction of spectral standard deviation features to the analysis significantly improved the performance of plot-based approach.