Quantifying tree mortality in a mixed species woodland using multitemporal high spatial resolution satellite imagery

Remote Sensing of Environment 129 (2013) 54–65

Contents lists available at SciVerse ScienceDirect

Remote Sensing of Environment journal homepage: www.elsevier.com/locate/rse

Quantifying tree mortality in a mixed species woodland using multitemporal high spatial resolution satellite imagery Steven R. Garrity a,⁎, Craig D. Allen b, Steven P. Brumby a, Chandana Gangodagamage c, Nate G. McDowell c, D. Michael Cai a a b c

International, Space & Response Division, Los Alamos National Laboratory, Los Alamos, NM 87545, Unites States U.S. Geological Survey, Jemez Mountains Field Station, Los Alamos, NM 87544, United States Earth & Environmental Sciences Division, Los Alamos National Laboratory, Los Alamos, NM 87545, United States

a r t i c l e

i n f o

Article history: Received 7 June 2012 Received in revised form 10 October 2012 Accepted 11 October 2012 Available online 23 November 2012 Keywords: Climate change Drought Image classification QuickBird Tree mortality Unsupervised data clustering WorldView-2

a b s t r a c t Widespread tree mortality events have recently been observed in several biomes. To effectively quantify the severity and extent of these events, tools that allow for rapid assessment at the landscape scale are required. Past studies using high spatial resolution satellite imagery have primarily focused on detecting green, red, and gray tree canopies during and shortly after tree damage or mortality has occurred. However, detecting trees in various stages of death is not always possible due to limited availability of archived satellite imagery. Here we assess the capability of high spatial resolution satellite imagery for tree mortality detection in a southwestern U.S. mixed species woodland using archived satellite images acquired prior to mortality and well after dead trees had dropped their leaves. We developed a multistep classification approach that uses: supervised masking of non-tree image elements; bi-temporal (pre- and post-mortality) differencing of normalized difference vegetation index (NDVI) and red:green ratio (RGI); and unsupervised multivariate clustering of pixels into live and dead tree classes using a Gaussian mixture model. Classification accuracies were improved in a final step by tuning the rules of pixel classification using the posterior probabilities of class membership obtained from the Gaussian mixture model. Classifications were produced for two images acquired post-mortality with overall accuracies of 97.9% and 98.5%, respectively. Classified images were combined with land cover data to characterize the spatiotemporal characteristics of tree mortality across areas with differences in tree species composition. We found that 38% of tree crown area was lost during the drought period between 2002 and 2006. The majority of tree mortality during this period was concentrated in piñon-juniper (Pinus edulis-Juniperus monosperma) woodlands. An additional 20% of the tree canopy died or was removed between 2006 and 2011, primarily in areas experiencing wildfire and management activity. -Our results demonstrate that unsupervised clustering of bi-temporal NDVI and RGI differences can be used to detect tree mortality resulting from numerous causes and in several forest cover types. © 2012 Elsevier Inc. All rights reserved.

1. Introduction Climate warming and recent severe droughts have resulted in vegetation mortality in various woody biomes across the globe (Allen et al., 2010; Breshears et al., 2005; Carnicer et al., 2011; Phillips et al., 2009; van Mantgem et al., 2009). Changes in ecosystem structure and function resulting from mortality have the potential to significantly alter biogeochemical cycles, energy fluxes, and landscape patterns of vegetation composition at local, regional, and possibly global scales (Adams et al., 2009; Amiro et al., 2010; Hanson & Weltzin, 2000; Huang et al., 2010a; Kane et al., 2011; van der Molen et al., 2011). Our ability to quantify the impacts of vegetation mortality

⁎ Corresponding author. Tel.: +1 505 606 0127. E-mail address: [email protected] (S.R. Garrity). 0034-4257/$ – see front matter © 2012 Elsevier Inc. All rights reserved. http://dx.doi.org/10.1016/j.rse.2012.10.029

on ecological, biogeochemical, and biosphere–atmosphere dynamics is currently limited due to a fundamental lack of information on where and when mortality events occur across the globe (Allen et al., 2010; McDowell et al., 2011). Even at local scales the extent and magnitude of shifts in vegetation composition and landscape structure resulting from mortality events is difficult to quantify and scale from limited, plot-based information. Remotely sensed imagery obtained from earth-observation satellites has potential to fill these knowledge gaps (Frolking et al., 2009). The utility of remotely sensed imagery for understanding vegetation mortality has been well documented in coniferous forests of the northern U.S. and Canadian Rockies where regional outbreaks of bark beetles are causing high rates of tree mortality. In these forests, remotely sensed imagery has been used to provide critical information on the extent and progression of bark beetle infestation (Coops et al., 2006; DeRose et al., 2011; Wulder et al., 2008), to predict forest

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

susceptibility and insect spread (Coggins et al., 2008; Wulder et al., 2006b), and as a tool for evaluating the performance of management strategies (Wulder et al., 2009). At the stand to landscape level, high spatial resolution (b3 m/pixel), single date (e.g., Dennison et al., 2010; Hicke & Logan, 2009; Meddens et al., 2011) and multitemporal (Wulder et al., 2008, 2009) color infrared imagery have been used for mapping beetle damage. Similarly, high resolution imagery has been applied in other forests to detect mortality from other insects and pathogens as well (e.g., Goodwin et al., 2005; Ismail et al., 2007; Kelly et al., 2004), although detection of bark beetle damage in North American forests has received the most research focus so far. Beetle- and other insect-caused mortality detection techniques often focus on identifying pixels containing green, red, or gray canopy foliage (Oumar & Mutanga, 2011). Detecting tree canopies at different stages of death requires that imagery be collected during the red attack phase when tree foliage turns red shortly after beetle attack (e.g. Hicke & Logan, 2009; White et al., 2005), or once trees turn gray following needle drop (e.g., Meddens et al., 2011), which typically occurs 2–3 years following infestation (Wulder et al., 2006a). Detecting red foliage, however, may not always be feasible depending on several factors. For example, tree foliage may not necessarily turn red, depending on mortality agent and tree species (e.g., Goodwin et al., 2005). Leaf drop may also occur quicker when the cause of mortality is abiotic (e.g., wind) or involves different insects (e.g., defoliators) and pathogens, limiting the time window for, or preventing altogether the detection of red trees. In cases of especially severe abiotic disturbance, such as hurricanes or harvest, whole trees may be removed rapidly making it impossible to detect changes in foliar coloration. It may also be the case that satellite imagery is contaminated by clouds and aerosols or is not acquired during periods of diagnostic foliar changes, limiting our ability to evaluate past occurrences of tree mortality in many areas with existing methods. Multitemporal image analysis may help overcome the challenges highlighted here, provided that imagery is available prior to the disturbance event so that premortality forest conditions can be quantified. The goal then would be to identify areas of change in subsequent images and determine what spectral change signals tree death. High resolution imagery has also been used to detect forest disturbances such as burn severity from wildfire (Holden et al., 2010) and the occurrence of standing dead wood or gaps that occur in continuous canopies when one or more trees die (Garbarino et al., 2012; Pasher & King, 2009). Detection of mortality not related to insects or pathogens, however, has only rarely been evaluated with high resolution imagery, and typically involves identification of unique spectral signatures associated with fire burn scars (Holden et al., 2010) or snags (Pasher & King, 2009), or the use of combined spectral and texture features to identify forest gaps (Rich et al., 2010). Forest gap formation is directly related to tree mortality and thus may be a useful indicator of tree death in forests with continuous canopy cover, but would not be suitable in more open forests and woodlands where tree canopy cover in discontinuous. Fire damage is also related to tree death, however, to our knowledge past studies have not used high resolution imagery to explicitly investigate fire caused tree mortality, but have rather focused on quantifying fire affected area. Knowledge of pre-burned forested area could be combined with detection of burned area post-fire to estimate the number of trees killed. Overall, new studies are required that evaluate the utility of high spatial resolution imagery for quantifying tree mortality in areas where the mortality occurs due to several different causes, affects diverse species, or happens during time periods when imagery is not available during periods of foliar fading associated with tree stress or insect and pathogen attack. The objective of our study was to evaluate the potential of high spatial resolution multispectral imagery for detailed mapping of tree mortality in a southwestern U.S. mixed species woodland. Regionalscale tree mortality was observed in southwestern U.S. woodlands

55

during a severe drought in the early 2000s (Breshears et al., 2005; Shaw et al., 2005; U S Forest Service, 2002, 2003, 2004, 2005). Coarse spatial resolution (≥30 m/pixel) imagery has been used to investigate the extent and ecological implications of this wide-scale die-off (e.g., Huang et al., 2010a; Rich et al., 2008; Vogelmann et al., 2009; Vogelmann et al., 2012) in targeted areas, however, detailed quantification of tree mortality and the landscape response following tree mortality is limited to a sparse number of intensively studied groundbased plots spread throughout the region (e.g., Breshears et al., 2005, Floyd et al., 2009; Shaw et al., 2005). In addition to drought, the study region has also experienced related insect outbreaks and wildfire (Allen, 2007). Here, we used multitemporal, archived satellite imagery acquired before and well after tree mortality occurred during drought and associated beetle outbreak, wildfire, and mechanical tree removal to develop a multistep classification approach for quantifying tree mortality across the landscape. Detailed tree mortality maps can be used for understanding landscape changes, for example, dynamics in stand structure and community composition, resulting from tree mortality. They may also be useful for tuning or evaluating analyses conducted with coarser resolution imagery or for evaluating the performance of hindcast ecosystem model simulations. 2. Methods 2.1. Study area Our study area is within the Pajarito Plateau region located on the eastern slope of the Jemez Mountains in northern New Mexico, USA (Fig. 1). We focused our study on a 46 km 2 area in Bandelier National Monument (Fig. 1). We selected this region because of the long history of well-documented disturbances (e.g., Allen, 1989, 2007; Allen & Breshears, 1998; Breshears et al., 2005; McDowell et al., 2010) and the relatively high variability in topography, elevation, and species composition. Average elevation of the Pajarito Plateau is approximately 2140 m. Topography of the Pajarito Plateau is characterized by a series of mesas divided by canyons. Piñon pine (Pinus edulis) and one-seed juniper (Juniperus monosperma) are the dominant overstory species on mesa tops in the mid- to lower-elevations that dominate our study area (Fig. 1). At higher elevations, mesa vegetation becomes more mesic, reflected by a change in vegetation composition with increasing frequency of ponderosa pine (Pinus ponderosa), mountain mahogany (Cercocarpus montanus), and Gambel oak (Quercus gambelii) interspersed with grasslands. Within the canyons, where water is more abundant, the overstory is primarily composed of a mix of ponderosa pine, Gambel oak, some Douglas-fir (Pseudotsuga menziesii) and white fir (Abies concolor), and narrowleaf cottonwood (Populus angustifolia). The Pajarito Plateau experiences bimodal annual precipitation with peaks as winter snowfall and summer monsoon rainfall. Average precipitation is approximately 400 mm per year. Mean annual air temperature is 9 °C with daily means of − 2 °C in January and 21 °C in June (Breshears et al., 2008). Anomalously high air temperatures and low precipitation were observed at a meteorological station located on the Pajarito Plateau in the early 2000s (Fig. 2a). Between 2002 and 2004, plot-level measurements within the Pajarito Plateau indicated that 97% of mature piñon died due to drought and associated piñon ips beetle (Ips confuses) attack, whereas a great majority of the juniper survived, with dispersed or patchy mortality in some areas (Allen, 1989, 2007; Breshears et al., 2005, 2008; McDowell et al., 2008). Time series of Advanced Very High Resolution Radiometer normalized difference vegetation index (NDVI), obtained from a 4.8 km 2 area of the Pajarito Plateau between 1990 and 2006, showed that NDVI declined and exhibited reduced annual variability in the years leading up to the mortality event (Fig. 2b). The lowest summer period NDVI was observed in 2002 coinciding with the lowest precipitation year, with almost no seasonality present during this time. In the years after 2002,

56

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

Fig. 1. Study area location in northern New Mexico, USA (state shaded yellow in inset). A Landsat image is shown (displayed as a combination of bands 5, 4, and 3) with Bandelier National Monument delineated by a white polygon and extent of high resolution imagery delineated by a black rectangle. For the area covered by the high resolution images, dominant vegetation cover types are ponderosa pine (green), piñon-juniper (yellow), and juniper (cyan). Data used to create the vegetation cover map were provided by the USGS and the New Mexico Natural Heritage Program and can be accessed at http://biology.usgs.gov/npsveg/band/index.html.

there was increased annual precipitation and reduced maximum air temperatures. Simultaneously, summer time NDVI gradually increased and experienced some recovery of the seasonal pattern observed in the 1990s. Plot-level measurements suggest that post-drought seasonal patterns of NDVI for the Pajarito Plateau have primarily been driven by recovery of herbaceous vegetation in response to increased soil moisture availability (Rich et al., 2008). Regionally, anomalies in NDVI time series during the early 2000s corresponded well with regional drought and forest drought stress (Williams et al., 2012), suggesting that the patterns observed at our study location were experienced across a broader geographical region. In addition to drought, there was a 50 ha wildfire in 2006 and ongoing ecological restoration treatments were conducted between 2006 and 2011 within the study area. The ecological restoration treatments were designed to mitigate accelerated soil erosion and protect embedded prehistoric cultural resources within piñon-juniper woodlands degraded by historic land use. Treatment consisted of a mechanical overstory thinning, targeting live juniper and dead piñon

b15–20 cm stem diameter, and using cut material as a coarse surface mulch on exposed inter-canopy soils to moderate runoff and improve site conditions for understory growth. 2.2. Satellite imagery acquisition and preprocessing High spatial resolution digital imagery was acquired from two satellite sensors: QuickBird and WorldView-2. QuickBird images were acquired prior to tree mortality on June 3, 2002 and again on March 26, 2006. In 2011, a WorldView-2 image was acquired on May 3 (Table 1). Each image covered an area of approximately 46 km 2. The satellite images were acquired following snowmelt and prior to the monsoon season, thus minimizing inter-image variability in the greenness of non-evergreen vegetation and facilitating multitemporal analysis of evergreen trees. Imagery was ordered as ortho-ready multispectral and panchromatic pairs with spectral information coded in units of digital numbers (DN). Multispectral images were pan sharpened by fusing the coarse

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

57

500

(a)

Tmax

Precip

18

400

17

300

16

200

15 1990

1992

1994

1996

1998

2000

2002

2004

2006

2000

2002

2004

2006

Annual precipitation (mm)

Mean annual Tmax (C)

19

100

Year 0.6

(b) AVHRR NDVI

0.5 0.4 0.3 0.2 0.1 1990

1992

1994

1996

1998

Year Fig. 2. (a) Mean annual precipitation and maximum air temperature for the Pajarito Plateau from 1990 to 2006. (b) Cubic-spline smoothed Advanced Very High Resolution Radiometer NDVI time series for a 4.8 km2 area of the Pajarito Plateau for the same observation period.

spatial resolution data with the high spatial resolution panchromatic band using the Gram-Schmidt (Laben & Brower, 2000) processes. Pan sharpened, 0.6 m spatial resolution multispectral images were then orthorectified by combining a 10 m USGS digital elevation model (DEM) with the rational polynomial coefficients (RPCs) supplied for each image. To facilitate multitemporal analysis, images acquired in 2006 and 2011 were co-registered to the 2002 QuickBird image using between 48 and 56 manually selected tie-points. Co-registration root mean squared error was b1.25 m for both images. Image-to-image radiometric normalization was performed using the iteratively re-weighted Multivariate Alteration Detection (IR-MAD) transformation (Canty & Nielsen, 2008). The IR-MAD algorithm identifies bi-temporal invariant pixels through iterative canonical correlation. Invariant pixels are used to fit a linear model for each spectral band. Fitted slopes and intercepts for each band are then applied to all image pixels. Normalization was performed on the DN values for the 2002–2006 and 2002–2011 image pairs independently. For the 2011 WorldView-2 image, DN values from bands 2, 3, 5, and 7 were used as the blue, green, red and near-infrared (NIR) bands, respectively, because they most closely matched the spectral characteristics of the blue, green, red and NIR bands from QuickBird. Although the spectral information contained in most high spatial resolution satellite imagery is relatively limited (i.e., four band colorinfrared), there are multiple vegetation indices that can be calculated using these bands that may enhance mortality detection. For example, several studies (Coops et al., 2006; Meddens et al., 2011; Wulder et al., 2008) have found that the NDVI and the red:green ratio (Red-Green

Index (RGI)) vegetation indices are useful for differentiating between live green trees and trees that have turned red or gray following mountain pine beetle attack. To increase the number of variables available for detecting tree mortality in our study area we used normalized image DN values to calculate NDVI and RGI for each image. 2.3. Image classification 2.3.1. Masking non-tree image elements Because the Pajarito Plateau landscape is comprised of several non-tree elements we chose to isolate image pixels containing trees from those containing other elements. The objective was to track the fate of pixels containing live tree foliage in 2002 through 2006 and in 2011. To do so, we first used a detailed vegetation map produced by the USGS and the New Mexico Natural Heritage Program for Bandelier National Monument (http://biology.usgs.gov/npsveg/ band/index.html). Pixels outside regions classified as white fir, Douglas-fir, ponderosa pine, piñon-juniper, or juniper were masked (Fig. 1c) in the 2002 image. Second, we used a supervised classification to mask out all non-tree elements within regions not already masked out in the 2002 image using the USGS vegetation map. Nontree elements in the 2002 image included herbaceous vegetation, exposed rock and soil, roads, clouds, and cloud shadows. The supervised classification was performed using Genetic Imagery Exploitation (GENIE; Observera, Inc., Chantilly, VA, USA) a machine learning software package that uses an evolutionary algorithm (i.e., continually evolving solution that seeks to optimize the resulting classification

Table 1 Satellite image acquisition parameters. Satellite sensor

Acquisition date

UTC time

Local time (MDT)

Sun azimuth

Sun elevation

Satellite azimuth

Satellite elevation

In track view angle

Cross track view angle

Off-nadir view angle

QuickBird QuickBird WorldView-2

2002–06–03 2006–03–26 2011–05–03

17:55:14 18:09:28 18:22:04

10:55:14 11:09:28 11:22:04

127.2° 153.4° 153.2°

70.1° 53.9° 68.2°

179.0° 65.4° 222.9°

78.7° 78.9° 68.4°

10.5° 5.8° −16.0

−1.8 8.6° −10.3

10.9° 10.5° 19.0°

58

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

Table 2 Confusion matrix of classification accuracy resulting from supervised classification of tree and non-tree pixels for the 2002 image. Classification Total

Commission error (%)

User's accuracy (%)

10869 224 11093 2.0 98.0

335 16417 16752 2.0 98.0

11204 16641 27845

3.0 1.3

97.0 98.7

(Holland, 1975)) to search through a space of pixel-wise spectral and patch-wise spatial textural image processing operations to generate a set of local image features that support a human-specified image classification task (Brumby et al., 1999; Perkins et al., 2001). This approach mimics and automates human exploration of pixel-wise spectral band combinations, like NDVI or RGI, and patch-wise spatial statistics and texture measures, like wavelet edge detection or local amplitude variance. GENIE uses a supervised linear discriminant classifier to combine these spectral and textural features into a pixel-level classification map over the input image. GENIE has previously been applied successfully to land-cover classification using satellite and aerial multi-spectral imagery (Harvey et al., 2002), including tree mortality due to catastrophic wildfires (Brumby et al., 2002). In the current study, we selected several thousand training pixels corresponding to tree and non-tree image elements. Results from the GENIE classification and the USGS vegetation map masking were combined so that only pixels classified as containing live tree foliage remained. The non-tree masks, including the cloud and cloud shadow regions, were overlaid on the 2006 and 2011 image for subsequent analyses.

ΔVItn ¼ VIt1 −VI tn

ð1Þ

2.3.3. Classification accuracy assessment Several thousand pixels containing live tree foliage and pixels containing non-tree elements were selected from the 2002 image using the GENIE graphical user interface. Marked up pixels were used for supervised classification training and to assess the accuracy of tree and non-tree classification. To assess mortality detection accuracy, pixels were extracted from a total of 300 manually delineated

700

1.00

(a) 600

0.86

500

0.71

400

0.57

300

0.43

200

0.29

100

0.14

0

0.00 Blue Green

Red

NIR

NDVI

RGI

700 600

1.00

(b)

0.86

500

0.71

400

0.57

300

0.43

200

0.29

100

0.14

0

Vegetation index values

where ΔVItn is the difference in the vegetation index from the first image (t1) to subsequent images (tn), so that if, for example, NDVIt1 is greater than NDVItn, ΔNDVItn will be positive. For each bi-temporal image pair, 20% of ΔNDVI and ΔRGI pixel values were randomly sampled and used to fit a Gaussian mixture model (GMM) using an iterative expectation maximization learning algorithm (Dempster et al., 1977). The objective of GMM modeling is to cluster observations according to n-dimensional multivariate Gaussian distributions, where n is the number of input variables. To determine the actual number of clusters contained in the data, we used a sweep of the cluster parameter, with values ranging from one to six. Model fitting for each sweep was performed iteratively until reaching convergence. Several replicates of the iterative fitting were performed for each sweep and the best model, determined through comparison of the Akaike Information Criterion (AIC) (Akaike, 1974), was retained. Once the optimal number of clusters was identified, again by comparing AIC among models, the fitted GMM was applied to the entire image.

tree crowns (150 live and 150 dead crowns) in each post-mortality image (2006 and 2011). Tree crowns were independently selected from the 2006 and 2011 image, thus providing a total of 600 tree crowns (300 live, 300 dead). False-color display (NIR, red, and green bands) of pan-sharpened images (0.6 m spatial resolution) facilitated visual selection of live and dead tree crowns and clusters in the imagery. Live and dead tree crown delineation was performed by a single person familiar with the study site to maintain consistency in crown selection. Manual selection of live and dead tree crowns was performed for two reasons. First, live and dead trees were visually distinguishable because of high spatial resolution and large differences in the NIR band. Second, ground-based data on tree mortality within the study area were limited to a single 1.0 ha plot. Data from the groundbased plot included a census of live and dead trees in 2011 which were used as an independent source of validation. Proportions of live and dead trees in 2006 and 2011 were calculated from the plot data and compared with co-located, area-based calculations of percent live and dead tree area in the 2006 and 2011 classification maps. However, plot data represented a limited number of trees over a very small spatial extent so it was necessary to obtain a larger sample for meaningful classification assessment. Furthermore, since we were evaluating tree mortality occurring several years in the past, it was not possible to

Vegetation index values

2.3.2. Unsupervised mortality detection Bi-temporal differencing of the NDVI (ΔNDVI) and RGI (ΔRGI) vegetation indices was used to characterize the temporal change in those pixels identified as containing live tree foliage in the 2002 image. For both NDVI and RGI, the bi-temporal differencing was calculated as

Overall accuracy = 98.0% Kappa = 0.958

DN values

Tree Non-tree Total Omission error (%) Producer's accuracy (%)

Non-tree

DN values

Reference

Tree

0.00 Blue Green

Red

NIR

NDVI

RGI

Fig. 3. Mean digital number (DN) values, NDVI, and RGI for live (open circles) and dead (closed circles) trees in (a) 2006 and (b) 2011. Error bars represent standard deviations.

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65 5

Akaike Information Criterion

−2.4

3. Results

x 10

2006 2011

−2.6

3.1. Single-date tree crown identification Supervised classification of tree and non-tree pixels in the 2002 image resulted in 98% overall accuracy when compared against our validation dataset (Table 2). Based on this classification, we estimated a tree crown area of 1074 ha immediately prior to the drought-induced mortality event. Tree pixels represented 23% of the entire image area; however, this estimate does not represent the actual tree area within the study region because portions of the image were masked due to cloud and cloud shadow.

−2.8 −3 −3.2 −3.4 −3.6 −3.8

59

3.2. Unsupervised multitemporal tree mortality detection 1

2

3

4

5

6

Number of clusters Fig. 4. Akaike Information Criterion (AIC) values from each Gaussian mixture model (GMM) fitted with a sweep of cluster numbers.

collect additional ground-based data. To assess classification accuracy, confusion matrices and kappa statistics were calculated by comparing the GMM-produced classifications with co-located pixels extracted from live and dead tree crown polygons.

Comparison of DN values and vegetation indices between live and dead tree pixels (Fig. 3a,b) demonstrated that dead trees tend to have higher reflectance in the visible bands, lower reflectance in the NIR band, lower NDVI values, and higher RGI values than live trees. These results were consistent across the imagery acquired in 2006 (Fig. 3a) and 2011 (Fig. 3b), although the magnitudes in the differences were noticeably larger in the visible bands and smaller in the NIR band in 2011 compared to 2006. Differences for the RGI and NDVI vegetation indices were the most consistent across images and were therefore used as input variables for the unsupervised GMM classification.

0.3

(a)

Posterior probability

0.2

0

liveB

Δ RGI

0.1

0.4

0.6

0.8

0.4

liveA

0

0.2

deadA

(a) 0.2

−0.1

Δ RGI

deadB −0.2

−0.2

−0.1

0

0.1

0.2

0.3

0

−0.2

Δ NDVI 0.3

(b) 0.2

−0.4 −0.4

Δ RGI

0

0.2

0.4

0.2

0.4

Δ NDVI 0.4

0.1

0

−0.2

liveB

liveA

(b)

deadA 0.2

Δ RGI

−0.1

dead

B

−0.2

−0.2

−0.1

0

0.1

0.2

0.3

0

−0.2

Δ NDVI Fig. 5. Clusters mean ΔRGI and ΔNDVI generated with a four cluster GMM. Observation densities are represented by contours where darker colors indicate high density and lighter colors indicate low density. GMM models were independently generated for the bi-temporal image pairs in (a) 2002–2006 and (b) 2002–2011. In both cases, observations belonging to clusters within the upper left quadrant were classified as live and those belonging to clusters in the lower right quadrant were classified as dead.

−0.4 −0.4

−0.2

0

Δ NDVI Fig. 6. Posterior probabilities of the liveA (a) and deadA (b) clusters from the GMM fitted to the 2002–2006 image pair.

60

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

1

(a) True live rate

0.98

0.96

0.94

0.92

0.9

0

0.02

0.04

0.06

0.08

0.1

0.08

0.1

False live rate 1

(b) True dead rate

0.98

0.96

0.94

0.92

0.9

0

0.02

0.04

0.06

False dead rate Fig. 7. Receiver operating characteristic (ROC) curves demonstrating effect of adjusting the posterior probability cutoff of belong to the liveA cluster in the 2002–2006 classification. Closed black circles represent iterative tuning adjustments, gray circles represent classification performance prior to tuning and the open circles represent classification performance after tuning.

A sweep of the cluster parameter in the GMM demonstrated that four clusters produced a minimum AIC for the 2006 image (Fig. 4). AIC results from models fitted to ΔNDVI and ΔRGI values from 2011 did not display an obvious minimum even though a noticeable break in the AIC–cluster slope was observed at four clusters. Although comparison of AIC results among fitted models did not provide a definitive answer as to how many clusters should be used, visual interpretation of the AIC–cluster curves suggested that four clusters were appropriate. Using four clusters, GMMs were fitted to the ΔNDVI and ΔRGI values independently for each bi-temporal image pair: 2002–2006

and 2002–2011. In each case cluster mean ΔNDVI–ΔRGI pairs showed a negative relationship where clusters with positive ΔRGI had negative ΔNDVI, and clusters with negative ΔRGI had positive ΔNDVI (Fig. 5a, b). Plotting cluster mean ΔNDVI and ΔRGI pairs revealed a logical clustering of the data, given the NDVI and RGI pixel values observed for live and dead tree pixels (Fig. 3). The density of observations was generally centered on ΔNDVI and ΔRGI values around zero. Initially, pixels belonging to clusters within the upper left quadrant of Fig. 5a, b were assigned to the live tree class (liveA and liveB) and pixels belonging to clusters in the lower right quadrant were assigned to the dead tree class (deadA and deadB). However, initial tests showed that this resulted in a high ‘false dead rate’, where pixels were incorrectly assigned to the dead tree class. Fig. 6 demonstrates that the posterior probabilities of belonging to the liveA (Fig. 6a) or deadA (Fig. 6b) class are approximately equal around ΔNDVI and ΔRGI values of zero (hereafter referred to as the ‘no change’ region). A reduced false dead rate was achieved through reducing the posterior probability cutoff point of belonging to the liveA class in the no change region where the liveA and deadA distributions overlapped. This effectively shifted the liveA distribution toward the lower right quadrant of Fig. 5a and b so that pixels around the no change region were assigned to the live class. The posterior probability cutoff point was tuned by incrementally shifting the posterior probability cutoff of the liveA class to increasingly lower values. Receiver operating characteristic (ROC) analysis was used to quantify the effect of these shifts on live and dead tree detection rates (Fig. 7a, b). The new, optimal cutoff was defined as the posterior probability cutoff that resulted in a true dead rate-false dead rate distance in the ROC curve space closest to the theoretical optimal classifier (i.e., true dead rate = 1 and false dead rate = 0) (see white circle in Fig. 7b). The cutoff tuning decreased the false dead rate from 7.1% to 1.8% in 2006 and from 0.5% to 0.2% in 2011. This also had the positive effect of increasing the true live rate from 92.8% to 98.2% in 2006 and from 99.4% to 99.8% in 2011. The tradeoff was a slight increase in the false live rate and decrease in the true dead rate, however these changes were small relative to the improvement in the false dead rate (Fig. 7a, b). Following tuning, the liveA and liveB clusters and deadA and deadB clusters were combined into two classes: live and dead, respectively. We found that 3% of pixels classified as dead in 2006 were classified as live in 2011. Areas where this occurred were concentrated along canyon walls and in riparian areas where deciduous trees, shrubs and herbaceous vegetation are common. Therefore, we removed any pixel classified as dead in 2006 and live in 2011 from subsequent analyses. 3.3. Classification accuracy Very high classification accuracies were observed once the cutoff between live and dead classes around the no change region was tuned.

