Abstract

The structure and function of oak Quercus spp. savanna ecosystems in the North American Midwest were originally maintained by an active disturbance regime (often fire). Subsequent reductions in the frequency of disturbance after European settlement have facilitated rapid, regional conversion of these ecosystems to more closed-canopy forest. Hence, regional-scale management strategies are now needed to restore critical spatial gradients of light, temperature, soil moisture, and soil organic matter for recovery and sustenance of the unique mosaic of understory grass and forb species assemblages that define oak savannas. Tree species composition, distribution, mortality, basal area, and canopy cover are important forest structural parameters that are intrinsically linked to oak savanna restoration ecology. In this benchmark study, we seek to determine whether Landsat-based monitoring protocols can be developed as a tool to guide and monitor regional-scale restoration and management efforts. Using the Sherburne National Wildlife Refuge in central Minnesota as a test case, ground-based forest-structure data were collected and used with multitemporal Landsat sensor data and iterative exclusion partial least-squares regression to calibrate six predictive overstory structure models. Model calibrations produced moderate- to high-accuracy results with respective adjusted coefficient of determination and root mean-squared error values as follows: 0.859, 9.3% (canopy cover); 0.855, 2.95 mha−1 (total basal area); 0.741, 11.6% (red oaks relative basal area); 0.781, 11.9% (bur oak relative basal area); 0.861, 3.20 mha−1 (living oak basal area); and 0.833, 9.1% (dead oak relative basal area). We used the resulting structure models for the Sherburne test site to demonstrate how these data could be applied to help managers prioritize areas within management zones for restorative treatments. Although our Sherburne oak savanna test ecosystem is small (12,424 ha) compared with the size of a full Landsat scene (3.4 million ha), resulting structure models can be extended to the whole Landsat scene, which demonstrates how such modeling protocols can be used for repeated (e.g., annual to decadal), regional-scale analysis and assessment to improve management, planning, and implementation of oak savanna restoration efforts elsewhere.

Introduction

The specific structure and function of oak Quercus spp. savanna ecosystems in the North American Midwest were maintained by frequent disturbances (often fire) in pre-European settlement times (Curtis 1959). Subsequent reductions in the frequency of disturbance after European settlement have facilitated rapid conversion of these ecosystems to more closed-canopy forest (Curtis 1959; Vogl 1974; White 1983; Faber-Langendoen and Tester 1993). Combined with land-use conversion and fragmentation, once abundant oak savannas are now rare (<0.02% of presettlement area) in this region (Nuzzo 1986), which increasingly threatens the survival of many savanna-specific ground-layer species (Packard 1988, 1993; Leach and Givnish 1996, 1999; Peterson et al. 2007). In the absence of a robust, natural disturbance regime among remaining oak savannas, prescribed burning as well as mechanical and chemical tree thinning are frequently recommended (often in combination) to control both overstory tree-cover composition and to limit woody plant encroachment including shade-tolerant and fire-intolerant tree species regeneration (Bowles and McBride 1998; Boyles and Aubrey 2006). Regional, cross-ownership management strategies of this nature are needed to restore critical spatial gradients of light, temperature, soil moisture, and soil organic matter (Ko and Reich 1993; Breashears et al. 1997) for recovery and sustenance of the unique mosaic of understory grass and forb species assemblages (Bowles and McBride 1998; Grundel et al. 1998; Bader 2001) and faunal associations that depend on them (Boyles and Aubrey 2006; Peterson et al. 2007; Olechnowski et al. 2009).

Tree basal area (BA, computed as the sum of cross-sectional area of trees measured at 1.37 m above ground per unit area) and canopy cover (CC) are common forest structural and ecological attributes that together affect ground-layer light environment in all forest ecosystems, including oak savannas (Apfelbaum and Haney 1991; Peterson and Reich 2001; Peterson et al. 2007). Basal area is typically used in forest management to describe occupancy of trees and to estimate wood volume (Ginrich 1967). However, BA is also an important ecological attribute used to estimate forest productivity (Burrows et al. 2003), understand ecosystem structure and function (Whittaker et al. 1974; Marshall and Waring 1986; Pacala and Deutschman 1995), as a surrogate for tracking forest carbon (Brown et al. 1989; Phillips et al. 1998), and as a metric to understand the ecology and habitat requirements of forest fauna (Blais 1958; Niemi and Hanowski 1984; Pastor et al. 1998; Sharov et al. 1999). Similarly, CC is used to distinguish different plant communities, assess forest-floor microclimate and light penetration (Jennings et al. 1999; Peterson et al. 2007), and as an indicator used for the conservation of endangered species (Packard 1993; Grundel et al. 1998).

Each of these forest parameters are routinely measured in situ, but are capable of being mapped regionally with the aid of satellite sensor data, such as those routinely acquired by Landsat. For local-scale monitoring and mapping efforts, it may be possible to reliably map some forest-structure parameters, such as CC (Liknes et al. 2010) and mortality (Allain et al. 2011), using National Agriculture Imagery Program (NAIP) data (http://datagateway.nrcs.usda.gov/). The NAIP effort began as a pilot program in 2001 and has been tasked with acquiring 1–2-m spatial-resolution imagery on a near-annual basis since 2003 for the continental U.S. NAIP's natural color imagery (three bands) is always collected during the peak agricultural growing season, with an option to add a near-infrared band if funding is available in a particular year. The true-color, county-wide NAIP mosaics are made freely available for download, but the complimentary near-infrared imagery, if available, is sold at a fee that scales with data volume requested.

Although NAIP is an excellent data source for many forest applications, regional discrimination of oak species is more complex as compared with CC or mortality mapping. The optimum time for spectral discrimination between hardwood tree species, including oaks, is during autumn leaf senescence when leaf pigmentation differences are at a maximum, which makes discrimination easy and unequivocal (Schwaller and Tkach 1985; Wolter et al. 1995). During peak summer months, red oaks and bur oaks are effectively indistinguishable in the electromagnetic wavelengths covered by NAIP. Compounding this, NAIP data are a composite of imagery collected by different contractors (varying sensor formats and acquisition methods), on different days (potential atmospheric differences), and at different times of day (shadow fraction differences). Collectively, this produces spectral artifacts in the NAIP mosaic products that require unique handling on a mosaic-to-mosaic basis (Allain et al. 2011), which effectively precludes automated, regional forest composition mapping that has become a hallmark of Landsat (Schriever and Congalton 1995; Wolter et al. 1995; Vogelmann et al. 2001; Reese et al. 2002; Huang et al. 2010). Although spatially coarse by comparison, Landsat data are frequently described as “synoptic”—capable of seeing large areas at once (one scene is 185 km × 185 km). Moreover, Landsat data are freely available and have spectral, radiometric, and temporal resolution (16-d repeat cycle) characteristic specifically designed for consistent, regional mapping and monitoring efforts. Hence, we explore this potential for regional mapping and monitoring of oak savannas in this research.

Spatially explicit estimates of BA and CC are increasingly sought for landscape analyses to inform regional-scale management decisions (He et al. 1998; Mäkelä and Pekkarinen 2004; McRoberts et al. 2007; Wolter et al. 2008). Satellite imagery can meet this need by providing spatially explicit, multispectral, reflectance measurements that can be related to tree species composition, tree size, stem density, and canopy cover (Franklin and McDermid 1993; Xu et al. 2003; Wolter et al. 2008; 2009; Wolter and Townsend 2011). Here we combine field measurements of oak BA and CC with satellite image data from the Sherburne National Wildlife Refuge (SNWR) in central Minnesota (Figure 1). The SNWR was chosen as a proof-of-concept test ecosystem to demonstrate the capability for repeated (annual to decadal), regional-scale mapping of oak savanna structure. We used iterative exclusion partial least-squares (xPLS) regression (Wolter et al. 2012; Singh et al. 2013) to evaluate 94 Landsat Thematic Mapper (Landsat-5) predictor variables to answer three main questions. First, can Landsat imagery be used to accurately predict critical aspects of oak savanna structure: forest composition, BA, CC, mortality? Second, given a set of hypothetical criteria, how much of the oak savanna within our proof of concept site may benefit from restorative treatments? Third, are Landsat data used in this fashion appropriate for guiding regional-scale management to prioritize and track prescriptive restoration treatments?

Figure 1.

Landsat-5 image (10 March 2008, band 3) of Sherburne NWR in central Minnesota, USA, showing field-plot locations (triangles). The refuge covers an area of 124 km2 (30,700 ac) and consists of oak woodlands, savannas, and prairies.

Figure 1.

Landsat-5 image (10 March 2008, band 3) of Sherburne NWR in central Minnesota, USA, showing field-plot locations (triangles). The refuge covers an area of 124 km2 (30,700 ac) and consists of oak woodlands, savannas, and prairies.

Methods

Oak savanna study site

The 12,424 ha (30,700 ac) SNWR is located on the gently rolling to flat Anoka Sandplain in Central Minnesota (Figure 1; N45.49° Lat., W93.73° Long.). Oak savanna habitat at SNWR is a fire-dependent, dynamic community characterized by scattered trees, or groves of trees, mostly consisting of oak species with absolute ranges in canopy cover from <1 to 100% and BA between <1–35 m2 ha−1; but these values are more typically between 33 and 78% for CC and 8–23 m2 ha−1 BA (Figure 2). Canopy cover can vary at small scales (stand-level) where areas of both scattered trees and areas with groves of trees are present within a single stand. Red oak Quercus rubra or northern pin oak Quercus ellipsoidalis, which commonly hybridize in this area (Swain 1972; Tester 1989), are typically the dominant tree species interspersed with bur oak Quercus macrocarpa (Law et al. 1994; Buchanan 1996). The oldest living trees on the refuge date back to ∼1800 and consist mainly of bur oak (Kipfmuller and Hepola 2007). Understory vegetation consists of both native grasses (25–100% cover) and forbs (5–50% cover; MNDNR 2005). Northern pin oak is occasionally present as a secondary tree species in the overstory or in the shrub layer. The shrub layer is typically composed of American hazelnut Corylus americana or oak sprouts and is patchy to continuous (30–100% cover). The ground layer is composed mostly of native forbs and woody plants characteristic of oak forests with some prairie and savanna forbs and graminoids as well (Wovcha et al. 1995). The spatial extent of prairie openings within this oak savanna ecosystem is <30% (USFWS 2005).

Figure 2.

Ground photo (taken 19 March 2011 at 1545 h UTC) of oak savanna at Sherburne National Wildlife Refuge in central Minnesota, USA, showing tree shadows cast onto snow ground cover.

Figure 2.

Ground photo (taken 19 March 2011 at 1545 h UTC) of oak savanna at Sherburne National Wildlife Refuge in central Minnesota, USA, showing tree shadows cast onto snow ground cover.

Field data

Savanna ecosystems were sampled in 2009 (60 plots) by the authors and other SNWR personnel (Figure 1). Field-plot data consisted of five variable-radius (Grosenbaugh 1952) subplots: one located at the intersection and one at each of the four end points of two crossing, orthogonal, 60-m transect lines placed near the center of large (≥4.4 ha), homogenous stands (Figure 3). We determined plot locations randomly using Geographic Information System (GIS) software and we used a digital map of relic oak savanna from the original, presettlement vegetation map of Minnesota compiled in 1930 (Marschner 1974) to stratify. These stratified random locations all fell in forested zones (BA >2.4 m2 ha−1), which resulted in the following numbers of plots within three BA ranges: 21 low-BA plots (range: 2.4–15.6 m2 ha−1), 16 medium-BA plots (range: 16–23 m2 ha−1), and 23 high-BA plots (range: 24–34.4 m2 ha−1). No two ground plots were closer than 60 m apart. Sufficient stand size and homogeneity assured that stand edge effects were minimized during analysis and that image georectification (image-to-earth) and coregistration (image-to-image) errors were inconsequential. We navigated to each plot center to within 3 m (2dRMS) using a Wide Area Augmentation System (satellite-based, real-time, differential correction) –enabled Garmin eTrex GPS receiver (Garmin International, Inc., Olathe, KS). We estimated overstory tree CC for the plot using 121 densitometer measurements (CC%  =  canopy hits/121) made in 1-m increments along the two plot axes (Figure 3). A densitometer is a hand-held instrument that allows one to view a discrete, vertical point in the canopy above a ground position, where observations are recorded as either canopy or sky. Living and dead overstory-tree-species BA, dominated by oak subgenre Leucobalanus (white oak group—mostly bur oak at SNWR) and Erythrobalanus (red oak group—northern red and pin oak), was measured at each of the five subplots using a metric basal area factor- (BAF-) 2 prism. Hereafter, members of oak subgenus Erythrobalanus are referred to simply as red oaks and Leucobalanus as bur oak. Averaged BA metrics from the five subplots became the respective BA estimates for the full plot. We performed conversion of bur oak BA, red oaks BA, and dead oak BA to relative BA (RBA) by dividing the respective partial BA measures by total plot BA to convert m2 ha−1 to a percent of total BA (bur RBA, red RBA, and dead RBA). We then used total BA, living BA, bur RBA, red RBA, dead RBA, and CC data as dependent variables to develop regression models to estimate and map these parameters using Landsat sensor data (Table S1).

Figure 3.

Field-plot configuration showing the orientation of subplots: one at the center and four the endpoints of two orthogonal 60-m axis, which was designed specifically for integration with 30-m satellite data. Orientation of the plot axis on the landscape was determined randomly. Basal area (sum of tree-trunk cross-section area per unit ground) estimates for each plot were collected at each subplot using angle-count sampling with a metric basal area factor two prism. Canopy-cover measurements were made every meter along the two 60-m axes of each plot using a densitometer (121 measurements/plot).

Figure 3.

Field-plot configuration showing the orientation of subplots: one at the center and four the endpoints of two orthogonal 60-m axis, which was designed specifically for integration with 30-m satellite data. Orientation of the plot axis on the landscape was determined randomly. Basal area (sum of tree-trunk cross-section area per unit ground) estimates for each plot were collected at each subplot using angle-count sampling with a metric basal area factor two prism. Canopy-cover measurements were made every meter along the two 60-m axes of each plot using a densitometer (121 measurements/plot).

Landsat data

Landsat-5 Thematic Mapper (hereafter referred to as Landsat) data have 30-m nominal pixel resolution in six optical bands (blue 450–520 nm [B1], green 520–600 nm [B2], red 630–690 nm [B3], near-infrared 760–900 nm [B4], and two shortwave infrared [SWIR] bands of 1550–1750 nm [B5] and 2080–2350 nm [B7]). We acquired Landsat images (one winter, two spring, one summer, and two autumn) for this study based on temporal proximity to field data collection (summer 2009). One older image (7 October 2003) was acquired to provide contrast between spectral signatures of living oak foliage from 2003 and signatures of leafless oaks that had recently died (prior to 2009). This basic strategy using satellite sensor data capturing leaf-on versus leaf-off (before and after) conditions is routine among studies involving aspects of forest change or mortality (Vogelmann and Rock 1989; Wolter et al. 1995; Collins and Woodcock 1996; Wolter and Townsend 2011). We included winter imagery because the spectral manifestation of tree shadows on snow ground cover was found to empirically relate to overall forest BA in this region (Wolter et al. 2012). Moreover, winter imagery has been used to accurately discriminate forest from nonforest (Wolter et al. 2008). Hence, we used a winter Landsat image (10 March 2008) to calculate the Tasseled Cap brightness index (Crist and Cicone 1984) for distinguishing forest from nonforest prior to structure model development according to Wolter et al. (2008). We selected October imagery to capture peak or near-peak autumn foliage coloration prior to leaf drop as an aid to distinguish oak species (Wolter et al. 1995). Senescent bur oak leaf coloration (typically pale brown) differs substantially from members of the red oaks group (e.g., Quercus velutina, Q. rubra, Q. ellipsoidalis, etc.), which are typically more reddish during autumn senescence because of greater expression of anthocyanin pigments (Boyer et al. 1988; Miller et al. 1991; Lee and Gould 2002). We included April imagery to capture potential leaf-flush timing difference between oak subgenres, while summer imagery provided stable growing-season conditions as spectral contrast to the autumn and spring imagery. Henceforth, the seven Landsat images are referred to using letter codes (Table 1): J (24 January 2009; Figure S1), P (14 April 2009; Figure S2), U (1 June 2009; Figure S3), A (17 August 2008; Figure S4), OF (4 October 2008; Figure S5), OS (7 October 2003; Figure S6), and OT (20 October 2008; Figure S7).

Table 1.

Dates and corresponding letter codes for the Landsat images (path 27, row 28) acquired for the Sherburne National Wildlife Refuge, as well as solar ephemeris at the time of sensor overpass.

Dates and corresponding letter codes for the Landsat images (path 27, row 28) acquired for the Sherburne National Wildlife Refuge, as well as solar ephemeris at the time of sensor overpass.
Dates and corresponding letter codes for the Landsat images (path 27, row 28) acquired for the Sherburne National Wildlife Refuge, as well as solar ephemeris at the time of sensor overpass.

We downloaded all Landsat images from the U.S. Geological Survey's Earth Resources Observation and Science Center web site (source: http://glovis.usgs.gov/) in Universal Transverse Mercator zone 15 coordinates. These 30-m Landsat images are precision-orthorectified (corrected for image horizontal displacement due elevation variation) and georectified by Earth Resources Observation and Science Center using Global Land Cover Facility's GeoCover data (source: www.landcover.org). Landsat images used in this study exhibited excellent pixel coregistration with each other. Each image was converted to top of atmosphere reflectance using published sensor-calibration coefficients (Thome et al. 2004). Because topography is known to influence Earth surface reflectance (Olyphant 1984; Dozier 1989; Culvenor and Coops 1999), all images were topographically adjusted with a lambertian correction methodology (Riaño et al. 2003) using a 30-m digital elevation model (source: http://ned.usgs.gov/) and scene-specific sun position information (Table 1).

In addition to each image's six reflective bands, established indices (simple ratio [SR], Jordan 1969; soil adjusted vegetation index [SAVI], Huete 1988; moisture stress index [MSI], Rock et al. 1986; shortwave infrared visible ratio [SVR], Wolter et al. 2008; Autumn Index [AI], Wolter and Townsend 2011; and Tasseled Cap transformations (brightness, greenness, and wetness; Crist and Cicone 1984) were calculated and used as predictive variables of forest structure in regression analysis (Table 2). These indices are known to exhibit sensitivity to tree species composition, forest BA, tree canopy cover, and tree mortality (Crist and Cicone 1984; Rock et al. 1986; Xu et al. 2003; Wolter et al. 2009; Wolter and Townsend 2011).

Table 2.

Landsat image variables used to calibrate iterative exclusion partial least-squares (xPLS) regression models of oak savanna forest structure at our Sherburne National Wildlife Refuge study site in central Minnesota. There are 42 individual Landsat bands: six bands per image (1–6 below with respective wavelength interval) and seven image dates (J, 24 January 2009; P, 14 April 2009; U, 1 June 2009; A, 17 August 2008; OF, 4 October 2008; OS, 7 October 2003; and OT, 20 October 2008). Image date codes are suffixed 1–6 to identify respective bands. For example, the 14 April 2009 (P) red band (3) is denoted as P3, while 17 August 2008 (A) blue band (1) is denoted A1, and so forth. In a similar fashion, band numbers 7–9 refer to the three Tasseled Cap transformations (Crist and Cicone 1984: brightness [7], greenness [8], and wetness [9]) calculated for each date of imagery. Five additional indices were used as predictors: autumn index, AI, (Wolter and Townsend 2011, calculated only for the three October images); moisture stress index, MSI, (Rock et al. 1986); soil adjusted vegetation index, SAVI, (Huete 1988); simple ratio, SR, (Jordan 1969); and shortwave infrared (SWIR) to visible ratio, SVR, (Wolter et al. 2008). Index abbreviations are suffixed with image date letters (e.g., SVR_A, MSI_P, or AI_OT). The oak mortality variable, MORT, is the 2008–2003 difference in October near-infrared (NIR) bands. Overall, there are 94 initial predictor variables for calibrating forest-structure models.

