Forest structure modeling with combined airborne hyperspectral and LiDAR data Hooman Latifi ⁎, Fabian Fassnacht, Barbara Koch Dept. of Remote Sensing and Landscape Information Systems, University of Freiburg, Tennenbacherstraße. 4, D-79106 Freiburg, Germany abstract article info Article history: Received 10 June 2011 Received in revised form 16 January 2012 Accepted 17 January 2012 Available online 17 February 2012 Keywords: LiDAR Hyperspectral Forest structure GA Spatial models The interest in the joint use of remote sensing data from multiple sensors has been remarkably increased for environmental applications. This is because a combined use is supposed to improve the results of e.g. forest modeling tasks compared to single-data use. To explore the ability of combined airborne 2D and 3D informa- tion to describe the forest structure in local level, we employed various height/intensity metrics from Light Detection and Ranging (LiDAR) data and original reflectance, indices, and linear transformations of airborne hyperspectral HyMap data to build spatial models of stem density, above ground total biomass, and biomass of coniferous species in a temperate forest site in Germany. The study area was stratified into coniferous, deciduous and mixed strata using the plot information from forest inventory data. Combinations of data sources were tested, and an evolutionary Genetic Algorithm (GA) was used to tailor the numerous predictor variables to final parsimonious sets. Most Similar Neighbor (MSN) approach based on variance-weighted ca- nonical correlations were used to make simultaneous single-Nearest Neighbor (NN) models of the attributes, where NN was searched either within the whole geographical domain or within the restricted forest strata. Results were evaluated by leave-one-out cross validations on 1000 bootstrap resample data. They showed that the LiDAR height metrics (descriptive statistics and percentiles) provided the most effective information amongst the entire data source combinations, while the HyMap metrics contributed only slightly to describe the variation beyond those explained by ALS data. Furthermore, restricted NN search improved the perfor- mance and returned approximately unbiased models of all the responses. The GA-screened HyMap predictors corresponded well to the atmospheric windows in visual and NIR domains, as well as to the mean reflectance curve of Scots Pine across the study area. It is concluded that GA-screened models featuring 9–12 predictors containing LiDAR height metrics and few HyMap original channels can be suggested for timely-efficient, unbiased modeling of area-based forest structural attributes. © 2012 Elsevier Inc. All rights reserved. 1. Introduction Remote sensing has been previously proven to be a functional tool to support ground based inventory systems. Different tools have shown potentials for the derivation of forest attributes, amongst which Light Detection and Ranging (LiDAR) has been reported as a promising data source for description of complex forest structure thanks to its ability to capture 3D features such as canopy height and density. While Hyyppä et al. (2008) summarized the studies employing LiDAR for for- est inventory in the Nordic countries, Koch (2010) reviewed the current status of optical and LiDAR remote sensing application for forest bio- mass assessment, in which the magnitude of joint application of those data was highlighted. The benefit of a combined use of LiDAR and mul- tispectral data for single tree or area-based models of forest attributes has been reported by a number of state-of-the-art studies in the litera- ture e.g. Breidenbach et al. (2010a), Latifi et al. (2010), Latifi et al. (2012) and Straub et al. (2010). Using different modeling schemes from the family of so called Nearest Neighbor (NN) spatial models is one common issue amongst the above-mentioned studies. In NN models a weighted mean of the response variables from the most sim- ilar neighbors (when the number of neighbors is greater than one) is used to assign the value of response to a target unit being predicted (Haapanen et al., 2004). The Most Similar Neighbor (MSN) is a special case of NN method in which the distances between the target and (neighboring) reference units are weighted using the canonical correla- tions of the predictor variables. Here an adaptation of MSN will be used, in which the canonical correlations are weighted by their variances (Crookston et al., 2002). Besides LiDAR technology, imaging spectroscopy (also called hyperspectral remote sensing) is a recent field which gained remark- ably increased attention in the last years with the introduction of airborne sensors such as AVIRIS (Green et al., 1998) and HyMap (Cocks et al., 1998). It offered sensors with a significantly increased spectral resolution compared to the multispectral sensors. The narrow spectral width of each individual channel in a hyperspectral sensor enables displaying of the complete spectral characteristics of each object. Moreover, the absence/presence of specific absorption Remote Sensing of Environment 121 (2012) 10–25 ⁎ Corresponding author. Tel.: + 49 7612033699; fax: + 49 7612033701. E-mail addresses: hooman.latifi@felis.uni-freiburg.de (H. Latifi), fabian.fassnacht@felis.uni-freiburg.de (F. Fassnacht), barbara.koch@felis.uni-freiburg.de (B. Koch). 0034-4257/$ – see front matter © 2012 Elsevier Inc. All rights reserved. doi:10.1016/j.rse.2012.01.015 Contents lists available at SciVerse ScienceDirect Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse