MULTIVARIATE ANALYSIS OF FOREST CHANGE AT AN ABANDONED MINE SITE USING AIRBORNE DIGITAL CAMERA IMAGERY

摘要

Forest structure data in research plots at an abandoned mine site, near Timmins, Ontario, demonstrated a significant spatial gradient with distance from an open taihngs area as well as significant changes between 1997 and 1999. Research is ongoing to develop image-based models for mapping and monitoring of these spatial and temporal dynamics. This paper presents a multivariate modelling methodology to develop a forest health index from spectral, textural and radiometric fraction image data extracted from multispectral airborne digital camera imagery. Canonical correlation analysis was used to relate principal components of forest structure and image variables acquired in 1997 (R= 0.727). The set of forest structure variables that comprised the first canonical variate was used as the forest health index. An image-based forest structural health model was then developed using stepwise regression with the health index as the dependent variable and the original images variables as independents. Image texture, texture variation, and mean plot brightness, as well as sunlit crown fraction were included in the best image model, which was capable of predicting 66% of the variation found in the forest health index. The model was applied to the 1999 airborne imagery to predict 1999 the health index. The 1999-1997 differences in the image-based index were analyzed and compared to the actual 1999-1997 differences in the forest health index determined from field data. The results showed that the model predicted positive health change and no health change well, but not increased damage. This was because as damage increases, greater forest opermess results in increased light penetration and flush of herbaceous and shrub vegetation, which adds to the signal from the damaged overstory. Further research is needed to separate these two signals in highly damaged areas.