Landsat image variables used to calibrate iterative exclusion partial least-squares (xPLS) regression models of oak savanna forest structure at our Sherburne National Wildlife Refuge study site in central Minnesota. There are 42 individual Landsat bands: six bands per image (1–6 below with respective wavelength interval) and seven image dates (J, 24 January 2009; P, 14 April 2009; U, 1 June 2009; A, 17 August 2008; OF, 4 October 2008; OS, 7 October 2003; and OT, 20 October 2008). Image date codes are suffixed 1–6 to identify respective bands. For example, the 14 April 2009 (P) red band (3) is denoted as P3, while 17 August 2008 (A) blue band (1) is denoted A1, and so forth. In a similar fashion, band numbers 7–9 refer to the three Tasseled Cap transformations (Crist and Cicone 1984: brightness [7], greenness [8], and wetness [9]) calculated for each date of imagery. Five additional indices were used as predictors: autumn index, AI, (Wolter and Townsend 2011, calculated only for the three October images); moisture stress index, MSI, (Rock et al. 1986); soil adjusted vegetation index, SAVI, (Huete 1988); simple ratio, SR, (Jordan 1969); and shortwave infrared (SWIR) to visible ratio, SVR, (Wolter et al. 2008). Index abbreviations are suffixed with image date letters (e.g., SVR_A, MSI_P, or AI_OT). The oak mortality variable, MORT, is the 2008–2003 difference in October near-infrared (NIR) bands. Overall, there are 94 initial predictor variables for calibrating forest-structure models.
Landsat image variables used to calibrate iterative exclusion partial least-squares (xPLS) regression models of oak savanna forest structure at our Sherburne National Wildlife Refuge study site in central Minnesota. There are 42 individual Landsat bands: six bands per image (1–6 below with respective wavelength interval) and seven image dates (J, 24 January 2009; P, 14 April 2009; U, 1 June 2009; A, 17 August 2008; OF, 4 October 2008; OS, 7 October 2003; and OT, 20 October 2008). Image date codes are suffixed 1–6 to identify respective bands. For example, the 14 April 2009 (P) red band (3) is denoted as P3, while 17 August 2008 (A) blue band (1) is denoted A1, and so forth. In a similar fashion, band numbers 7–9 refer to the three Tasseled Cap transformations (Crist and Cicone 1984: brightness [7], greenness [8], and wetness [9]) calculated for each date of imagery. Five additional indices were used as predictors: autumn index, AI, (Wolter and Townsend 2011, calculated only for the three October images); moisture stress index, MSI, (Rock et al. 1986); soil adjusted vegetation index, SAVI, (Huete 1988); simple ratio, SR, (Jordan 1969); and shortwave infrared (SWIR) to visible ratio, SVR, (Wolter et al. 2008). Index abbreviations are suffixed with image date letters (e.g., SVR_A, MSI_P, or AI_OT). The oak mortality variable, MORT, is the 2008–2003 difference in October near-infrared (NIR) bands. Overall, there are 94 initial predictor variables for calibrating forest-structure models.

Predictive models

Partial least-squares (PLS) regression is frequently used in forest ecology studies to calibrate models of biochemical and biophysical forest parameters using full-spectrum (i.e., hyperspectral) imagery (Coops et al. 2003; Townsend et al. 2003; Martin et al. 2008), multispectral imagery (Wolter et al. 2008, 2009, 2012), and multisensor data fusion (Wolter and Townsend 2011). Partial least-squares regression has become routine in such studies because the algorithm does not assume zero error in the predictor data—often falsely assumed in remote sensing-based studies (Curran and Hay 1986)—and addresses the problem of collinearity (dependence) among multiple independent and dependent variables (Helland 1988). For a detailed discussion of the theory and application of PLS regression, see Geladi and Kowalski (1986); Wold et al. (2009); and Wolter et al. (2008).

Partial least-squares regression was recently adapted to run in an iterative fashion to eliminate weak predictors and to find the most parsimonious model (i.e., fewest predictor variables) that accounts for the largest amount of variation among dependent variables (Wolter et al. 2008, 2009; Wolter and Townsend 2011). This variable-reduction strategy is now known as iterative exclusion PLS or simply xPLS (Wolter et al. 2012; Singh et al. 2013). Partial least-squares regression differs from stepwise and Akaike's Information Criteria or Bayesian Information Criteria methods in that it uses cross-validation to reduce the importance of (but not eliminate) intra- and intercorrelated variables. What distinguishes xPLS from PLS is that xPLS governs the execution of thousands of PLS regressions in an iterative fashion such that predictors showing no or low sensitivity to the dependent variable(s) are completely eliminated; iterations continue until model improvement ceases (reductions in model root mean-squared error; RMSE). Wolter et al. (2008, 2009) and Wolter and Townsend (2011) used xPLS regression with field-plot data and multispectral, multisensor satellite-image predictor variables to calibrate models for estimating forest composition and structure.

In this study, we use xPLS regression to evaluate 94 initial multispectral image predictor variables (Table 2) to calibrate models for estimating CC (Figure S8), total BA (living + dead, Figure S9), living BA (Figure S10), and the relative BA (RBA) of bur oak (Figure S11), red oaks (Figure S12), and dead oak (Figure S13), where RBA  =  specific BA/total BA. Basal area measurements include predominantly oak species (bur, northern red, and northern pin) with an assortment of other forest species, such as Populus tremuloides, Acer saccharum, Fraxinus pennsylvanica, Ulmus americana, Ostrya virginiana, Prunus serotina, with low, overall, relative occurrence at SNWR (Kipfmuller and Hepola 2007).

Spatial autocorrelation is known to weaken some statistical tests by lowering the effective degrees of freedom (Koenig 1999). If present, spatial autocorrelation among regression residuals produces spatially dependent trends that may violate basic assumptions of parametric statistical tests. However, model coefficients of determination (R2) or F-statistics are not rendered invalid (Kerr et al. 2001). We performed isotropic spatial autocorrelation analyses on the regression residuals for total forest BA to determine whether critical levels of spatial autocorrelation were present.

Leave-one-out cross-validation (Gong 1986), using the prediction residual sum-of-squares prediction residual sum-of-squares procedure (Myers 1986), was used to validate all model calibrations using two-thirds of the field-plot data (n  =  40). This method of model validation is analogous to applying the equation to an independent data set as the prediction residual sum-of-squares residual is obtained using observations that are excluded during equation derivation (Sun et al. 2003; Woods et al. 2008). However, independent model validations were also performed using plot data (n  =  20) that were withheld from the xPLS model-calibration process.

Results

Forest-structure model calibration and validation

The isotropic range of spatial autocorrelation among the regression residuals of total forest BA was 30.0 m. Hence, spatial dependence among regression residuals was not a factor in these analyses, because the minimum ground-plot separation distance was 60 m. Leave-one-out cross-validation (henceforth referred to as simply cross-validation; Table 3) of forest-structure models calibrated using ground-plot data and multitemporal Landsat imagery resulted in adjusted coefficients of determination, R2, that ranged between 0.741 for red oaks RBA (RMSE  =  11.6%) and 0.861 for living oak BA (RMSE  =  3.198 m2 ha−1). Cross-validation results for bur oak RBA and red oaks RBA models were similar, with bur oak RBA slightly better (adj. R2  =  0.781, RMSE  =  11.9%) than red oaks RBA. Living BA, CC, and total BA were the three best models according to cross-validation (adj. R2  =  0.861, 0.859, and 0.855; RMSE  =  3.198 m2 ha−1, 9.3%, and 2.948 m2 ha−1, respectively). While the living oak BA model had the highest adjusted coefficient of determination overall, the RMSE (3.198 m2 ha−1) for this model was slightly greater than that reported for total BA (Figure 4). Dead oak RBA calibration and cross-validation resulted in estimation errors of <10% (Figure 4), while adjusted R2 (0.833) was intermediate among the other model calibrations (Table 3).

Figure 4.

Model-calibration results for canopy cover, total basal area, bur oak Quercus macrocarpa relative basal area, red oaks Quercus rubra and Q. ellipsoidalis relative basal area, living oak basal area, and dead oak relative basal area. Relative basal area (RBA) estimates represent percentages of the 2009 plot-wise total basal area (BA) on a scale from 0 to 1, where BA is the sum of all tree-trunk cross-section areas per unit ground area.

Figure 4.

Model-calibration results for canopy cover, total basal area, bur oak Quercus macrocarpa relative basal area, red oaks Quercus rubra and Q. ellipsoidalis relative basal area, living oak basal area, and dead oak relative basal area. Relative basal area (RBA) estimates represent percentages of the 2009 plot-wise total basal area (BA) on a scale from 0 to 1, where BA is the sum of all tree-trunk cross-section areas per unit ground area.

Table 3.

Results of iterative exclusion partial least-squares (xPLS) regression model development and leave-one-out cross-validation using summer 2009 field data from 40 randomly selected plots (predicted residual sum of squares [PRESS]  =  0.5075) for our oak savanna study site in central Minnesota. Independent model validation (IV) results (performed using 20 withheld plots) are shown in italics. Both the coefficient of determination (R2) and adjusted (Adj.) R2 values are presented for the cross-validations and the independent validations (IV), along with cross-validated model root mean-squared errors (RMSE), IV residual errors (Pred. E), IV F-statistics (F-stat), and IV goodness-of-fit (P-value). Relative basal area (RBA) is a percentage of the total basal area (BA) that a tree species or forest type represents, calculated as the species BA divided by the total BA (e.g., red oaks Quercus rubra and Q. ellipsoidalis RBA  =  red oaks BA/total BA).

Results of iterative exclusion partial least-squares (xPLS) regression model development and leave-one-out cross-validation using summer 2009 field data from 40 randomly selected plots (predicted residual sum of squares [PRESS]  =  0.5075) for our oak savanna study site in central Minnesota. Independent model validation (IV) results (performed using 20 withheld plots) are shown in italics. Both the coefficient of determination (R2) and adjusted (Adj.) R2 values are presented for the cross-validations and the independent validations (IV), along with cross-validated model root mean-squared errors (RMSE), IV residual errors (Pred. E), IV F-statistics (F-stat), and IV goodness-of-fit (P-value). Relative basal area (RBA) is a percentage of the total basal area (BA) that a tree species or forest type represents, calculated as the species BA divided by the total BA (e.g., red oaks Quercus rubra and Q. ellipsoidalis RBA  =  red oaks BA/total BA).
Results of iterative exclusion partial least-squares (xPLS) regression model development and leave-one-out cross-validation using summer 2009 field data from 40 randomly selected plots (predicted residual sum of squares [PRESS]  =  0.5075) for our oak savanna study site in central Minnesota. Independent model validation (IV) results (performed using 20 withheld plots) are shown in italics. Both the coefficient of determination (R2) and adjusted (Adj.) R2 values are presented for the cross-validations and the independent validations (IV), along with cross-validated model root mean-squared errors (RMSE), IV residual errors (Pred. E), IV F-statistics (F-stat), and IV goodness-of-fit (P-value). Relative basal area (RBA) is a percentage of the total basal area (BA) that a tree species or forest type represents, calculated as the species BA divided by the total BA (e.g., red oaks Quercus rubra and Q. ellipsoidalis RBA  =  red oaks BA/total BA).

