Resource data, such as, stems per hectare, tree crown diameter, and tree species, are useful for forest inventory and planing. The collection is mainly dominated by interpretation of aerial photographs combined with field work or by field work alone. Such processes are very time consuming and labour. Image analysis, with its automatic or semi-automatic methods, can make the interpretation become more efficient.
A summary of my PhD project made at Centre for Image Analysis, Uppsala, Sweden.
Tree crown extraction (see link above) can only be done if there are any tree crowns to extract. In an aerial image which covers a large region there are often non-forested areas included in the image. These areas must be excluded before the extraction process can start. Moreover, the results shown above can only be achieved if the forest is dense enough. If that is not the case, there will be a lot of false positives. However, a method that can handle sparse forests may not necessarily (most likely can not) be able handle a dense forests. Thus, there is also a need to classify the forest as sparse or dense and choose the extraction algorithm according to the outcome.
Before a classification of the forest can be done it is necessary to divide to image into different parts or segments. Ideally, there should be one segment for each stand as well as for the each non-forested area. However, the ideal case is very hard to achieve and since there is a second step involved (the classification step) a perfect segmentation is not necessary.
A multi-resolution region growing method is used to segment the image into regions with similar colour. The classification is based on a k-means clustering. The image below shows (from left to right) the original image provided by IFN, the French forest inventory, the segmentation result, and the classification result. The brighter the regions are in the classification the denser is the region (in sense of forest). Black regions contains no forest at all.
Last modified: Fri Aug 25 18:20:34 CEST 2006