International Journal of Applied Earth Observation and Geoinformation
Integrating optical satellite data and airborne laser scanning in habitat classification for wildlife management
Graphical abstract
Introduction
Wildlife respond to a large number of factors when selecting habitat, involving complex behavioral decisions which are made at multiple spatial scales (Ciarniello et al., 2007, Herfindal et al., 2009, Johnson et al., 2002). Broad scale spatial variation in biodiversity is thought to respond to three major drivers; climatic stability, productivity, and habitat structure (MacArthur, 1972) – with empirical evidence demonstrating the importance of each of these variables (Coops et al., 2008). Bioclimatic models are often applied to estimate broad-scale distribution of species (Guisan and Zimmermann, 2000, Rahbek and Graves, 2001, Willis and Whittaker, 2002). However, at finer spatial scales land cover, disturbance, and habitat heterogeneity are more important factors affecting local distribution and habitat selection of species (Iverson and Prasad, 1998, Thuiller, 2004).
The vertical and horizontal structure of vegetation plays a critical role in defining suitable wildlife habitat and can do so in a variety of ways. For certain species, vegetation structure drives food quality, diversity, and availability (Hamer and Herrero, 1987, Johnson et al., 2002, Månsson et al., 2007). Access to high quality forage in early successional stage forest stands, deciduous overstorey stands, or open areas with grass, forb, herb and berry species (Allen et al., 1987, Dussault et al., 2005, Munro et al., 2006) decrease energy required for foraging and digestion in Grizzly bear (Ursus arctos), and thus, maximise energy intake (White, 1983). Vegetation structure also provides protection and/or cover which provides security against predation and can protect species from heat stress when ambient temperature exceeds optimal levels (Schwab and Pitt, 1991), or deep snow during winter; with snow accumulation often adversely impacting species mobility and food intake, and thus, the survival and reproductive rates (Cederlund et al., 1991, Mech and McRoberts, 1987, Post and Stenseth, 1998). Vegetation structure is also inextricably linked to disturbances; especially fire, harvesting, and insect defoliation. As a result, disturbances potentially increase future habitat suitability for bears (Nielsen et al., 2008, Nielsen et al., 2004b, Rempel et al., 1997, Stewart et al., 2012). Heterogeneity in vegetation structure also provides access to forest edges, where forage and protection are amplified (i.e., the cover-food edge concept) which is a key habitat type selected by many species (Courtois et al., 2002, Dussault et al., 2005, Stewart et al., 2013), although edges can also represent attractive sinks where survival is low (Nielsen et al., 2006, Nielsen et al., 2004a).
Grizzly bears have diverse seasonal habitat requirements with three distinct foraging seasons, hypophagia, early hyperphagia, and later hyperphagia (Nielsen et al., 2006). In hypophagia they forage on roots (such as alpine sweetvetch), early herbaceous material and ungulate kills, in early hyperphagia their main diet is green herbaceous material (such as cow parsnip, sedges, and horsetails) with some insect matter, whereas in late hyperphagia berries make up a the majority of their diet (Hamer and Herrero, 1987, Munro et al., 2006). The optimal habitat for Grizzly bears, therefore, changes significantly throughout the season and contains herbaceous areas, wetlands, and open forest, as well as proximity to forest stands for other habitat requirements including bedding and hiding cover. Over the past 40 years, since the launch of the first Earth observation satellites, satellite-based image classification techniques have been used to map species habitat and has become an important tool in large area mapping and management of wildlife habitat (McDermid et al., 2005, Wang et al., 2009). The Landsat series of sensors in particular have set the standard for regional classification projects because of their combination of spatial and spectral resolution, consistent long term record, and excellent data availability (Cohen and Goward, 2004, Franklin and Wulder, 2002, Leimgruber et al., 2005). However, considerable limitations exist in the application of optical satellite imagery specifically involving the detection of detailed forest structural characteristics beyond initial canopy closure (Franklin et al., 2003, Wang et al., 2009). The issue of signal saturation on optical remote sensing imagery with increasing leaf area is well known. Studies have shown both theoretically and practically that estimation of canopy parameters can be difficult beyond a leaf area index of 3–5 (Baret and Guyot, 1991, Song, 2012, Turner et al., 1999) and that canopy parameter estimation also varies between conifer and deciduous canopy types (Song, 2012). As a result while classification schemes often recognize the importance of forest structure in the class definition (Franklin and Wulder, 2002, McDermid et al., 2009, Wulder et al., 2008b), they are often generalized or have considerable uncertainty in forest density classes caused by the inherent limitations of the optical sensor system.