Independent validation of the forest-structure models using data from 20 plots withheld from the xPLS calibration procedure show slight deviations from leave-one-out cross-validation results (Table 3). The three RBA models (bur oak, red oaks, and dead oak) showed similar error levels between the two validations used, but the adjusted R2 values under the independent validation for the red oaks RBA model was 9.1% greater and the dead oak RBA model 7.2% lower than that reported under cross-validation. However, the adjusted R2 values under each validation method were similar for the bur oak RBA model (Table 3). Independent model validation showed lower adjusted R2 values (−7.0 to −14.2%) than cross-validation for living BA, total BA, and CC accompanied by slight to modest increases in error levels: CC, +4.9%; living BA, +0.69 m2 ha−1; and total BA, +1.39 m2 ha−1.

Model reduction and important image-based predictor variables

Of the original set of 94 multispectral, image-based predictor variables (Table 2), 23 were retained for final model calibrations. This represents a model reduction of 75.5% over the initial full model. The six dependent structure variables (ground-plot data) were all placed together in a matrix and regressed simultaneously against the independent, image-based, predictor variable matrix (Wolter et al. 2008; Wold et al. 2009); therefore, all of the 23 retained image variables are used in each of the six structure models. However, individual predictor variable coefficients vary in sign and magnitude among the different models, which gives an indication of the relative importance of the retained predictors among the six respective structure model calibrations (Figure 5). Image predictor variables are reported below with a + or − sign preceding the variable name to indicate direction of coefficient loading.

Figure 5.

Scaled component loadings for the reduced set of 23 image variables used by iterative exclusion partial least-squares (xPLS) regression to calibrate forest-structure models. Positive and negative loadings are depicted as red and blue filled circles, respectively, while circle size indicates loading magnitude. Landsat shortwave infrared band 7 (2.08–2.35 µm) is indicated with suffix number 6, while Tasseled Cap brightness, greenness, and wetness transformations are denoted with number suffixes 7, 8, and 9, respectively. Image dates are indicated with letter codes: J, 24 January 2009; P, 14 April 2009; U, 1 June 2009; A, 17 August 2008; OF, 4 October 2008; OS, 7 October 2003; and OT, 20 October 2008.

Figure 5.

Scaled component loadings for the reduced set of 23 image variables used by iterative exclusion partial least-squares (xPLS) regression to calibrate forest-structure models. Positive and negative loadings are depicted as red and blue filled circles, respectively, while circle size indicates loading magnitude. Landsat shortwave infrared band 7 (2.08–2.35 µm) is indicated with suffix number 6, while Tasseled Cap brightness, greenness, and wetness transformations are denoted with number suffixes 7, 8, and 9, respectively. Image dates are indicated with letter codes: J, 24 January 2009; P, 14 April 2009; U, 1 June 2009; A, 17 August 2008; OF, 4 October 2008; OS, 7 October 2003; and OT, 20 October 2008.

For CC, 1 June 2009 SVR (−SVR_U) and 4 October 2008 Tasseled Cap wetness (+OF9) were the two strongest image predictors, followed by 14 April 2009 Tasseled Cap greenness (−P8) and 1 June 2009 Tasseled Cap brightness (−U7). In terms of sign and magnitude, these top two variables were also the leading variables for estimation of total BA, living BA, and bur oak RBA (Figure 5). The top two predictor variables for red oak RBA (northern red and pin oak), were 20 October 2008 Tasseled Cap greenness (+OT8) and 14 April 2009 Tasseled Cap greenness (+P8), followed by 20 October 2008 shortwave infrared band 6 (−OT6) and 1 June 2009 near infrared (−U4). The strongest loadings for calibrating the dead oak RBA model were +SVR_U and −OF9, which were similar to that observed for other models, but with opposite signs (Figure 5). The next two strongest predictor variables were 1 June 2009 Tasseled Cap brightness (+U7) and 7 October 2003 moisture stress index (−MSI_OS). Of the stronger variables from the set of 23 for calibrating the dead oak RBA model, three were from the 7 October 2003 image (−MSI_OS, +OS9, and –SVR_OS). These 2003 image variables were stronger for calibrating dead oak RBA than for any other structure model (Figure 5).

Current oak savanna structure at Sherburne NWR

In all, 2,898 trees were tallied within our 60 field plots, with red oaks consisting of 33.6%, bur oak 39.7%, other species 4.0%, and dead oaks 22.6%. Most of the dead trees tallied were red oaks (68.2%), followed by bur oak (22.9%) and other species (8.9%). Within the 12,424 ha of land at SNWR, ∼45.3% supports some level of tree cover. Across 10 CC intervals (each representing a range in CC of 10%) between 1 and 100 (referred to hereafter as CC classes), the 41–50% CC class encompasses the greatest proportion of oak savanna at SNWR, with a mean total BA  =  11.7 m2 ha−1 (Figure 6). Greater proportions of oak savanna at SNWR occur in the next three higher CC percent classes (51–60 CC class, 14.3%; 61–70 CC class, 13.4%; and 71–80 CC class, 12.8%) than that falling in the lower CC classes (11–20 CC class, 4.2%; 21–30 CC class, 9.0%; and 31–40 CC class, 13.4%). Bur oak RBA exceeds that of the red oaks RBA (northern red and pin) within CC levels from ∼21 to 57% (total BA: 5.7–15.5 m2 ha−1); otherwise, red oaks RBA is greater than that of bur oak (Figure 6).

Figure 6.

Total basal area (BA) and the three relative basal area (RBA) structure estimates (bur oak Quercus macrocarpa, red oaks Quercus rubra and Q. ellipsoidalis, and dead oak [all Q. spp.]) averaged across canopy cover (CC) intervals in increments of 10% (referred to here as classes) for oak savanna at Sherburne National Wildlife Refuge (SNWR). The 41–50% CC class is the largest in area (859.7 ha) and accounts for ∼15.3% of the oak savanna wooded area at SNWR, while the 1–10% CC class has the least area (70 ha) at 1.2%. All relative basal area (RBA) estimates represent percentages of the 2009 plot-wise total basal area (BA) on a scale from 0 to 1, where BA is the sum of all tree-trunk cross-section areas per unit ground area.

Figure 6.

Total basal area (BA) and the three relative basal area (RBA) structure estimates (bur oak Quercus macrocarpa, red oaks Quercus rubra and Q. ellipsoidalis, and dead oak [all Q. spp.]) averaged across canopy cover (CC) intervals in increments of 10% (referred to here as classes) for oak savanna at Sherburne National Wildlife Refuge (SNWR). The 41–50% CC class is the largest in area (859.7 ha) and accounts for ∼15.3% of the oak savanna wooded area at SNWR, while the 1–10% CC class has the least area (70 ha) at 1.2%. All relative basal area (RBA) estimates represent percentages of the 2009 plot-wise total basal area (BA) on a scale from 0 to 1, where BA is the sum of all tree-trunk cross-section areas per unit ground area.

Based on visual interpretation of modeled results, CC and total BA distribution patterns at SNWR appear to be similar (Figure 7), while distributions of bur oak and red oaks RBA appear to differ substantially (Figure 7). Distribution of bur oak with higher RBA levels is concentrated in the north-central and northwestern regions of SNWR, while red oaks at similar RBA levels are concentrated in the west-central and southeastern areas. The distribution of high red oaks RBA appears to correspond with high percent CC levels, while bur RBA distribution is less distinct with reference to percent CC. The RBA of dead oak is generally low across SNWR, with concentrations of higher RBA in the southeastern areas where forest total BA is generally low (Figure 7).

Figure 7.

Modeled oak savanna structure results for Sherburne National Wildlife Refuge: percent canopy cover (CC), total basal area (BA, m2 ha−1), bur oak Quercus ellipsoidalis percent relative basal area (RBA  =  species BA/total BA), red oaks percent RBA, and an example of how these Landsat-derived forest-structure data could be used to inform management decisions and monitor management outcomes. Basal area refers to the sum of tree-trunk cross-section areas (measured 1.37 m above ground) per unit of ground area. These maps are projected using Universal Transverse Mercator (World Geodetic System 1984) zone 15 coordinates, with northings (y-axes) and eastings (x-axes) shown for scale and north orientation.

Figure 7.

Modeled oak savanna structure results for Sherburne National Wildlife Refuge: percent canopy cover (CC), total basal area (BA, m2 ha−1), bur oak Quercus ellipsoidalis percent relative basal area (RBA  =  species BA/total BA), red oaks percent RBA, and an example of how these Landsat-derived forest-structure data could be used to inform management decisions and monitor management outcomes. Basal area refers to the sum of tree-trunk cross-section areas (measured 1.37 m above ground) per unit of ground area. These maps are projected using Universal Transverse Mercator (World Geodetic System 1984) zone 15 coordinates, with northings (y-axes) and eastings (x-axes) shown for scale and north orientation.

Discussion

Landsat and oak savanna structure mapping

Landsat sensor data have been used in numerous studies to quantify forest structural parameters (Franklin 1986; Horler and Ahern 1986; Woodcock et al. 1994; Franco-Lopez et al. 2001; Xu et al. 2003; McRoberts et al. 2007; McRoberts 2008; Wolter et al. 2009), but use of iterative PLS regression (xPLS) as a data-fusion and model-calibration optimization technique for Landsat data analysis is relatively new (Wolter et al. 2009, 2012; Wolter and Townsend 2011; Singh et al. 2013). Wolter et al. (2012) used xPLS regression with multispectral, multitemporal winter Landsat data to model one element of oak savanna structure (total BA) with strong results (R2  =  0.898, RMSE  =  2.73 m2 ha−1). However, our use of xPLS regression with multiseason Landsat sensor data for more comprehensive oak savanna structure mapping, including oak mortality, is novel. Moreover, simultaneous calibration of the six forest-structure models (CC, total BA, living BA, bur oak RBA, red oaks RBA, and dead oak RBA) using 23 of the original 94 image variables (75.5% model reduction) is notable because the process is fully automatic and repeatable. However, success of these analyses also hinges on the dates of Landsat imagery and indices used as inputs to the xPLS algorithm, which should always be coupled with knowledge of the biology and phenology of forest species under investigation (Boyer et al. 1988; Miller et al. 1991; Wolter et al. 1995; Wolter et al. 2012).

