Estimation of Forest Stand Characteristics Using Spectral Histograms Derived from an Ikonos Satellite Image

The aim of this paper was to examine the potential of Ikonos satellite images for estimating boreal forest stand characteristics using frequency distributions of radiometric values. The spectral features selected for use in the estimation were medians, standard deviations, and the parameters of the two-parametric Weibull distribution derived from the standwise spectral histograms of the Ikonos image. Ancillary map information, such as land-use and peatland classes, was also included. The method of estimation was nonparametric k-most similar neighbors (K-MSN) method. The most accurate results were achieved using spectral features that were derived from the multispectral images. The lowest RMSEs for the mean total stem volume, basal area, and mean height were 52.2 m3/ha (31.3 percent), 5.6 m2/ha (25.3 percent), and 3.1 m (20.6 percent), respectively. When only the panchromatic image was used in the analysis, the RMSEs for the mean total stem volume and basal area were about 3 percentage points higher. No differences in the mean height estimates were observed between the multispectral and panchromatic images. The most efficient predictor variables were the medians and the scale parameters of the Weibull distribution. The use of classified map information did not improve the results. The findings suggest that Ikonos satellite images can be used in to estimate forest stand characteristics giving an accuracy that corresponds to that achieved with aerial photographs. Introduction Valid information on forest resources and characteristics is essential for the planning forest use and management. Intensive field inventories for collecting stand-level information are expensive and time-consuming. However, recent progress made in remote sensing technology, e.g., the better availability of higher-resolution satellite images, offers an opportunity to acquire the necessary information at lower costs and in a reasonably short time. Numerical interpretation of remote sensing material at the stand level includes the extraction of stand-level features and their use for defining variables in models that predict the stand attributes of interest. There have been several studies concerning the estimation of forest stand characteristics from Estimation of Forest Stand Characteristics Using Spectral Histograms Derived from an Ikonos Satellite Image Jussi Peuhkurinen, Matti Maltamo, Lauri Vesa, and Petteri Packalén remote sensing material of different kinds in recent decades the most commonly used approaches being regression models (e.g., Hyyppä et al., 2000; Næsset, 2004; Kayitare et al., 2006) and non-parametric methods (e.g., Holmgren et al., 2000; Muinonen et al., 2001; Anttila, 2002). Non-parametric methods do not rely on the estimation of parameters (such as the mean or the standard deviation) describing the distribution of the variable of interest in the population. Using medium resolution (pixel size 10 to 30 meters) satellite imagery’s spectral information, the estimation accuracies of stand level forest characteristics achieved in boreal conditions have not been in sufficient level for forest planning purposes. The standard errors of volume estimates have varied from 36 percent (Holmgren et al., 2000) to 56 percent (Hyyppä et al., 2000). Basal area estimates’ accuracies have been from 24 percent (Holmgren et al., 2000) to 47 percent (Hyyppä et al., 2000) and those for the mean tree height 36 to 39 percent (Hyyppä et al., 2000). Holmgren et al. (2000) used the non-parametric k-nearest neighbor (K-NN) method whereas Hyyppä et al. (2000) utilized multiple regression and neural networks. Digitized aerial photographs have been widely studied in forest stand characteristics prediction (e.g., Muinonen et al., 2001; Anttila, 2002; Couteron et al., 2005). Muinonen et al. (2001) used the k-most similar neighbor (K-MSN) method to estimate mean forest stand volumes from a scanned aerial photograph. They achieved RMSE of 25 percent using only spectral information and 18 percent using additional textural information. Anttila (2002) used the same method with a larger dataset that contained several aerial photographs. He reported the RMSE of mean volume 37 percent as its best. Using only spectral information the RMSE was slightly worse (39 percent). Couteron et al. (2005) used two-dimensional Fourier analysis for obtaining characterization of canopy texture in tropical forests. The used remote sensing material was black and white aerial photographs with one-meter pixel size, which can be considered very similar to Ikonos panchromatic imagery. They reported significant correlations between the obtained canopy structure index and tree density, diameter of the tree of mean basal area, and mean canopy height. The use of Ikonos imagery in forest characteristics estimation has been studied using textural features PHOTOGRAMMETRIC ENGINEER ING & REMOTE SENS ING Novembe r 2008 1335 Jussi Peuhkurinen, Matti Maltamo, and Petteri Packalén are with the University of Joensuu, P.O. Box 111, FI-80101 Joensuu, Finland ([email protected]). Lauri Vesa is with Company LTV Forest, Koski-Jaakonkatu 15, FI-80230 Joensuu, Finland. Photogrammetric Engineering & Remote Sensing Vol. 74, No. 11, November 2008, pp. 1335–1341. 0099-1112/08/7411–1335/$3.00/0 © 2008 American Society for Photogrammetry and Remote Sensing 06-027.qxd 11/10/08 3:43 AM Page 1335