A Comparison of Object-based and Pixel-based Approaches to Estimate Lidar-measured Forest Canopy Height Using Quickbird Imagery

Canopy surface height (CSH) is a significant forest biophysical parameter to estimate above-ground biomass and carbon content. Highspatial resolution optical remotely sensed data have shown promising results to delineate various forest biophysical properties; though few studies have evaluated the accuracy of forest height information from such data. In this study, we compare several strategies using high-resolution Quickbird imagery to estimate CSH measured from small-footprint lidar data in a forest scene. Two main approaches were tested: 1) geographic object-based image analysis (GEOBIA), where the areal units are the objects from a segmentation-derived partition, which are akin to forest patches; and 2) pixel-based, where the areal units used to estimate CSH are the cells of a grid-shaped partition, which are akin to square field plots. Multiple linear regression models between within areal unit spectral response and lidar-measured CSH were developed for these two types of approaches using various areal unit sizes (AUSs). The best results (derived from the optimal AUSs) illustrated a better fitting model employing the GEOBIA approach (R = 0.605, RMSE = 2.86 m) than the pixel-based approach (R = 0.544, RMSE = 2.97 m). To develop more representative models when using their optimal AUSs, texture (i.e., standard deviation, skewness and kurtosis) and tree-ray-shadow geometry were investigated and applied to GEOBIA and pixel-based approaches. For the GEOBIA approach, the addition of texture and tree-ray-shadow geometry explained more variance of lidar-measured CSH by 5 percent and 10 percent respectively. The best performance (R = 0.739, RMSE = 2.60 m) was achieved using the combination of all three types of variables. For the pixel-based approach, only slight improvements were made with the best result (R = 0.577, RMSE = 2.88 m) achieved using all types of variables in the regression analysis. The comparisons in this study illustrate the potential of using meaningful image-objects instead of traditional fixed-size square grids to achieve higher accuracies in estimating the vertical structure of tree canopies. * Corresponding author.