Automated selection of image variables by xPLS regression during model calibration highlights the strengths of several image variables for quantifying oak savanna structure at SNWR—specifically, the importance of SVR (Wolter et al. 2008) and Tasseled Cap wetness (Crist and Cicone 1984) for nearly all elements of structure modeled (Figure 5). Landsat wetness is a linear contrast of SWIR with the near-infrared and the visible portion of the electromagnetic spectrum and is known to provide highly accurate information on forest canopy structure due, in part, to relative insensitivity to topographical effects (Cohen and Spies 1992; Cohen et al. 1995; Hansen et al. 2001). Similarly, SVR is the ratio of SWIR to visible, while MSI (Rock et al. 1986) is a ratio of SWIR to near-infrared (Table 2); both are known to be strong predictors of forest basal area and tree species relative to basal area in subboreal forests (Wolter et al. 2008; Wolter and Townsend 2011). The SWIR component of these indices is known to vary inversely with forest biomass due to increases in shadowing with increasing heterogeneity in forest structure (Berner et al. 2012). Percent CC increases linearly with total BA (Jennings et al. 1999); therefore, it is not surprising that SWIR-based variables were of similar magnitude and sign for estimating these structural parameters (Figure 5).

Predictor variable loading (magnitude and sign) for bur oak RBA and red oaks RBA may indicate phenology differences between stands of the two oak subgenres at SNWR. For instance, of all the image variables retained during xPLS model calibration, each of the seven variables from 14 April 2009 (most for any one date) show opposite loading signs between the two oak subgenres (Figure 5). For living versus dead oak, the sign of respective loadings on almost all the retained image predictors were opposed, which is logical. The opposed April image (P) predictor loading signs for bur and the red oaks, however, is puzzling because both subgenres appeared, upon visual inspection of April imagery, as if they had not yet flushed leaves. Alternatively, it is possible that ground-layer green-up had begun prior to image acquisition. Hence, the Landsat sensor may be detecting potential differences in understory vegetation structure, composition, and phenology associated with the two oak subgenres (Leach and Givnish 1999; Schulte et al. 2011), which would be optimal in the absence of fully flushed oak leaves. However, we have neither ground-layer vegetation data nor specific phenology characteristics required to test this hypothesis.

Oak savanna management

Similar to oak savanna trends reported elsewhere in Minnesota (Margoles 2009), the apparent increase in the proportion of red oaks over bur oak at SNWR has resulted in a disproportionate spatial extent of canopy cover classes >61% (Figure 6), which agrees with tree composition (n  =  55 point samples) reported by Kipfmuller and Hepola (2007). Similar to other oak savanna ecosystems (Leach and Givnish 1999), the distributions of bur oak at SNWR appears, based on visual inspection, to correspond with soils having higher silt content, such as Becker loam and Isanti loamy fine sand. Soils at SNWR are predominantly sandy, but the distribution of red oaks does not appear to be distinctly associated with any one soil type or series (see USFWS 2005). Patterns of red oak abundance at SNWR (especially northern pin) are likely aligned more closely to fire-disturbance history (White 1983; Crow 1988) and, potentially, ground-water availability (P. Drobney, U.S. Fish and Wildlife Service, Neal Smith National Wildlife Refuge, pers. comm.) than soil type alone. Whether bur oak regeneration has declined in favor of fire-intolerant northern red and northern pin oaks at SNWR over past decades due to changes in fire disturbance, competition for light from invasive understory species, herbivory from white-tailed deer Odocoileus virginianus, climate trends, or a combination of these factors, as observed elsewhere (Kipfmuller and Hepola 2007; Ziegler et al. 2008; Schulte et al. 2011), is unclear. What is clear is that red oaks are increasing in proportion and generally outcompeting bur oaks at SNWR.

Regardless of causal factors, there are no precise analogs of oak savanna remaining in this region from which to model or guide restorative efforts (Bowles and McBride 1998); although, a few highly diverse savannas do exist (e.g., Timberhill in southern Iowa; Wilhelm and Rericha 2007) that provide good benchmark examples of species–light relationships and ecosystem functioning. In the absence of extant savanna benchmarks, restoration targets are often identified by old aerial photos, General Land Office maps, surveyor's comments from original land surveys, herbaria, dendrochronology, evidence of plant communities in archeological collections and soil sedimentation, and other historical records (USFWS 2012).

In order to achieve regional restoration objectives based on historical references, appropriate management tools should be implemented broadly and, ideally, across ownership boundaries to mimic historical disturbances that influenced and maintained savanna ecosystems. For example, because many savannas are highly degraded, with fire-intolerant species dominating the overstory and understory, prescribed burning is used as a tool to mimic historical fires. Prescribed burning accomplishes several things: it promotes the growth of native wildflowers and native grasses; retards competition from exotic grasses and fire-intolerant trees and shrubs; promotes higher water-table levels; and converts heavily wooded or closed-canopy forest ecosystems to a more open-canopy condition that more closely approximates oak savannas in this region (USFWS 2005). The severity of degradation may warrant additional restoration techniques such as tree thinning operations. At our SNWR case-study ecosystem, tree-thinning operations are often implemented prior to prescribed burning to expedite restoration, focusing efforts on closed-canopy oak savanna forest with high BA values and favoring bur oak trees for retention. Removal of undesirable trees from the overstory creates open canopy conditions facilitating light penetration and changes species composition. Depending on specific management objectives or definitions (Bray 1955; Curtis 1959; Botts et al. 1994), these Landsat-based structure maps can be developed relatively quickly, and repeatedly (annual to decadal intervals), for large regions and serve as a framework from which to help target or monitor regional- or local-scale restorative treatments.

In our local-scale test ecosystem for example, one of the management objectives in the Comprehensive Conservation Plan for SNWR is conversion of existing woodlands to oak savanna (USFWS 2005). Basal area (BA) and CC attributes among oak savannas in this region typically ranged from 1 to 11.5 m2 ha−1 (5–50 ft2 acre−1) and 10–70%, respectively. If an upper BA limit of 11.5 m2 ha−1 is used as the threshold above which restorative treatments are applied, then 66.5% (i.e., 3,698 ha) of the oak savanna at SNWR qualifies for treatment. Rather, if a 50% upper CC threshold is used, then 57.7% (3,205 ha) qualifies for treatment. Using the intersection of these two ground areas (11.5 m2 ha−1 BA and 50% CC) qualifies 55.9% (3,106 ha) of the oak savanna at our SNWR test study for restorative treatments. Furthermore, these Landsat-based structure maps can provide spatial reference to assist managers in focusing restoration efforts, such as tree-thinning operations, to areas with high BA and CC values. Managers at SNWR may also wish to apply their resources to those locations where red oaks make up >50% of the total forest BA. In this case, 28% (1,556 ha) of forest at SNWR qualifies for savanna restoration treatment (Figure 7). Use of these Landsat-derived forest-structure data and other ancillary GIS data can effectively help managers prioritize oak savanna ecosystems for restorative treatments, at both the local scale (such as our SNWR study site), but also for regional-scope management. With this said, it is important to note that while overstory CC and BA are certainly salient factors affecting understory light environment, specific determination of whether a site is truly “degraded,” is also a function of the quality and composition of understory vegetation present, regardless of overstory factors (P. Drobney, pers. comm.). Hence, these Landsat data should be used as a template to guide local- and regional-scale restoration efforts.

Within our SNWR test-study ecosystem, managers (E. Berkley and G. Dehmer, U.S. Fish and Wildlife Service, pers. obs.) estimate that approximately 50% of the mortality mapped using 2003 and 2009 Landsat image data (Figure 7) was due to oak wilt disease caused by the fungus Ceratocystis fagacearum, which affects red oaks more readily than white oaks (Brinkman 1952; Boyce 1959; Jacobi and MacDonald 1980). Once oak wilt spreads to an area, such as SNWR, the amount of damage and rate of spread is a function of stem density, tree age, and site differences (Menges and Kuntz 1985). Although practical methods for retarding the spread of oak wilt outside of urban settings do not exist, the Landsat-based forest-structure maps produced in this study could be used to facilitate a better understanding of the variability in rates of disease spread (Juzwik 2000; Downing et al. 2009).

Conclusions

Tree-species composition, distribution, mortality, basal area (BA), and canopy cover (CC) are important forest-structure parameters that are intrinsically linked to oak savanna restoration ecology in the upper Midwest region of the United States. We have demonstrated that predictive forest-structure models, calibrated using xPLS regression, simple ground measurements, and readily accessible Landsat image data are capable of producing moderate to highly accurate, spatially explicit, estimates for these parameters. Although we chose a relatively small oak savanna ecosystem for this proof of concept, our intent is for repeated, regional application. With regard to these Landsat-derived oak structure estimates (in the context of savanna restoration), one can only speculate whether a site is truly “degraded,” because much also depends on specific understory vegetation composition. However, these Landsat-based structure estimates can be used both locally and regionally in a variety of ways to 1) visualize and quantify forest composition trends, 2) prioritize potential areas for oak savanna restoration efforts (e.g., identification of overgrown oak savanna for thinning), and 3) monitor BA and CC of oak savanna through space and time as restoration treatments are implemented. Forest mortality models predicting the relative BA of living versus dead oaks may also be useful for determining rates of mortality (whether from prescribed fire or disease). This Landsat-based method of forest-structure estimation is robust and may be extended as a large-scale assessment tool to guide and improve management, planning, and implementation of oak savanna restoration efforts at multiple scales.

Supplemental Material

Please note: The Journal of Fish and Wildlife Management is not responsible for the content or functionality of any supplemental material. Queries should be directed to the corresponding author for the article.

Table S1. Sherburne National Wildlife Refuge 2009 field-plot data—average values calculated from the five subplots per plot. Also, the associated Universal Transverse Mercator zone 15 (WGS84) plot center coordinates (meters) used to extract Landsat band and index values.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S1 (14 KB XLSX).

Figure S1. Subset of the 24 January 2009 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S2 (13.8 MB TIF).

Figure S2. Subset of the 14 April 2009 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S3 (13.8 MB TIF).

Figure S3. Subset of the 1 June 2009 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S4 (13.8 MB TIF).

Figure S4. Subset of the 17 August 2008 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S5 (13.8 MB TIF).

Figure S5. Subset of the 4 October 2008 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S6 (13.8 MB TIF).

Figure S6. Subset of the 7 October 2003 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S7 (13.8 MB TIF).

Figure S7. Subset of the 20 October 2008 Landsat-5 Thematic Mapper imagery that was used to calibrate biophysical models of oak forest structure for the Sherburne National Wildlife Refuge test site in central Minnesota. These Landsat data are provided in Universal Transverse Mercator zone 15 projection and WGS84 datum at 30-m spatial resolution.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S8 (13.8 MB TIF).

Figure S8. Modeled percent canopy cover estimates for the Sherburne National Wildlife Refuge test site in central Minnesota. Percent canopy cover estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S9 (4097 KB TIF).

