Automatic individual tree crown delineation using a valley-following algorithm and a rule-based system

In remotely sensed aerial images or digitized aerial photographs of high sp atial resolution (10-100 cm/pixel) tree crowns are typically visible as alm ost distinct items. To produce semi-automatic forest inventories from such data (even stand-based inventories), it is better for the computer to deal directly with this predominant structural element of the images. A first ob vious step is thus, the delineation of the individual tree crowns, preferab ly in a completely automatic fashion. This step is to be followed by an ind ividual tree crown species recognition and then, a regrouping (if need) int o relatively conventional forest stands. Such a comprehensive approach can provide very detailed forest inventories, specially considering that regrou pings can be organized to follow criteria other than that of the present in ventories. This paper addresses the automatic individual tree crown delinea tion process. The individual tree crown delineation is accomplished by first following th e valleys of shade that typically exist between the crowns of dense forest stands. The image from a selected spectral band is first smoothed. Areas th at are completely in the shade are masked out by thresholding and local min ima are found in the remaining areas, except for non-forested areas that we re masked out a priori. From the local minima, valleys are followed to othe r local minima. Visually, this delineates most coniferous tree crowns rathe r well. However, not all tree crowns are fully delineated and tree clusters still exist, but most importantly, the process does not yet have the conce pt of tree crowns as distinct elements. This "awareness" is made possible b y the subsequent use of a rule-based system which attempts to follow in a c lock-wise fashion the internal boundaries of each individual tree crown (IT C) until a distinct closed area is obtained. Additional rules attempt to re cognize tree clusters and separate them into ITCs. From there, additional I TC-based processes, such as an ITC-based classifier (Gougeon et nl., 1988), are used to produce detailed information about the given forested area. This paper will first describe briefly the valley-following algorithm and t he rule-based system and then, comment on their strengths and weaknesses, a nd potential remedies to these weaknesses. Examples will be shown throughou t. The paper concludes with a list of possible improvements and future deve lopments.