Many have tried to bridge the gap between the need for structural information and the inability of direct optical classification to provide this information. Solutions may include the use of ancillary data, texture information, object based analysis, post classification procedures, or other remotely sensed data like radar (Lu and Weng, 2007, Roberts et al., 2007). The most common source of ancillary data is elevation models (Franklin et al., 2002, Johnson et al., 2003, McDermid et al., 2009) and topographic derivatives like slope and aspect. Texture information is used in the form of gray-level co-occurrence matrices (Franklin et al., 2002), spatial autocorrelation (Magnussen et al., 2004), or variogram functions (Zhang et al., 2004), based on homogeneity assumptions within the forest stand and the information content of shaded vs. sunlit parts in the canopy. In post classification methods the fine scale patterning of simple land-cover types (e.g., treed, herb, bare) or vegetation indices can be used to define habitat classes (Sluiter et al., 2004). Radar in particular is able to partially penetrate vegetation canopies, but the efficacy in detecting structure is highly dependent on the microwave wavelength, vegetation height and moisture content (Imhoff et al., 1997). All of these potential solutions can improve classification results in certain cases, but can be laborious, costly and require extensive training data or manual steps which may lead to interpreter-related differences and locally optimised but regionally less applicable results.
Airborne laser scanning (ALS) uses discrete return small footprint airborne lidar to map the elevation of the ground surface and canopy elements. ALS provides high accuracy measurement of canopy heights and density through the separation of the terrain model from canopy returns. Terrain height and landforms are used to model hydrological and soil processes (White et al., 2012) and are shown to be key drivers of plant species distribution (Nijland et al., 2014). The potential of ALS to detect structural forest characteristics has been shown in many studies, and it has quickly become an operational technology for estimation of forest height, cover and structure around the world (Lim et al., 2008, Wulder et al., 2008a). ALS data can provide specific information on forest structure, such as understory and midstory cover assessment, topographic morphological variables, such as slope and aspect, as well as the presence of old, tall trees or snags. As a result, the use of ALS technology has increased for assessments of wildlife habitat. Hyde et al. (2005) utilized ALS data to characterize montane forest canopy structure in the Sierra National Forest for large-area habitat mapping. They found that the accurate prediction of canopy height, canopy cover, and biomass was an important prerequisite predicting wildlife habitat showing significant promise in its use. Vierling et al. (2008) provide a review of the current status of ALS remote sensing and habitat characterization and conclude that, although a growing number of studies highlight interest in ALS advances, few studies have actually used the data to quantitatively address these relationships.
Western Alberta, Canada is a highly dynamic region where widespread resources extraction from the forestry and fossil fuels industries occurs on important habitat for species at risk (Roever et al., 2008). Coal, oil, gas, and timber extraction, in addition to related population growth, urban development and expanding demands for outdoor recreation impact biodiversity through habitat alteration and fragmentation (Schneider et al., 2003). Western Alberta represents the eastern limit of Grizzly bear habitat in Southern Canada (Nielsen et al., 2009) and has an important population of woodland caribou (Tarandus rangifer) (Bradshaw and Hebert, 1996, Festa-Bianchet et al., 2011). Effective management of wildlife habitat is of paramount importance for sustainable support of both ecological values and resource extraction in the region. To support wildlife and habitat management, we need a detailed account of habitat status and a thorough understanding of the habitat requirements of different key species. The availability of accurate habitat maps is crucial for both objectives.