Figure S9. Modeled estimates of total oak basal area for the Sherburne National Wildlife Refuge test site in central Minnesota. Basal area estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates and represent the trunk cross-section area (square meters measured at 1.37 m above ground) of all forest trees per unit ground area (m2 ha–1).

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S10 (4097 KB TIF).

Figure S10. Modeled estimates of living oak basal area for the Sherburne National Wildlife Refuge test site in central Minnesota. Basal area estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates and represent the trunk cross-section area (square meters measured at 1.37 m above ground) of living oak trees (Q. rubra, Q. ellipsoidalis, and Q. macrocarpa) per unit ground area (m2 ha−1).

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S11 (4097 KB TIF).

Figure S11. Modeled estimates of relative basal area of bur oak for the Sherburne National Wildlife Refuge test site in central Minnesota. Relative basal area estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates and represent the ratio of trunk cross-section area (square meters measured at 1.37 m above ground) of bur oak Q. macrocarpa per unit area to the total cross-section area of all tree trunks (live + dead) in the same units (m2 ha−1).

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S12 (4097 KB TIF).

Figure S12. Modeled estimates of relative basal area of red oaks for the Sherburne National Wildlife Refuge test site in central Minnesota. Relative basal area estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates and represent the ratio of cross-section area (square meters measured at 1.37 m above ground) of red oaks Q. rubra and Q. ellipsoidalis per unit area to the total cross-section area of all tree trunks (live + dead) in the same units (m2 ha−1).

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S13 (4097 KB TIF).

Figure S13. Modeled estimates of relative basal area of dead oaks for the Sherburne National Wildlife Refuge test site in central Minnesota. Relative basal area estimates were generated at 30-m spatial resolution in Universal Transverse Mercator zone 15, WGS84 coordinates and represent the ratio of cross-section area (square meters; measured at 1.37 m above ground) of dead oak trees per unit area to the total cross-section area of all stems (live + dead) in the same units (m2 ha−1).

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S14 (4097 KB TIF).

Reference S1. Buchanan H. 1996. Oak savanna restoration at Myre–Big Island State Park. Student On-Line Journal 1(2):1–4.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S15; also available at http://conservancy.umn.edu/bitstream/58421/1/1.2.Buchanan.pdf (269 KB PDF).

Reference S2. Law JR, Johnson PS, Houf G. 1994. A crown cover chart for oak savannas. Columbia, MO: USDA Forest Service, North Central Forest Experiment Station. Technical brief TB-NC-2.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S16; also available at http://ncrs.fs.fed.us/pubs/tb/tb2/techbrf2.html (211 KB PDF).

Reference S3. Kipfmuller KF, Hepola T. 2007. Fire history and age structure analysis in the Sherburne National Wildlife Refuge: establishing reference conditions in a remnant oak savanna woodland. Fort Collins, CO: USDA Forest Service, Rocky Mountain Research Station. USDA Forest Service Proceedings RMRS-P-46CD, 507–514.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S17; also available at http://www.fs.fed.us/rm/pubs/rmrs_p046/rmrs_p046_507_514.pdf (890 KB PDF).

Reference S4. [USFWS] U.S. Fish and Wildlife Service. 2005. Comprehensive conservation plan: Sherburne National Wildlife Refuge. Washington, D.C.: U.S. Department of Interior.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S18; also available at http://www.fws.gov/Midwest/Planning/sherburne/ccp/FinalCCP.pdf (2563 KB PDF).

Reference S5. Margoles SS. 2009. Perspectives on oak savanna restoration in Minnesota: a dendroecological approach. Doctoral dissertation. Minneapolis: University of Minnesota.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S19; also available at http://conservancy.umn.edu/bitstream/60108/1/Margoles_Sarah_December%202009.pdf (3641 KB PDF).

Reference S6. Wilhelm G, Rericha L. 2007. Timberhill savanna assessment of landscape management. Elmhurst, IL: Conservation Research Institute.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S20; also available at http://wwx.inhs.illinois.edu/files/1913/4021/3314/Timberhill_Final.pdf (5553 KB PDF).

Reference S7. [USFW] U.S. Fish and Wildlife Service. 2012. Working definition of “oak savanna” for restoration efforts at Sherburne NWR. Unpublished addendum to USFWS (2005) Comprehensive Conservation Plan.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S21 (144 KB PDF).

Reference S8. Downing MC, Thomas V, Juzwik J, Appel DN, Reich RM, Camilli K. 2009. Using classification tree analysis to predict oak wilt distribution in Minnesota and Texas. Pages 67–83 in Billings RF, Appel DN, editors. Proceedings of the national oak wilt symposium. Newtown Square, PA: USDA Forest Service, Northern Research Station. Texas Forest Service Publication 166:61–69.

Found at DOI: http://dx.doi.org/10.3996/022013-JFWM-010.S22; also available at http://texasoakwilt.org/Professionals/NOWS/conference_assets/NOWS_Proceedings.pdf#page=6 (2785 KB PDF).

Acknowledgments

This research was supported by funding from the U.S. Fish and Wildlife Service. The authors would also like to acknowledge the three anonymous Journal reviewers as well as the Subject Editor whose comments and suggestions greatly improve the quality of this manuscript.

Any use of trade, product, or firm names is for descriptive purposes only and does not imply endorsement by the U.S. Government.

