A cross-comparison of field, spectral, and lidar estimates of forest canopy cover

A common challenge when comparing forest canopy cover and similar metrics across different ecosystems is that there are many fieldand landscape-level measurement methods. This research conducts a cross-comparison and evaluation of forest canopy cover metrics produced using unmixing of reflective spectral satellite data, light detection and ranging (lidar) data, and data collected in the field with spherical densiometers. The coincident data were collected across a ~25 000 ha mixed conifer forest in northern Idaho. The primary objective is to evaluate whether the spectral and lidar canopy cover metrics are each statistically equivalent to the field-based metrics. The secondary objective is to evaluate whether the lidar data can elucidate the sources of error observed in the spectral-based canopy cover metrics. The statistical equivalence tests indicate that spectral and field data are not equivalent (slope region of equivalence = 43%). In contrast, the lidar and field data are within the acceptable error margin of most forest inventory assessments (slope region of equivalence = 13%). The results also show that in plots where the mean lidar plot heights are near zero, each of modeled remotely sensed estimates continues to report canopy cover >21% for lidar and >30% for all investigated spectral methods using near-infrared bands. This suggests these metrics are sensitive to the presence of herbaceous vegetation, shrubs, seedlings, saplings, and other subcanopy vegetation. Résumé. Un défi rencontré fréquemment en comparant les mesures du couvert forestier ou autres mesures semblables à travers différents écosystèmes vient du fait qu’il existe plusieurs méthodes de mesure sur le terrain ainsi qu’au niveau du paysage. Dans cette recherche, on compare et on évalue les mesures du couvert forestier réalisées à l’aide de trois approches différentes, c.-à-d. le démixage des données spectrales satellitaires, les données lidar (« light detection and ranging ») et l’acquisition de données sur le terrain avec des densitomètres sphériques. Des données simultanées ont été acquises sur l’ensemble d’une forêt mixte de conifères de ~25 000 ha dans le nord de l’Idaho. L’objectif premier est d’évaluer si les mesures du couvert à l’aide des données spectrales et lidar sont statistiquement équivalentes par rapport aux mesures sur le terrain. Le second objectif est d’évaluer si les données lidar peuvent élucider les sources d’erreur observées dans les mesures du couvert dérivées des mesures spectrales. Les tests d’équivalence statistique indiquent que les données spectrales et de terrain ne sont pas équivalentes (région d’équivalence de la pente = 43 %). Par contre, les données lidar et de terrain se situent à l’intérieur de la marge d’erreur acceptable de la plupart des évaluations d’inventaires forestiers (région d’équivalence de la pente = 13 %). Les résultats montrent également que, dans les parcelles où les hauteurs lidar moyennes sont près de zéro, chacune des estimations modélisées par télédétection continue de donner un couvert de >21 % pour le lidar et de >30 % pour toutes les méthodes spectrales analysées en utilisant les bandes du moyen infrarouge. Ceci laisse supposer que ces mesures sont sensibles à la présence de végétation herbacée, d’arbustes, de semis, de gaules ou d’autre végétation présente sous le couvert. [Traduit par la Rédaction] Smith et al 459 Introduction Forest canopy cover (CCForest), which is commonly defined as a projection of the vertical profile of canopy foliage onto a horizontal plane (Fiala et al., 2006), is a useful metric for several biophysical and natural resource management applications (Hopkinson and Chasmer, 2009). These application areas include the assessment of wildlife habitat (Koy et al., 2005; Fiala et al., 2006), parameterization of fire behavior simulation models (Finney, 1998), characterization of carbon pools and sources (Chopping et al., 2008), quantification of canopy light transmission (Lieffers et al., © 2009 CASI 447 Can. J. Remote Sensing, Vol. 35, No. 5, pp. 447–459, 2009 Received 9 May 2009. Accepted 10 November 2009. Published on the Web at http://pubservices.nrc-cnrc.ca/cjrs on 12 April 2010. A.M.S. Smith.1 Forest and Rangeland Measurements Laboratory, University of Idaho, Moscow, ID 83844, USA. M.J. Falkowski. School of Forest Resources and Environmental Science, Houghton, MI 49931, USA. A.T. Hudak. Rocky Mountain Research Station, US Forestry Service, Moscow, ID 83844, USA. J.S. Evans. The Nature Conservancy, Fort Collins, CO 80524-2863, USA. A.P. Robinson. Department of Mathematics and Statistics, University of Melbourne, Melbourne Australia. C.M. Steele. Jornada Experimental Range, USDA Agriculture Research Service, New Mexico State University, Las Cruces, NM 88003, USA. 1Corresponding author (e-mail: [email protected]). 1999), and ecosystem structure classification (Lovell et al., 2003; Fiala et al., 2006; Lee and Lucas 2007), among others (Fiala et al., 2006; Chopping et al., 2008). Furthermore, the United Nations Food and Agriculture Organization (FAO) definition of forest includes a canopy cover parameter, and therefore improvement of such estimates is of global significance, especially in areas with a high canopy cover (FAO, 2000). One of the challenges with comparing canopy cover estimates across studies or ecosystems is the number of fieldand landscape-level measurement methods that exist. As outlined in recent studies, field-based measurement of CCForest can be obtained using a wide range of equipment, including hemispherical photography, spherical densiometers, the moosehorn densitometer, point counts, stem maps, and line intercept methods (Fiala et al., 2006; Korhonen et al., 