Regression modeling and mapping of coniferous forest basal area and tree density from discrete-return lidar and multispectral satellite data

We compared the utility of discrete-return light detection and ranging (lidar) data and multispectral satellite imagery, and their integration, for modeling and mapping basal area and tree density across two diverse coniferous forest landscapes in north-central Idaho. We applied multiple linear regression models subset from a suite of 26 predictor variables derived from discrete-return lidar data (2 m post spacing), advanced land imager (ALI) multispectral (30 m) and panchromatic (10 m) data, or geographic X, Y, and Z location. In general, the lidar-derived variables had greater utility than the ALI variables for predicting the response variables, especially basal area. The variables most useful for predicting basal area were lidar height variables, followed by lidar intensity; those most useful for predicting tree density were lidar canopy cover variables, again followed by lidar intensity. The best integrated models selected via a best-subsets procedure explained ~90% of variance in both response variables. Natural-logarithm-transformed response variables were modeled. Predictions were then transformed from the natural logarithm scale back to the natural scale, corrected for transformation bias, and mapped across the two study areas. This study demonstrates that fundamental forest structure attributes can be modeled to acceptable accuracy and mapped with currently available remote sensing technologies. Résumé. Nous avons comparé l’utilité du lidar à retour discret et de l’imagerie satellitaire multispectrale et leur intégration pour la modélisation et la cartographie de la surface terrière et la densité des arbres pour deux paysages diversifiés de forêts de conifères dans le centre-nord de l’Idaho. Nous avons appliqué les sous-ensembles des modèles de régression linéaire multiple d’une série de 26 variables prédictives dérivées de données lidar à retour discret (post-espacement de 2 m), de données multispectrales (30 m) et panchromatiques (10 m) du capteur ALI (« advanced land imager ») ou de localisation géographique en X, Y et Z. En général, les variables dérivées du lidar étaient d’une plus grande utilité que les variables ALI pour la prévision des variables dépendantes, particulièrement la surface terrière. Les variables les plus utiles pour la prévision de la surface terrière des arbres étaient les variables lidar de la hauteur des arbres suivies par l’intensité lidar ; les plus utiles pour la prévision de la densité des arbres étaient les variables lidar du couvert, là aussi suivies par l’intensité lidar. Les meilleurs modèles intégrés sélectionnés via une procédure du meilleur sous-ensemble a permis d’expliquer ~90% de la variance pour les deux vaiables dépendantes. Les variables dépendantes transformées par logarithme naturel ont été modélisées. Les prévisions ont alors été transformées de l’échelle ln, puis à l’échelle naturelle, corrigées pour le biais lié à la transformation et cartographiées sur l’ensemble des deux régions d’étude. Cette étude démontre que les attributs fondamentaux de la structure forestière peuvent être modélisés avec une précision acceptable et cartographiés au moyen de technologies de télédétection disponibles à l’heure actuelle. [Traduit par la Rédaction] Hudak et al. 138 Introduction Measures of stand structure are needed to manage forested landscapes for multiple purposes, including timber production, wildlife habitat, and fire hazard. Remote sensing of forest structure has proven challenging for forest operational managers and planners, many of whom still rely on aerial photograph surveys to meet user accuracy requirements. Although moderate-resolution satellite imagery (e.g., Landsat) is reasonably sensitive to variation between managed forest stands, it is insensitive to canopy height variation within stands compared to aerial photography. Laser altimetry and light detection and ranging (lidar) systems, on the other hand, actively measure height to the reflective surface. Most commercially available discrete-return lidar systems can accurately measure top-of-canopy height and ground height, as well as canopy layers in between. Recognizing that passive imaging and active lidar systems sense fundamentally different 126 © 2006 CASI Can. J. Remote Sensing, Vol. 32, No. 2, pp. 126–138, 2006 Received 30 September 2005. Accepted 26 January 2006. A.T. Hudak,1 N.L. Crookston, and J.S. Evans. Rocky Mountain Research Station, US Department of Agriculture Forest Service, 1221 South Main Street, Moscow, ID 83843, USA. M.J. Falkowski, A.M.S. Smith, P.E. Gessler, and P. Morgan. Department of Forest Resources, University of Idaho, Sixth & Line Streets, Moscow, ID 83844-1133, USA. 1Corresponding author (e-mail: [email protected]). aspects of forest structure, and that probably no single remote sensor can provide all of the information useful and relevant to forest managers, the integration of image and lidar data for the purpose of predicting, mapping, managing, and monitoring forest structure attributes is a logical and worthwhile pursuit (Lefsky et al., 1999; Hudak et al., 2002). Landsat imagery has become the standard relied upon by many forest ecologists and managers who use remotely sensed data (Cohen and Goward, 2004). Landsat data coverage began with the launch of Landsat-1 in 1972. Landsat-5 operated far beyond its expected lifespan, from 1984 until 26 November 2005, when the appearance of a solar array drive anomaly briefly halted imaging (http://landsat.usgs.gov/technical_details/ investigations/l5_solar_drive.php). Landsat-7 was launched and has operated since 1999, although with reduced utility since a scan line corrector anomaly began on 31 May 2003 (http://landsat. usgs.gov/programnews.html). Considering the declining availability of new Landsat imagery, there is justifiable concern for maintaining Landsat data continuity, particularly until the launch of the Landsat data continuity mission (LDCM) operational land imager (OLI), which will provide Landsat-like imagery but is expected no sooner than late 2009 (http://ldcm. usgs.gov/). The advanced land imager (ALI) satellite sensor was designed in part to provide data continuity with the Landsat-5 thematic mapper (TM) and Landsat-7 enhanced thematic mapper plus (ETM+) sensors (http://eo1.usgs.gov/ali.php). Although the ALI swath width (37 km) is more restricted than that of Landsat (185 km), and ALI acquisitions must be scheduled in advance, the ALI sensor is pointable. The ALI measures solar irradiance in nine multispectral bands between 0.433 and 2.350μm in the electromagnetic spectrum, matching the six multispectral bands of Landsat TM or ETM+, plus an additional three bands. The spatial resolution of the panchromatic (PAN) band is 10 m, an improvement over the 15 m resolution of the ETM+ panchromatic band. Furthermore, ALI data are 16-bit rather than 8-bit, offering greater dynamic range. In a comparative study, Bryant et al. (2003) found no disadvantages of the ALI sensor relative to the TM or ETM+ sensors and recommended the ALI sensor for a potential Landsat-8 payload. Efforts to model and map height and related attributes from satellite imagery alone have generally been too inaccurate for forest operational managers. Canopy height is particularly valued by foresters because it relates strongly to other structure attributes of interest, such as basal area and biomass. Numerous studies have demonstrated the utility of lidar for characterizing various attributes of forest canopy structure from discretereturn lidar data (Nelson, 1984; Nilsson, 1996; Means et al., 2000). Enthusiasm for lidar-based forest inventory is driving expansion of the commercial lidar industry (Flood, 2001). As the costs of managing forested landscapes increase in a competitive environment, commercial timber and paper companies are increasingly turning to lidar for potentially more accurate and efficient inventory and assessment of their forest resources. Our objective was to compare the relative utility of discretereturn lidar data and ALI satellite imagery, and their integration, for modeling and mapping basal area and tree density across two spatially disjunct coniferous forest landscapes situated along a single biomass and productivity gradient in northern Idaho. Many researchers have recognized the potential of remote sensing data integration, making “data integration” a broad term that needs to be more narrowly defined. Lefsky et al. (2001) compared the utility of several remote sensing data types for accurately characterizing high-biomass forest structure in western Oregon and found that lidar outperformed digital aerial photography, hyperspectral aerial imagery, and multispectral satellite imagery. Rather than evaluate many remote sensing products, we used single acquisitions of discrete-return lidar data and multispectral satellite imagery, much like a commercial forester with limited time and resources might do. Popescu and Wynne (2004) fused lidar and multispectral image data to improve estimates of individual tree height in eastern forests. Rather than examine individual tree attributes, we focused on stand attributes of interest to planners and managers of large forested landscapes. Lastly, rather than “fuse” remotely sensed data layers, or test a variety of data integration methods, we focused on the simple and widely applicable method of multiple linear regression. Hence the data integration conducted in this analysis is purely statistical but provides an accessible means of selecting remotely sensed predictor variables and evaluating alternative models. This study is intended to demonstrate to forest planners and operational managers that it is within their means to model and map fundamental stand structure variables of interest to acceptable accuracy with current lidar and imaging technologies.