References

Allain
M
,
Nguyen
A
,
Johnson
E
,
Williams
E
,
Tsai
S
,
Prichard
S
,
Freed
T
,
Skiles
JW
.
2011
.
A geospatial assessment of mountain pine beetle infestations and their effect on forest health in Okanogan–Wenatchee National Forest
. Pages
20
31
in
D
,
Lichti
editor
.
Proceedings of the American Society of Photogrammetry and Remote Sensing
.
Bethesda, Maryland
:
American Society for Photogrammetry and Remote Sensing
.
Apfelbaum
SI
,
Haney
AW
.
1991
.
Management of degraded oak savanna remnants in the upper Midwest: preliminary results from three years of study
. Pages
81
90
in
Burger
GV
,
Ebinger
JE
,
Wilhelm
GS
,
editors
.
Proceedings of the oak woods management workshop
.
Charleston
:
Eastern Illinois University
.
Bader
BJ
.
2001
.
Developing a species list for oak savanna/oak woodland restoration
.
Ecological Restoration
19
:
242
250
.
Berner
LT
,
Beck
PSA
,
Loranty
MM
,
Alexander
HD
,
Mack
MC
,
Goetz
SJ
.
2012
.
Cajander larch (Larix cajanderi) biomass distribution, fire regime and post-fire recovery in northeastern Siberia
.
Biogeosciences Discussions
9
:
7555
7600
.
Blais
JR
.
1958
.
The vulnerability of balsam fir to spruce budworm attack in northwestern Ontario, with special reference to the physiological age of the tree
.
Forestry Chronicle
34
:
405
422
.
Botts
P
,
Haney
A
,
Holland
K
,
Packard
S
.
1994
.
Midwest oak ecosystems recovery plan. Technical report for the 1993 Midwest Oak Savanna Conference.
.
Chicago
:
The Nature Conservancy, Illinois Field Office
.
Bowles
ML
,
McBride
LL
.
1998
.
Vegetation, composition, structure, and chronological change in a decadent Midwestern North American savanna remnant
.
Natural Areas Journal
18
:
14
27
.
Boyce
JS
.
1959
.
Oak wilt spread in control-treated and untreated counties in the southern Appalachians
.
Journal of Forestry
57
:
660
661
.
Boyer
M
,
Miller
J
,
Belanger
M
,
Hare
E
,
Wu
J
.
1988
.
Senescence and spectral reflectance in leaves in northern pin oak (Quercus palustris Muench.)
.
Remote Sensing of Environment
25
:
71
87
.
Boyles
JG
,
Aubrey
DP
.
2006
.
Managing forests with prescribed fire: implications for a cavity-dwelling bat species
.
Forest Ecology and Management
222
:
108
115
.
Bray
JR
.
1955
.
The savanna vegetation of Wisconsin and an application of the concepts of order and complexity to the field of ecology. Doctoral dissertation
.
Madison
:
University of Wisconsin
.
Breashears
DD
,
Rich
PM
,
Barnes
FJ
,
Cambell
K
.
1997
.
Overstory-imposed heterogeneity in solar radiation and soil moisture in a semi-arid woodland
.
Ecological Applications
7
:
1201
1215
.
Brinkman
A
.
1952
.
Oak wilt mortality in an unmanaged Iowa forest
.
Journal of Forestry
50
:
941
942
.
Brown
S
,
Gillespie
AJR
,
Lugo
AE
.
1989
.
Biomass estimation methods for tropical forests with applications to forest inventory data
.
Forest Science
35
881
902
.
Buchanan
H
.
1996
.
Oak savanna restoration at Myre–Big Island State Park
.
Student On-Line Journal
1
(
2
):
1
4
. .
Burrows
SN
,
Gower
ST
,
Norman
JM
,
Diak
G
,
Mackay
DS
,
Ahl
DE
,
Clayton
MK
.
2003
.
Spatial variability of aboveground net primary production for a forested landscape in northern Wisconsin
.
Canadian Journal of Forest Research
33
:
2007
2018
.
Cohen
WB
,
Spies
TA
.
1992
.
Estimating structural attributes of douglas-fir/western hemlock forest stands from Landsat and SPOT imagery
.
Remote Sensing of Environment
41
:
1
17
.
Cohen
WB
,
Spies
TA
,
Fiorella
M
.
1995
.
Estimating the age and structure of forests in a multi-ownership landscape of western Oregon, USA
.
International Journal of Remote Sensing
16
:
721
746
.
Collins
JB
,
Woodcock
CE
.
1996
.
An assessment of several linear change detection techniques for mapping forest mortality using multitemporal Landsat TM data
.
Remote Sensing of Environment
56
:
66
77
.
Coops
NC
,
Smith
ML
,
Martin
ME
,
Ollinger
SV
.
2003
.
Prediction of eucalypt foliage nitrogen content from satellite-derived hyperspectral data
.
IEEE Transactions on Geoscience and Remote Sensing
41
:
1338
1346
.
Crist
EP
,
Cicone
RC
.
1984
.
A physically based transformation of Thematic Mapper data—the TM Tasseled Cap
.
IEEE Transactions on Geoscience and Remote Sensing
GE-22
:
256
263
.
Crow
TR
.
1988
.
Reproductive mode and mechanisms for self-replacement of northern red oak (Quercus rubra): a review
.
Forest Science
34
:
19
40
.
Culvenor
DS
,
Coops
NC
.
1999
.
A forest scene simulation model based on geometric optics
.
Abstracts of the second international workshop on multiple-angle measurements and models, Poster Presentation, Ispra, Italy
.
Curran
P
,
Hay
A
.
1986
.
The importance of measurement error for certain procedures in remote sensing at optical wavelengths
.
Photogrammetric Engineering and Remote Sensing
52
:
229
241
.
Curtis
JT
.
1959
.
The vegetation of Wisconsin: an ordination of plant communities
.
Madison
:
University of Wisconsin Press
.
Downing
MC
,
Thomas
V
,
Juzwik
J
,
Appel
DN
,
Reich
RM
,
Camilli
K
.
2009
.
Using classification tree analysis to predict oak wilt distribution in Minnesota and Texas
. Pages
67
83
in
Billings
RF
,
Appel
DN
,
editors
.
Proceedings of the national oak wilt symposium
.
Newtown Square, PA
:
USDA Forest Service, Northern Research Station
.
Texas Forest Service Publication 166:61–69. (see Supplemental Material, Reference S8, http://dx.doi.doi.org/10.3996/022013-JFWM-010.S22); also available: http://texasoakwilt.org/Professionals/NOWS/conference_assets/NOWS_Proceedings.pdf#page=67 (June 2014)
.
Dozier
J
.
1989
.
Spectral signature of alpine snow cover from the Landsat thematic mapper
.
Remote Sensing of Environment
28
:
9
22
.
Faber-Langendoen
D
,
Tester
JR
.
1993
.
Oak mortality in sand savannas following drought in east-central Minnesota
.
Bulletin of the Torrey Botanical Club
120
:
248
256
.
Franco-Lopez
H
,
Ek
AR
,
Bauer
ME
.
2001
.
Estimation and mapping of forest stand density, volume, and cover type using the k-nearest neighbors method
.
Remote Sensing of Environment
77
:
251
274
.
Franklin
J
.
1986
.
Thematic Mapper analysis of coniferous forest structure and composition
.
International Journal of Remote Sensing
7
:
1287
1301
.
Franklin
SE
,
McDermid
GJ
.
1993
.
Empirical relations between digital SPOT HRV and CASI imagery and lodgepole pine (Pinus contorta) forest stand parameters
.
International Journal of Remote Sensing
14
:
2331
2348
.
Geladi
P
,
Kowalski
BR
.
1986
.
Partial least-squares regression: a tutorial
.
Analytica Chimica Acta
185
:
1
17
.
Ginrich
SF
.
1967
.
Measuring and evaluating stocking and stand density in upland hardwood forests in the central States
.
Forest Science
13
:
38
53
.
Gong
G
.
1986
.
Cross-validation, the jackknife, and the bootstrap: excess error estimation in forward logistic regression
.
Journal of the American Statistical Association
81
:
108
113
.
Grosenbaugh
LR
.
1952
.
Plotless timber estimates—new, fast, easy
.
Journal of Forestry
50
:
322
337
.
Grundel
R
,
Pavlovic
NB
,
Sulzman
CL
.
1998
.
Habitat use by the endangered Karner blue butterfly in oak woodlands: the influence of canopy cover
.
Biological Conservation
85
:
47
53
.
Hansen
MJ
,
Franklin
SE
,
Woudsma
C
,
Peterson
M
.
2001
.
Forest structure classification in the north Columbia Mountains using the Landsat TM Tasseled Cap Wetness component
.
Canadian Journal of Remote Sensing
27
:
20
32
.
He
HS
,
Mladenoff
DJ
,
Radeloff
VC
,
Crow
TR
.
1998
.
Integration of GIS data and classified satellite imagery for regional forest assessment
.
Ecological Applications
8
:
1072
1083
.
Helland
IS
.
1988
.
On the structure of partial least squares regression
.
Communications in Statistics — Simulation and Computation
17
:
581
607
.
Horler
DNH
,
Ahern
FJ
.
1986
.
Forestry information content of thematic mapper data
.
International Journal of Remote Sensing
7
:
405
428
.
Huang
C
,
Goward
SN
,
Masek
JG
,
Thomas
N
,
Zhu
Z
,
Vogelmann
JE
.
2010
.
An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks
.
Remote Sensing of Environment
114
:
183
198
.
Huete
AR
.
1988
.
A soil adjusted vegetation index SAVI
.
Remote Sensing of the Environment
25
:
295
309
.
Jacobi
WR
,
MacDonald
WL
.
1980
.
Colonization of resistant and susceptible oaks by Ceratocystis fagacearum
.
Phytopathology
70
:
618
623
.
Jennings
SB
,
Brown
ND
,
Sheil
D
.
1999
.
Assessing forest canopies and understory illumination: canopy closure, canopy cover and other measures
.
Forestry
72
:
59
73
.
Jordan
CF
.
1969
.
Derivation of leaf area index from quality of light on the forest floor
.
Ecology
50
:
663
666
.
Juzwik
J
.
2000
.
An oak wilt primer
.
International Oaks
11
:
14
20
.
Kerr
JT
,
Southwood
TRE
,
Cihlar
J
.
2001
.
Remotely sensed habitat diversity predicts butterfly species richness and community similarity in Canada
.
Proceedings of the National Academy of Sciences
98
:
11365
11370
.
Kipfmuller
KF
,
Hepola
T
.
2007
.
Fire history and age structure analysis in the Sherburne National Wildlife Refuge: establishing reference conditions in a remnant oak savanna woodland
.
Fort Collins, CO
:
USDA Forest Service, Rocky Mountain Research Station
.
USDA Forest Service Proceedings RMRS-P-46CD, 507–514. (see Supplemental Material, Reference S3, http://dx.doi.doi.org/10.3996/022013-JFWM-010.S17); also available: http://www.fs.fed.us/rm/pubs/rmrs_p046/rmrs_p046_507_514.pdf (June 2014)
.
Ko
LJ
,
Reich
PB
.
1993
.
Oak tree effects on soil and herbaceous in savannas and pastures in Wisconsin
.
American Midland Naturalist
130
:
31
42
.
Koenig
WD
.
1999
.
Spatial autocorrelation of ecological phenomena
.
Trends in Ecology & Evolution
14
:
22
26
.
Law
JR
,
Johnson
PS
,
Houf
G
.
1994
.
A crown cover chart for oak savannas
.
Columbia, MO
:
USDA Forest Service, North Central Forest Experiment Station
.
Technical brief TB-NC-2. (see Supplemental Material, Reference S2, http://dx.doi.doi.org/10.3996/022013-JFWM-010.S16); also available: http://ncrs.fs.fed.us/pubs/tb/tb2/techbrf2.html (June 2014)
.
Leach
MK
,
Givnish
TJ
.
1996
.
Ecological determinants of species loss in remnant prairies
.
Science
273
:
1555
1558
.
Leach
MK
,
Givnish
TJ
.
1999
.
Gradients in the composition, structure, and diversity of remnant oak savannas in southern Wisconsin
.
Ecological Monographs
69
:
353
374
.
Lee
DW
,
Gould
KS
.
2002
.
Why leaves turn red
.
American Scientist
90
:
524
531
.
Liknes
GC
,
Perry
CH
,
Meneguzzo
DM
.
2010
.
Assessing tree cover in agricultural landscapes using high-resolution aerial imagery
.
Journal of Terrestrial Observation
2
:
5
.
Mäkelä
H
,
Pekkarinen
A
.
2004
.
Estimation of forest stand volumes by Landsat TM imagery and stand-level field-inventory data
.
Forest Ecology and Management
196
:
245
255
.
Margoles
SS
.
2009
.
Perspectives on oak savanna restoration in Minnesota: a dendroecological approach. Doctoral dissertation
.
Minneapolis
:
University of Minnesota
. .
Marschner
FJ
.
1974
.
