Machine Vision Based Automation in Forest Inventory.Remote sensing saves money in forest inventoryAccurate forest inventory data saves money in the planning of forestry and forest industry. At the same time, the high costs of terrain inventory encourages attempt to employ remote sensing methods in inventory operations. More accurate remote sensing methods in measuring timber volume, breast height diameter and stem number can be based on recognition of single trees in aerial images. The University of Joensuu has developed a process for stand based semi-automated forest inventory. The project was funded by the Finnish Technology Development Centre. The research work has been carried out in the Department of Mathematics and the Faculty of Forestry in the University of Joensuu and the Laboratory of Space Technology in the Helsinki University of Technology. Pattern recognition enables single tree crown delineation Pattern recognition methods give tools to blob out single tree canopies from an image. An example of the results of pattern recognition algorithm are shown in the Image 3. In practical use, the segmentation proved to be reliable when the resolution of the image was better than 0.5m. Main steps in segmentation process:
After segmenting the area
of one stand from the aerial image, we used regression models to convert
the individual segmented tree crown areas to breast height diameters (dbh).
The data for the model was collected from Eastern Finland in 1998. It was
clear in advance that it is not possible in most cases to see the whole
tree crown from aerial images, because crowns cover each other and shadows
make it impossible to see the lower parts of the crowns. Therefore, we
used also regression model to calibrate the estimated dbh to field data.
After these steps we have the dbh distribution for one stand and we can
calculate stand characteristics like total volume, stem number, basal area
and tree height. In individual tree volume estimation we used the LaasasenahoÕs
(1982) two-parameter model and tree height was calculated using the NŠslundÕs
(1936) model.
 Figure 1. Left : the original
aerial image. The sample plot is marked with white broken line.
 Figure 2. Left : the Gaussian smoothed image and local maxima marked with red crosses. Right: Image after segmentation Table 1. Ground truth measured
on sample plot vs. semi automatic forest inventory based on pattern recognition.
A tool for semi-automated forest resource inventory and management Compared with field inventory, a substantially more efficient tool can be developed based on semi-automatic use of the algorithms developed. The role of the user is to regulate the steps of the algorithm and to control the results step by step. Some parameters have to be adjusted for different stand types. In forestry operations, remote sensing has been used for large-scale inventories, like land-use classifications. With traditional methods, the accuracy of the information is acceptable at the level of a whole municipality, but poor in small invention units like stands. Single tree recognition from high-resolution images gives a tool to expand the use of remote sensing to reliable single stand invention. The same process can be adjusted for different image data. Aerial images, aerial video and radar images have been tested during the project. At the moment, digital aerial false colour images form the most practical data source for inventory operations. Potential method for operational use in the near future In operative use, it is estimated than an accuracy of 25 % RMSE per hectare can be achieved using this method for timber volume, breast height diameter and growth density. After semi-automated product development, the methods are anticipated to penetrate quickly into the forest inventory market. Contacts:
Mikko Lehikoinen
Juha Hyyppä
|