2006). An important distinction between these field methods is that hemispherical photographs and spherical densiometers integrate information from the sky hemisphere over a single point on the ground, which could be considered as a measure of canopy closure, whereas the other field methods measure canopy presence–absence above a spatially distributed twodimensional (2D) sample of points on the ground, which is a measure of canopy cover (Jennings et al., 1999). Furthermore, light detection and ranging (lidar) data add a third dimension to the problem by including a distribution of points within a threedimensional (3D) volume of space above the ground, perhaps making the term canopy density more appropriate. Clearly these different terms are confusing, especially given their slightly different interpretations, requiring a need for ease and consistency. Therefore, throughout this paper we use the term CCForest to describe each metric. Landscape-level assessments of CCForest often rely on satellite and aircraft sensor imagery or, more recently, laser altimetry and light detection and ranging (lidar) data. From the perspective of these landscape-level remote sensing approaches, two types of canopy cover estimates are commonly derived: metrics describing the 2D horizontal extent of canopy, which is often expressed for a given cover type as a percentage of pixels (Asner et al., 2003; Falkowski et al., 2005), subpixel proportions (Pu et al., 2003; Xu et al., 2003; Sommers et al., 2009), and discrete image objects (Greenberg et al., 2005; Strand et al., 2006; 2008; Smith et al., 2008); or as 3D lidar metrics that represent the transmission of light through the canopy (Means et al., 1999; Chen et al., 2004; Hyde et al., 2005; Lefsky et al., 2005; Hopkinson and Chasmer, 2009). A range of metrics have been proposed for lidar data to represent CCForest in forested ecosystems, including simple binary classifications of rasterized lidar data (i.e., pixel contains a canopy return or pixel does not contain a canopy return) (Chen et al., 2004) and other lidar metrics that relate to the proportion of returns penetrating the canopy (Hopkinson and Chasmer, 2009). Previous lidar studies have also evaluated the relationships between the mean and maximum lidar heights at a given plot with field-derived CCForest (Thomas et al., 2006; Hopkinson and Chasmer, 2009) to investigate the potential of modeling CCForest at landscape scales. However, Hopkinson and Chasmer (2009) observed poor relationships when using lidar-derived maximum plot heights to predict CCForest. This may be due to skewed perspectives of the overall structure of a plot, especially if remnant, open-grown–isolated trees or single dominant trees are present; stronger relationships have been noted when using mean lidar plot heights (Thomas et al., 2006). Although many studies have attempted to characterize CCForest via spectral remote sensing, few have compared the remotely derived CCForest estimates with coincident field measurements (e.g., Falkowski et al., 2005). Instead, datasets with higher spatial resolution (1–4 m pixel size) have been employed to validate estimates of CCForest derived from coarser resolution spectral data (Pu et al., 2003; Xu et al., 2003). In contrast, studies deriving CCForest from lidar data often compare lidar estimates with coincident field measurements (Magnussen and Boudewyn, 1998; Riano et al., 2004a; 2004b; Morsdorf et al., 2006; Hopkinson and Chasmer, 2009). This is largely due to the desire to augment traditional forest inventories with lidar data, which requires the collection of coincident field inventory data. Lidar data have proven useful for estimating CCForest; however, because of logistical and financial constraints associated with acquiring lidar data across large areas, lidarbased characterizations of canopy metrics are often limited in spatial extent. Spectral remote sensing datasets (e.g., the Landsat series) have also been employed to estimate CCForest. However, the insensitivity of these spectral datasets to the 3D structure of vegetation canopies (Falkowski et al., 2005) often degrades the relationship between CCForest and metrics calculated from the spectral data (e.g., band ratios or vegetation indices). A further challenge when using spectral indices is that they will always produce poorer relationships than those produced simply using a multiple regression of the individual bands (Lawrence and Ripple, 1998). Nevertheless, the affordability and large area availability of spectral datasets still make them an attractive data source for characterizing CCForest across large spatial extents. Prior to using any spectral dataset to estimate CCForest across large spatial extents, the relationship between the remotely sensed data and three-dimensional forest structure must be quantified and understood. In addition to conducting a spectral, lidar, and field crosscomparison of CCForest, the research presented in this paper also aims to quantify the relationship between spectral remotely sensed data and the 3D structure of forest canopies. This is primarily achieved by comparing estimates of CCForest derived from imagery obtained using the nadir bands of the Advanced Spaceborne Thermal Emission and Reflection Radiometer (ASTER) with lidar metrics describing the 3D structure of forest canopies. Understanding the magnitude and source of errors in CCForest metrics produced from ASTER imagery will enable an evaluation of the potential uncertainties within landscapeto global-level remote sensing products (e.g., LANDFIRE and FAO products) that use or produce similar metrics. This study seeks to answer the following specific questions: (1) Are CCForest estimates derived from both spectral 448 © 2009 CASI Vol. 35, No. 5, October/octobre 2009