In this research, we introduce an approach to integrate ALS and multispectral satellite images to develop a new heuristic wildlife habitat classifier for western Alberta. The classifier uses vegetation structure, species composition, and terrain characteristics derived from available ALS and multispectral data directly in a decision tree. We evaluate the accuracy of the habitat layers and discuss the added value of the created products for the classification. Based on our results, we look at implications of increased ALS availability for habitat mapping and wildlife management, and make recommendations on the application of ALS in regional habitat mapping efforts.
Section snippets
Study area
Our focus areas encompasses the western Rocky Mountains in Alberta, Canada constrained by the Upper and Lower foothills Natural subregions, with the higher elevations in the Alpine natural subregion (Downing and Pettapiece, 2006) (Fig. 1). Elevations range from 700 to 3000 m ASL with steep montane topography in the west, transitioning to a gently rolling landscape in the eastern parts of the area. The natural vegetation in the sub-alpine areas is forested with Lodgepole Pine (Pinus contorta),
Pre classification models
ALS derived maximum canopy height and points above 2 m were selected to represent canopy height and cover directly without using any models. The relations between the selected variables and field based measurements of canopy structure were strong (R2 vegetation height: 0.87, vegetation cover over 2 m: 0.60) and visual evaluation of the relations reveal no bias in the estimators (Fig. 3). The accuracy of using direct ALS variables (RMSE canopy height: 3.08 m, canopy cover: 16.2%) is acceptable for
Discussion
Traditional spectral classifiers rely on training data (for supervised classifications), or the discretion of the interpreter (for labeling unsupervised clustering), and maximise class separability for classes that are spectrally different. By adding ancillary data sources such as terrain information, by stratifying datasets, or including the spatial domain into the classification it is possible to improve the classification accuracy.
Availability of ALS data into habitat classifications allows
Conclusions
In this paper, we present a new habitat classification for Grizzly bear management in Alberta, Canada. We combine optical satellite images and ALS into a heuristic, decision tree based habitat classifier. Based on the integrated use of optical and ALS data we are able to describe the major axes of landscape variability including species composition and vegetation structure and to use these data directly in the landcover classifier. The classifier allows for more detailed habitat classes in
Acknowledgements
Funding for this research was generously provided by the Grizzly bear program of the Foothills Research Institute located in Hinton, Alberta, Canada, with additional information available at: www.foothillsri.ca. Additional funds were provided by an NSERC Discovery grants to Dr. Nicholas Coops and Dr. Scott Nielsen. Lidar data was made available by the Government of Alberta. The authors thank Adam Erickson and Ilia Parshakov for their help with the field program, and Dr. Txomin Hermosilla for
References (77)
- et al.
Potentials and limits of vegetation indices for LAI and APAR assessment
Remote Sens. Environ.
(1991) - et al.
The development of a Canadian dynamic habitat index using multi-temporal satellite estimates of canopy light absorbance
Ecol. Indic.
(2008) - et al.
Predictive habitat distribution models in ecology
Ecol. Modell.
(2000) - et al.
Remotely sensed indicators of habitat heterogeneity: use of synthetic aperature radar in mapping vegetation strucutre and bird habitat
Remote Sens. Environ.
(1997) - et al.
Characterizing woodland caribou habitat in sub-boreal and boreal forests
For. Ecol. Manage.
(2003) - et al.
Remote sensing and forest inventory for wildlife habitat assessment
For. Ecol. Manage.
(2009) - et al.
Grizzly bears and forestry I: selection of clearcuts by Grizzly bears in west-central Alberta
For. Ecol. Manage.
(2004) - et al.
A habitat-based framework for Grizzly bear conservation in Alberta
Biol. Conserv.
(2006) - et al.
Grizzly bears and forestry: II. distribution of Grizzlybear foods in clearcuts of west-central Alberta, Canada
For. Ecol. Manage.
(2004) - et al.
Can natural disturbance-based forestry rescue a declining population of Grizzly bears?
Biol. Conserv.
(2008)