A Support Vector Regression Approach to Estimate Forest Biophysical Parameters at the Object Level Using Airborne Lidar Transects and QuickBird Data

A potential solution to reduce high acquisition costs for airborne lidar (light detection and ranging) data is to combine lidar transects and optical satellite imagery to characterize forest vertical structure. Although multiple regression is typically used for such modeling, it seldom fully captures the complex relationships between forest variables. In an effort to improve these relationships, this study investigated the potential of Support Vector Regression (SVR), a machine learning technique, to generalize (lidar-measured) forest canopy height from four lidar transects (representing 8.8 percent, 17.6 percent, 26.4 percent and 35.2 percent area of the site) to the entire study area using QuickBird imagery. The best estimated canopy height was then linked with field measurements to predict actual canopy height, above-ground biomass (AGB) and volume. GEOgraphic Object-Based Image Analysis (GEOBIA) was used to generate all estimates at a small tree/cluster level with a mean object size (MOS) of 0.04 ha for conifer and deciduous trees. Results show that for all lidar transect samples, SVR models achieved better performance for estimating canopy height than multiple regression. By using SVR and a single lidar transect (i.e., 8.8 percent of the study area), the following relationships were found between predicted and field-measured canopy height (R2: 0.81; RMSE: 4.0 m), AGB (R2: 0.76; RMSE: 63.1 Mg/ha) and volume (R2: 0.64; RMSE: 156.9 m3/ha). Introduction Forests play a critical role in the global carbon budget, as they dominate the dynamics of the terrestrial carbon cycle (Dong et al., 2003); where, for example, 90 percent of above-ground carbon is stored in tree stems (Hese et al., 2005). Airborne lidar (light detection and ranging), a recent remote sensing tool, has demonstrated the ability to characterize forest vertical structure (e.g., canopy height), leading to the accurate estimation of forest above-ground biomass (AGB) and timber volume (Means et al., 1999; Lefsky et al., 2002; Lim et al., 2003). However, the current PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING J u l y 2011 733 A Support Vector Regression Approach to Estimate Forest Biophysical Parameters at the Object Level Using Airborne Lidar Transects and QuickBird Data Gang Chen and Geoffrey J. Hay cost of lidar data collection still remains high. This prohibits the wall-to-wall airborne lidar forest mapping of large areas, such as Canada, which is covered by 402.1 million hectares of forest and other wooded land (Natural Resources Canada, 2009). In an effort to overcome this cost limitation, recent studies report on the integration of lidar transects with optical remotely sensed data to estimate forest vertical structure (Hudak et al., 2002; Wulder and Seemann; 2003; Hilker et al., 2008; Chen and Hay, 2011). The primary strategy is to generalize canopy height information from a relatively small area covered by lidar transects to the entire study area, covered by the optical scene. Multiple regression, as a standard statistical technique, is widely used in these studies. However, linear or other simple nonlinear (e.g., logarithmic or exponential) multiple regression models seldom fully characterize forest complexity, especially at fine scales, due to the high structural variability within small tree clusters when using high spatial resolution imagery (i.e., less than 5.0 m). Support vector machines (SVMs), originating from statistical learning theory, provide the capability to deal with highly nonlinear problems (Vapnik, 1995 and 1998) such as estimating complex forest structures. Additionally, SVMs are (a) robust in generalization, even when the training data are noisy, and (b) are guaranteed to have a unique global solution, that is not trapped in multiple local minima (Cristianini and Shawe-Taylor, 2000). In forest remote sensing studies, SVMs have proven their use in the domain of classification (Huang et al., 2008; Kuemmerle et al., 2009). However, few studies have investigated the application of support vector regression (also known as SVMs for regression, hereafter SVR) to estimate forest biophysical parameters, especially their vertical characteristics. To reduce high costs for lidar acquisition for forest parameterization while improving model accuracies, the primary objective of this study is to investigate the potential of SVR machine learning models to estimate forest biophysical parameters (i.e., canopy height, AGB and volume) for a full study site by combining (smaller-area) Photogrammetric Engineering & Remote Sensing Vol. 77, No. 7, July 2011, pp. 733–741. 0099-1112/11/7707–0733/$3.00/0 © 2011 American Society for Photogrammetry and Remote Sensing Foothills Facility for Remote Sensing and GIScience, Department of Geography, University of Calgary, 2500 University Dr. NW, Calgary, AB TN 1N4, Canada ([email protected]). lidar transects and QuickBird imagery. To do so, we proceed by: (a) applying a GEOBIA (geographic object-based image analysis) approach to extract forest characteristics at a small tree cluster level, (b) developing SVR models to estimate forest biophysical parameters, and (c) comparing the model performance between SVR and multiple regression. Support Vector Regression (SVR) Support vector regression (SVR) essentially transforms the nonlinear regression problem into a linear one by using kernel functions to map the original input space into a new feature space with higher dimensions (Cristianini and Shawe-Taylor, 2000). A brief description of the SVR basic principles is addressed below. Please refer to Gunn (1998), Cristianini and Shawe-Taylor (2000), and Smola and Schölkopf (2004) for details. In SVR, if we consider training samples as (xi, yi), (i 1, . . ., n), where xi is a multivariate input, yi is a scalar output, and n is the number of training samples; then a linear model can fit this new high-dimensional feature space as follows: