Fa. Gougeon et al., Individual tree crown species recognition: The Nahmint study, AUTOMATED INTERPRETATION OF HIGH SPATIAL RESOLUTION DIGITAL IMAGERY FOR FORESTRY, INTERNATIONAL FORUM, 1999, pp. 209-223
The species content of forest stands is an information of paramount importa
nce in conventional forest inventories. Typically, the stands and their con
tent are assessed by human interpretation of aerial photographs. However, u
sing remotely sensed aerial images or digitized aerial photographs of high
spatial resolution (10-100 cm/pixel), it is now becoming possible to automa
tically delineate most of the visible individual tree crowns (ITC) in such
images. This led to the development of several ITC multispectral signatures
and of an ITC-based supervised classification system making possible ITC s
pecies recognition. The resulting information on the individual trees can b
e preserved, where very detailed information is needed; or collated, to gen
erate very precise information on existing forest stands, or regrouped (sta
tically or dynamically) using new criteria, to help with multi-resource for
est management.
This paper primarily addresses the species recognition aspects of this new
paradigm for generating precise information useful to forest inventories. T
he ITC-based delineation and classification system is tested with a geometr
ically corrected 60 cm/pixel casi image of the Nahmint Lake species demonst
ration area, Vancouver Island, British Columbia. Simple correction curves t
o compensate for bi-directional reflectance function (BRDF) effects were ap
plied to the multispectral image in spite of the fact that the image had be
en previously geometrically corrected by the supplier. The ITC-based superv
ised classification of five western Canadian coniferous species and a gener
ic hardwood class led to an overall classification accuracy of 59.8% when t
ested with a conventional confusion matrix approach, These low classificati
on results are attributed to the lack of purity of the training and testing
areas. A comparison of the species content of more sizable testing areas w
ith their corresponding field transects led to an overall error of 12%, 19.
5% when only the dominant species is considered. The paper concludes with a
discussion of the research and operational problems to be resolved before
the goal of semi-automatic generation of precise forest management inventor
ies is achieved.