Date of this Version
Presented at the 26th Canadian Symposium on Remote Sensing in Wolfville, Nova Scotia, June 14-16, 2005
Multispectral satellite imagery are appealing for their relatively low cost, and have demonstrated utility at the landscape level, but are typically limited at the stand level by coarse resolution and insensitivity to variation in vertical canopy structure. In contrast, lidar data are less affected by these difficulties, and provide high structural detail, but are less available due to their comparatively high cost. Two forest structure attributes measured at the plot level, basal area and trees per hectare, were predicted using stepwise multiple regression on 40 predictor variables derived from discrete-return lidar data (2 m post spacing), Advanced Land Imager (ALI) multispectral (30 m resolution) and panchromatic (10 m resolution) images, and geographic X,Y,Z location. Square root and natural logarithm transforms were applied to normalize the positively skewed response variables. Stepwise variable selection used the AIC statistic to guard against overfitting. Models predicting the transformed variables explained 80-93% of variance, based on 20-22 predictor variables. Lidar-derived variables had the most explanatory power; especially height and intensity variables for predicting plot basal area, and cover and intensity variables for predicting tree density. The ALI variables were less useful for predicting these attributes of forest structure, but could prove more helpful for predicting attributes of forest composition.