Table 3 Confusion matrix accuracy assessment for live and dead tree pixels in images acquired in 2006 and 2011. 2006 Classification

Reference

Live Dead Total Omission error (%) Producer's accuracy (%)

Live

Dead

Total

Commission error (%)

User's accuracy (%)

8500 210 8710 2.4 97.6

160 8487 8647 1.9 98.1

8660 8697 17357

1.8 2.4

98.2 97.6

Live

Dead

Total

Commission error (%)

User's accuracy (%)

5714 215 5929 3.6 96.4

11 9183 9194 0.1 99.9

5725 9398 15123

0.2 2.3

99.8 97.7

Overall accuracy = 97.9% Kappa = 0.957

2011 Classification

Reference

Live Dead Total Omission error (%) Producer's accuracy (%)

Overall accuracy = 98.5% Kappa = 0.969

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

Overall accuracies of 97.9% (kappa = 0.96) and 98.5% (kappa= 0.97) were observed for the 2006 and 2011 classifications, respectively (Table 3). In each case, commission and omission errors were less than 4%. Ground-based assessment of 413 trees with diameter at breast height greater than 10 cm were collected from a 1.0 ha plot within the piñon-juniper woodland. These measurements showed that 47% of the tree canopy area (253 trees, all piñon) had died by 2011. For the same

61

plot area, the image classifications resulted in area-based mortality estimates of 45% in 2006 and 54% in 2011. 3.4. Spatiotemporal mortality characteristics The tree mortality classifications (Fig. 8a-i, a-ii) were used to assess the spatiotemporal characteristics of tree death within Bandelier

Fig. 8. (a) Mortality classifications for 2006 (i) and 2011 (ii). Black pixels were unclassified because of cloud, cloud shadow, or non-tree vegetation, green pixels contain live tree crown, and red pixels contain dead tree crown. For the 2011 classification, blue pixels represent tree crown that died before 2006 and red pixels represent tree crown that died between 2006 and 2011. In panel a-ii a wildfire occurring in May 2006 is outlined in yellow and restoration areas are outlined in white. (b) False color (NIR, red, green) satellite image subsets of the area denoted by an X in a-ii acquired in 2002 (i), 2006 (ii), and 2011 (iii). Panels iv–vi are the classifications associated with the image subsets in i-iii, where the color scheme is the same as (a). (c) Ground-based images of the Bandelier study area acquired in 2004.

62

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

Dead crown area (ha)

400

occurred, and allowed us to infer rates of mortality in different cover classes.

(a)

300

4.1. Image analysis method 200

100

0 Fir

Ponderosa

P−J

Juniper

Vegetation cover class 1

Fraction dead pixels

(b)

2006 2011

0.8 0.6 0.4 0.2 0

Fir

Ponderosa

P−J

Juniper

Vegetation cover class Fig. 9. (a) Dead crown area by vegetation cover type. (b) Fraction of pixels classified as dead for each cover type.

National Monument. The greatest amount of mortality was observed between 2002 and 2006, with a loss of approximately 407 ha, or 38%, of the total tree crown area. An additional crown area loss of 215 ha (20%) was observed between 2006 and 2011. The majority of crown area lost between 2002 and 2006 was in the piñon-juniper woodlands (248 ha) followed by ponderosa (95 ha), juniper (45 ha) and white fir/Douglas-fir (0.5 ha) (Fig. 9a). In 2011, additional mortality was observed in each of the vegetation classes: piñon-juniper (130 ha), ponderosa (35 ha), juniper (45 ha), and white fir/Douglas-fir (0.2 ha) (Fig. 9a). Across both classification years (2006 and 2011) the highest percent tree crown area experiencing mortality was observed in piñon-juniper regions and the lowest percent tree crown area death occurred in the white fir/Douglas-fir regions (Fig. 9b). A substantial portion of the tree mortality occurring between 2006 and 2011 was in areas receiving restoration treatments conducted between 2007 and 2010 and wildfire in the summer of 2006. Within burned areas, our classification results showed that 5 ha of tree crown area died. Within areas receiving restoration treatments 125 ha of crown area was removed or died. Mortality occurring outside of wildfire or restoration areas accounted for the remaining 85 ha of crown area lost between 2006 and 2011.

4. Discussion Our objective was to evaluate the potential of high spatial resolution multispectral imagery for detailed mapping of tree mortality in a southwestern US mixed species woodland experiencing drought, bark beetle outbreak, fire, and ecological restoration treatments. Using a time series of NDVI and RGI values from pixels classified as tree in the pre-mortality image, we demonstrated that an unsupervised GMM was able to identify areas where trees or tree canopy had died or had been removed with a high level of accuracy. Classification accuracies were improved even further by automated adjustment of posterior probabilities of class membership. We showed that, when combined with an existing land cover classification, the tree mortality maps were useful for quantifying landscape-level information about where and when mortality

There were several elements of our approach that lead to high levels of classification performance. The first element was masking of non-tree image components. This was essential because it restricted all subsequent analysis to the dynamics of evergreen tree pixels. This is especially critical in multitemporal studies where confounding factors, such as variation in herbaceous vegetation greenness and soil brightness may reduce classification accuracy. Future multitemporal studies focused on evergreen trees should consider isolation of tree pixels as an initial step to increase discrimination between important and irrelevant changes, although this may be difficult in forests with low contrast between trees and surrounding vegetation. The second key factor was that we selected a reduced number of model input variables, namely ΔNDVI and ΔRGI values, as opposed to using individual spectral bands. Between the two vegetation indices, the green, red, and NIR bands were used, meaning that only information in the blue band was discarded. Reducing the inputs to two spectral dimensions facilitated straightforward interpretation of classification model output, tuning, and performance. We chose to use ΔNDVI and ΔRGI values for several reasons. First, multitemporal ΔRGI has been shown to be a good indicator of tree mortality in other forest systems (Wulder et al., 2008). Second, the differences between live and dead tree NDVI and RGI were more consistent than differences in any of the visible or NIR bands. In 2006, live and dead tree pixels exhibited very small differences in the visible bands and a large difference in the NIR band (see Fig. 3a). The opposite was observed in the 2011 image, where spectral differences between live and dead pixels were large in the visible bands and small in the NIR band (see Fig. 3b). The difference in live and dead tree spectral signatures between 2006 and 2011 was likely due to a number of factors, including differences in sun-surface-sensor geometry among images (see Table 1) and differences in spectral characteristics between QuickBird and WorldView-2 sensors. Furthermore, the optical properties of dead trees can be expected to change over the years due to degradation of structure and surface properties that occur during the decomposition process. Band ratioing involved in the calculation of NDVI and RGI effectively removed the influence of these factors. Another reason we chose to use NDVI and RGI is because their behavior can be reasonably well predicted from a theoretical standpoint. For example, leaf chlorophyll content is commonly observed to decline in dying or dead leaves (Eitel et al., 2011; Hilker et al., 2009). Because chlorophyll most strongly absorbs photons in red wavelengths, dying or dead leaves should have higher RGI and lower NDVI values relative to unstressed or live leaf material. Furthermore, stressed trees can experience partial or complete defoliation (Dobbertin & Brang, 2001) and dead trees eventually lose all of their leaves, resulting in lower leaf area and, therefore, low NDVI values. Taking these factors into consideration we expected RGI values to increase and NDVI values to decrease through time as a given tree (or pixel containing tree canopy) progresses from live to stressed to dead, as was observed in the current study. Another important element of our approach was the use of the unsupervised GMM clustering algorithm. This is not to imply that supervised classification approaches are necessarily inferior. On the contrary, several studies, especially those focused on detection of various stages of mountain pine beetle attack in forests of the northern U.S. and Canadian Rockies have used supervised classification with good success (e.g., Dennison et al., 2010; Hicke & Logan, 2009; Meddens et al., 2011). One advantage of the unsupervised GMM approach, however, is that it does not require training data. Instead the GMM expectation maximization algorithm is data-driven. In other words the underlying multivariate distributions of ΔNDVI and ΔRGI values do

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

not need to be known a priori, but are instead learned by the expectation maximization algorithm. An advantage of the GMM in particular is that it is a soft clustering method, where pixel class assignment is determined by posterior probability of cluster membership. As we show in the current study, posterior probabilities associated with each pixel can be used to tune final class membership and optimize classification results. The high level of automation in GMM clustering and class membership tuning represents an advance in efforts to detect tree mortality with high spatial resolution imagery across broad areas using large amounts of remotely sensed imagery. Although not tested in the current study, we hypothesize that GMMs can be fit to a number of different classes, such as red canopy, gray canopy, standing dead wood, and burned vegetation. Once developed and validated, the model may be transferred to other sets of imagery to find features of interest. For example, in bark beetle studies, once a model is developed for identifying green, red, and gray canopies, it may be applied to subsequent imagery or imagery from other areas where groundbased observations are lacking to detect whether or not beetle damage has occurred. Such an approach may be useful for automated forest monitoring systems that use aerial and satellite imagery (e.g., Wulder et al., 2008, 2012), either as a standalone technique or in support of existing systems that use coarser spatial resolution imagery to cover broader areas (e.g., Huang et al., 2010b; Kennedy et al., 2010; Mildrexler et al., 2009). 4.2. Ecological application The image classifications showed that substantial tree mortality occurred between 2002 and 2006. Piñon-juniper woodlands experienced the highest total loss of crown area relative to other vegetation cover types (see Fig. 9). Evidence from ground-based plots in Bandelier National Monument and elsewhere on the Pajarito Plateau suggest that the majority of local mortality occurred in piñon trees due to drought-induced physiological stress and an associated piñon ips beetle outbreak during 2002 and 2003 (Allen, 2007, and unpublished data; Breshears et al., 2008; McDowell et al., 2008), which has been validated by further experimental drought manipulations (Plaut et al., 2012). Our image classifications also show mortality in areas dominated by ponderosa pine and juniper. Ground-based measurements have documented significant ponderosa and juniper mortality in the study area (McDowell et al., 2010; C.D. Allen, unpublished data) and elsewhere in the region (e.g., Bowker et al., 2012; Floyd et al., 2009; Ganey & Vojta, 2012; Williams et al., 2012). The accuracy of mortality classification in white fir/Douglas-fir areas is less certain because they only occur in very isolated areas of the Pajarito Plateau, typically in areas with abundant soil water availability. In our study area, white fir/Douglas-fir occupied a very small area (~ 2 ha) in the northwestern portion of our study area (see Fig. 1c) and occurred in a canyon where Gambel oak and herbaceous vegetation is common. Therefore, our mortality detection in white fir/Douglas-fir areas could also have been due to phenological variability or death of shrubs and non-woody herbaceous vegetation, all of which may have been erroneously classified as tree mortality. Much of the post-drought mortality during the 2006–2011 period was due to extensive ecological restoration projects conducted by Bandelier National Monument land managers, as well as a wildfire that occurred in summer 2006 (see Fig. 8a-ii). Mortality during the 2006 to 2011 period outside of restoration and burned areas may have been due to a number of methodological and ecological factors. For example, differences in sun-surface-sensor geometry (see Table 1) and image-to-image georegistration errors could have led to classification errors (Wulder et al., 2008). Geometry-based errors should have the greatest impact in areas with tall statured trees, such as ponderosa pine and white fir/Douglas-fir, however they should be less pronounced in areas with short-statured trees such as