The original vegetation of Minnesota, a map compiled in 1930 by F.J. Marschner under the direction of M.L. Heinselman of the US Forest Service
.
St. Paul
:
Cartography Laboratory of the Department of Geography, University of Minnesota
.
Available: http://collections.mnhs.org/cms/largerimage.php (ID Number: G4141 .D2 1930 .M3 c.7 8F)
.
Marshall
JD
,
Waring
RH
.
1986
.
Comparison of methods of estimating leaf-area index in old-growth Douglas-fir
.
Ecology
67
:
975
979
.
Martin
ME
,
Plourde
LC
,
Ollinger
SV
,
Smith
M-L
,
McNeil
BE
.
2008
.
A generalizeable method for remote sensing of canopy nitrogen across a wide range of forest ecosystems
.
Remote Sensing of Environment
112
:
3511
3519
.
McRoberts
RE
.
2008
.
A two-step nearest neighbors algorithm using satellite imagery for predicting forest structure within species composition classes
.
Remote Sensing of Environment
113
:
532
545
.
McRoberts
RE
,
Tomppo
EO
,
Finley
AO
,
Heikkinen
J
.
2007
.
Estimating areal means and variances of forest attributes using the k-Nearest Neighbors technique and satellite imagery
.
Remote Sensing of Environment
111
:
466
480
.
Menges
ES
,
Kuntz
JE
.
1985
.
Predictive equations for local spread of oak wilt in Southern Wisconsin
.
Forest Science
31
:
43
51
.
Miller
JR
,
Wu
J
,
Boyer
MG
,
Belanger
M
,
Hare
EW
.
1991
.
Seasonal patterns in leaf reflectance red-edge characteristics
.
International Journal of Remote Sensing
12
:
1509
1523
.
[MDNR] Minnesota Department of Natural Resources.
.
2005
.
Field guide to the native plant communities of Minnesota: the eastern broadleaf forest province
.
St. Paul
:
Minnesota Department of Natural Resources. Ecological Land Classification Program, Minnesota County Biological Survey, and Natural Heritage and Nongame Research Program
.
Myers
RH
.
1986
.
Classical and modern regression with applications
.
Boston
:
PWS-Kent Publishing
.
Niemi
GJ
,
Hanowski
JM
.
1984
.
Relationships of breeding birds to habitat characteristics in logged areas
.
Journal of Wildlife Management
48
:
438
443
.
Nuzzo
VA
.
1986
.
Extent and status of Midwest oak savanna: presettlement and 1985
.
Natural Areas Journal
6
:
6
36
.
Olechnowski
BFM
,
Debinski
DM
,
Drobney
P
,
Viste-Sparkman
K
,
Reed
WT
.
2009
.
Changes in vegetation structure through time in a restored tallgrass prairie ecosystem and implications for avian diversity and community composition
.
Ecological Restoration
27
:
449
457
.
Olyphant
GA
.
1984
.
Insolation topoclimates and potential ablation in alpine snow accumulation basins: Front Range, Colorado
.
Water Resources Research
20
:
491
498
.
Pacala
SW
,
Deutschman
DH
.
1995
.
Details that matter: the spatial distribution of individual trees maintains forest ecosystem function
.
Oikos
74
:
357
365
.
Packard
S
.
1988
.
Just a few oddball species: restoration and rediscovery of the tallgrass savanna
.
Restoration and Management Notes
6
:
13
24
.
Packard
S
.
1993
.
Restoring oak ecosystems
.
Restoration Management Notes
11
:
5
16
.
Pastor
J
,
Dewey
B
,
Moen
R
,
Mladenoff
DJ
,
White
M
,
Cohen
Y
.
1998
.
Spatial patterns in the moose–forest–soil ecosystem on Isle Royale, Michigan, USA
.
Ecological Applications
8
:
411
424
.
Peterson
DW
,
Reich
PB
.
2001
.
Prescribed fire in oak savanna: fire frequency effects on stand structure and dynamics
.
Ecological Applications
11
:
914
27
.
Peterson
DW
,
Reich
PB
,
Wrage
KJ
.
2007
.
Plant functional group responses to fire frequency and tree canopy cover gradients in oak savannas and woodlands
.
Journal of Vegetation Science
18
:
3
12
.
Phillips
OL
,
Malhi
Y
,
Higuchi
N
,
Laurance
WF
,
Núñez
PV
,
Vásquez
RM
,
Laurance
SG
,
Ferreira
LV
,
Stern
M
,
Brown
S
,
Grace
J
.
1998
.
Changes in the carbon balance of tropical forests: evidence from long-term plots
.
Science
282
:
439
442
.
Reese
HM
,
Lillesand
TM
,
Nagel
DE
,
Stewart
JS
,
Goldmann
RA
,
Simmons
TE
,
Chipman
JW
,
Tessar
PA
.
2002
.
Statewide land cover derived from multiseasonal Landsat TM data: a retrospective of the WISCLAND project
.
Remote Sensing of Environment
82
:
224
237
.
Riaño
D
,
Chuvieco
E
,
Salas
J
,
Aguado
I
.
2003
.
Assessment of different topographic corrections in Landsat-TM data for mapping vegetation types
.
IEEE Transactions on Geoscience and Remote Sensing
41
:
1056
1061
.
Rock
BN
,
Vogelmann
JE
,
Williams
DL
,
Vogelmann
AF
,
Hoshizaki
T
.
1986
.
Remote detection of forest damage
.
Bioscience
36
:
439
445
.
Schriever
JR
,
Congalton
RG
.
1995
.
Evaluating seasonal variability as an aid to cover-type mapping from Landsat Thematic Mapper data in the Northeast
.
Photogrammetric Engineering and Remote Sensing
61
:
321
327
.
Schulte
LA
,
Mottl
EC
,
Palik
BJ
.
2011
.
The association of two invasive shrubs, common buckthorn (Rhamnus cathartica) and tartarian honeysuckle (Lonicera tatarica), with oak communities in the midwestern United States
.
Canadian Journal of Forest Research
41
:
1981
1992
.
Schwaller
MR
,
Tkach
SJ
.
1985
.
Premature leaf senescence; remote-sensing detection and utility for geobotanical prospecting
.
Economic Geology
80
:
250
255
.
Sharov
AA
,
Pijanowski
BC
,
Liebhold
AM
,
Gage
SH
.
1999
.
What affects the rate of gypsy moth (Lepidoptera: Lymantriidae) spread: winter temperature or forest susceptibility
?
Agriculture and Forest Entomology
1
:
37
45
.
Singh
A
,
Jakubowski
AR
,
Chidister
I
,
Townsend
PA
.
2013
.
A MODIS approach to predicting stream water quality in Wisconsin
.
Remote Sensing of Environment
128
:
74
86
.
Sun
SJ
,
Chumlea
WC
,
Heymsfield
SB
,
Lukaski
HC
,
Schoeller
D
,
Friedl
K
,
Kuczmarski
RJ
,
Flegal
KM
,
Johnson
CL
,
Hubbard
VS
.
2003
.
Development of bioelectrical impedance analysis prediction equations for body composition with the use of a multicomponent model for use in epidemiologic surveys
.
American Journal of Clinical Nutrition
77
:
331
40
.
Swain
PC
.
1972
.
An analysis of the morphological differences among oaks in selected stands of the Quercus borealis-Q. ellipsoidalis complex. Master's thesis
.
Minneapolis
:
University of Minnesota
.
Tester
JR
.
1989
.
Effects of fire frequency on oak savannah in east-central Minnesota
.
Bulletin of the Torrey Botanical Club
121
:
266
276
.
Thome
KJ
,
Helder
DL
,
Aaron
D
,
Dewald
JD
.
2004
.
Landsat-5 TM and Landsat-7 ETM+ absolute radiometric calibration using the reflectance-based method
.
IEEE Transactions on Geoscience and Remote Sensing
42
:
2777
2785
.
Townsend
PA
,
Foster
JR
,
Chastain
RA
,
Currie
WS
.
2003
.
Application of imaging spectroscopy to mapping canopy nitrogen in the forests of the central Appalachian Mountains using Hyperion and AVIRIS
.
IEEE Transactions on Geoscience and Remote Sensing
41
:
1347
1354
.
[USFWS] U.S. Fish and Wildlife Service
.
2005
.
Comprehensive conservation plan: Sherburne National Wildlife Refuge
.
Washington, D.C.
:
U.S. Department of Interior
.
(see Supplemental Material, Reference 4, http://dx.doi.doi.org/10.3996/022013-JFWM-010.S18); also available: http://www.fws.gov/Midwest/Planning/sherburne/ccp/FinalCCP.pdf (December 2013)
.
[USFWS] U.S. Fish and Wildlife Service
.
2012
.
Working definition of “oak savanna” for restoration efforts at Sherburne NWR
.
Unpublished addendum to USFWS (2005) Comprehensive Conservation Plan (see Supplemental Material, Reference S7, http://dx.doi.doi.org/10.3996/022013-JFWM-010.S21)
.
Vogelmann
JE
,
Howard
SM
,
Yang
L
,
Larson
CR
,
Wylie
BK
,
Van Driel
N
.
2001
.
Completion of the 1990's National Land Cover Data Set for the conterminous United States
.
Photogrammetric Engineering and Remote Sensing
67
:
650
662
.
Vogelmann
JE
,
Rock
BN
.
1989
.
Use of thematic mapper data for the detection of forest damage caused by the pear thrips
.
Remote Sensing Environment
30
:
217
225
.
Vogl
RJ
.
1974
.
Effects of fire on grasslands
. Pages
139
194
in
Kozlowski
TT
,
Ahlgren
CE
,
editors
.
Fire and ecosystems
.
New York
:
Academic Press
.
White
AS
.
1983
.
The effects of thirteen years of annual prescribed burning on a Quercus ellipsoidalis community in Minnesota
.
Ecology
64
:
1081
1085
.
Whittaker
RH
,
Bormann
FH
,
Likens
GE
,
Siccama
TG
.
1974
.
The Hubbard Brook ecosystem study: forest biomass and production
.
Ecological Monographs
44
:
233
254
.
Wilhelm
G
,
Rericha
L
.
2007
.
Timberhill savanna assessment of landscape management
.
Elmhurst, IL
:
Conservation Research Institute
. .
Wold
S
,
Hoy
M
,
Martens
H
,
Trygg
J
,
Westad
F
,
MacGregor
J
,
Wise
BM
.
2009
.
The PLS model space revisited
.
Journal of Chemometrics
23
:
67
68
.
Wolter
PT
,
Berkley
EA
,
Peckham
SD
,
Singh
A
,
Townsend
PA
.
2012
.
Exploiting tree shadows on snow for estimating forest basal area using Landsat data
.
Remote Sensing of Environment
121
:
69
79
.
Wolter
PT
,
Mladenoff
DJ
,
Host
GE
,
Crow
TR
.
1995
.
Improved forest classification in the northern Lake State using multi-temporal Landsat imagery
.
Photogrammetric Engineering and Remote Sensing
61
:
1129
1143
.
Wolter
PT
,
Townsend
PA
.
2011
.
Estimating forest species composition using a multi-sensor fusion approach
.
Remote Sensing of Environment
115
:
671
691
.
Wolter
PT
,
Townsend
PA
,
Kingdon
CC
,
Sturtevant
BR
.
2008
.
Remote sensing of the distribution and abundance of host species for spruce budworm in northern Minnesota and Ontario
.
Remote Sensing of Environment
112
:
3971
3982
.
Wolter
PT
,
Townsend
PA
,
Sturtevant
BR
.
2009
.
Estimation of forest structural parameters using 5 and 10 meter SPOT-5 satellite data
.
Remote Sensing of Environment
113
:
2019
2036
.
Woodcock
CE
,
Collins
JB
,
Gopal
S
,
Jakabhazy
V
,
Li
X
,
Macomber
S
,
Ryherd
S
,
Wu
Y
,
Harward
VJ
,
Levitan
J
,
Warbington
R
.
1994
.
Mapping forest vegetation using Landsat TM imagery and a canopy reflectance model
.
Remote Sensing of Environment
50
:
240
254
.
Woods
M
,
Lim
K
,
Treitz
P
.
2008
.
Predicting forest stand variables from LiDAR data in the Great Lakes–St. Lawrence forest of Ontario
.
The Forestry Chronicle
84
:
827
839
.
Wovcha
DS
,
Delaney
BC
,
Nordquist
GE
.
1995
.
Minnesota's St. Croix River Valley and Anoka Sandplain: a guide to native habitats
.
Minneapolis
:
University of Minnesota Press
.
Xu
B
,
Peng
G
,
Pu
R
.
2003
.
Crown closure estimation of oak savannah in a dry season with Landsat TM imagery: a comparison of various indices through correlation analysis
.
International Journal of Remote Sensing
24
:
1811
1822
.
Ziegler
SS
,
Larson
ER
,
Rauchfuss
J
,
Elliott
GP
.
2008
.
Tree establishment during dry spells at an oak savanna in Minnesota
.
Tree-Ring Research
64
:
47
54
.

Author notes

Wolter PT, Berkley EA, Peckham SD, Singh A. 2014. Satellite-based management tool for oak savanna ecosystem restoration. Journal of Fish and Wildlife Management 5(2):252–269; e1944-687X. doi: 10.3996/022013-JFWM-010

The findings and conclusions in this article are those of the author(s) and do not necessarily represent the views of the U.S. Fish and Wildlife Service.