Automated forest structure mapping from high resolution imagery based on directional semivariogram estimates

A new segmentation approach that allows forest stands identification on high spatial resolution (⪕1 m) optical imagery is presented. Texture information was first derived by measuring the range of the semivariogram of monochrome image values in three different directions using a moving window. The semivariogram ranges were then used to predict, on a per-pixel basis, three stand structure parameters through regression equations developed for crown diameter, stand density, and crown closure. A region growing algorithm was applied to these three regression estimate images to identify the limits of the forest stands. Calibration of the prediction equations was made using artificial images created by a geometrical-optical process. It was found that forest stands boundaries can be adequately identified on artificial images and that average forest structure estimates within each delineated stand are close to the actual values. Preliminary application of the proposed method to real images acquired with the MEIS-II airborne sensor yielded good segmentation and per stand structure estimates. Some errors were generated due to the fact that the moving window sometimes overlapped two different forest stands because of the presence of areas covered by nonforest vegetation or human made structures. The issue of the moving window size and means to increase the precision of the method are discussed.