63

piñon and juniper. From an ecological perspective, mortality not caused by fire or restoration activity during the 2006–2011 time window may have numerous causes. For example, delayed mortality may have occurred in trees that were weakened by drought-induced physiological stress during the severe drought of the early 2000s but did not die until several years later (cf. Franklin et al., 1987). In addition, juniper routinely experiences partial crown die-back, though this often does not result in death of the entire tree (Plaut et al., 2012). Background mortality (i.e., non-anomalous tree death) rates of 1–5% have been observed across a wide range of forest ecosystems (Hurst et al., 2012; Laarmann et al., 2009; Metsaranta et al., 2008; van Mantgem et al., 2009). Tree mortality occurring outside of restoration and burned areas between 2006 and 2011 translates to an average annual mortality rate of approximately 1.3%, which is within the expected background rate of forests in general. 5. Conclusions We have demonstrated that highly accurate discrimination between live and dead tree pixels is possible using multitemporal, high spatial resolution satellite imagery. We found that unsupervised clustering of bi-temporal differences in NDVI and RGI worked well, even though post-mortality imagery was acquired several years after leaf abscission and in many cases tree-fall. When combined with land cover information, our classified imagery was useful for quantifying landscape-scale spatial and temporal dynamics of tree mortality occurring in an area that experienced drought, bark beetle outbreak, fire, and management activities across areas with different species composition. These results demonstrate a strength of our methodology — it is capable of detecting tree mortality arising from several causes and in several different forest types. These results also demonstrate the potential to increasingly automate baseline image classification work that is an essential precursor to change detection analyses at broader spatial scales. Although our work was focused on pixel-level analyses, future studies can shift the unit of analysis from the pixel to the tree by fusing object-based tree delineation and pixel-level mortality detection, thus increasing the relevance of tree mortality maps for ecological studies. Acknowledgments This project was funded by LANL-LDRD. The authors thank Kay Beeley and Collin Haffey for collecting and processing field data, Mort Canty and Allan Nielsen for providing open access to the IR-MAD code, and Alexei Skurikhin and Park Williams for helpful discussion. We are also grateful for the helpful comments and suggestions provided by three anonymous reviewers. Appendix A. Supplementary data Supplementary data associated with this article can be found in the online version, at http://dx.doi.org/10.1016/j.rse.2012.10.029. These data include Google maps of the most important areas described in this article. References Adams, H. D., Guardiola-Clarmonte, M., Barron-Gafford, G. A., Villegas, J. C., Breshears, D. D., Zou, C. B., et al. (2009). Temperature sensitivity of drought-induced tree mortality portends increased regional die-off under global-change-type drought. Proceedings of the National Academy of Sciences, 106, 7063–7066. Akaike, H. (1974). A new look at the statistical model identification. IEEE Transactions on Automatic Control, AC-19, 716–723. Allen, C.D. 1989. Changes in the Landscape of the Jemez Mountains, New Mexico. Ph.D. dissertation, Dept. of Forestry and Natural Resources, University of California, Berkeley, CA. 346 pp. Allen, C. D. (2007). Interactions across spatial scales among forest dieback, fire, and erosion in northern New Mexico landscapes. Ecosystems, 10, 797–808.

64

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65

Allen, C. D., & Breshears, D. D. (1998). Drought-induced shift of a forest-woodland ecotone: Rapid landscape response to climate variation. Proceedings of the National Academy of Sciences of the United States of America, 95, 14839–14842. Allen, C. D., Macalady, A. K., Chenchouni, H., Bachelet, D., McDowell, N., Vennetier, M., et al. (2010). A global overview of drought and heat-induced tree mortality reveals emerging climate change risks for forests. Forest Ecology and Management, 259, 660–684. Amiro, B. D., Barr, A. G., Barr, J. G., Black, T. A., Bracho, R., Brown, M., et al. (2010). Ecosystem carbon dioxide fluxes after disturbance in forests of North America. Journal of Geophysical Research, 115, G00K02. http://dx.doi.org/10.1029/2010JG001390. Bowker, M. A., Muñoz, A., Martinez, T., & Lau, M. K. (2012). Rare drought-induced mortality of juniper is enhanced by edaphic stressors and influenced by stand density. Journal of Arid Environments, 76, 9–16. Breshears, D. D., Cobb, N. S., Rich, P. M., Price, K. P., Allen, C. D., Balice, R. G., et al. (2005). Regional vegetation die-off in response to global-change-type drought. Proceedings of the National Academy of Sciences, 102, 15144–15148. Breshears, D. D., Myers, O. B., Meyer, C. W., Barnes, F. J., Zou, C. B., Allen, C. D., et al. (2008). Tree die-off in response to global change-type drought: Mortality insights from a decade of plant water potential measurements. Frontiers in Ecology and the Environment, 7, 185–189. Brumby, S. P., Koch, S. W., & Hansen, L. A. (2002). Evolutionary computation and postwildfire land-cover mapping with multispectral imagery. Proceedings of SPIE, 4545, 174–183. Brumby, S. P., Theiler, J., Perkins, S. J., Harvey, N. R., Szymanski, J. J., Block, J. J., et al. (1999). Investigation of image feature extraction by a genetic algorithm. Proceedings of SPIE, 3812, 24–31. Canty, M. J., & Nielsen, A. A. (2008). Automatic radiometric normalization of multitemporal satellite imagery with the iteratively re-weighted MAD transformation. Remote Sensing of Environment, 112, 1025–1036. Carnicer, J., Coll, M., Ninyerola, M., Pons, X., Sánchez, G., & Peñuelas, J. (2011). Widespread crown condition decline, food web disruption, and amplified tree mortality with increased climate change-type drought. Proceedings of the National Academy of Sciences of the United States of America, 108, 1474–1478. Coggins, S., Coops, N. C., & Wulder, M. A. (2008). Initialization of an insect infestation spread model using tree structure and spatial characteristics derived from high spatial resolution digital aerial imagery. Canadian Journal of Remote Sensing, 34, 485–502. Coops, N. C., Johnson, M., Wuler, M. A., & White, J. C. (2006). Assessment of QuickBird high spatial resolution imagery to detect red attack damage due to mountain pine beetle infestation. Remote Sensing of Environment, 103, 67–80. Dempster, A. P., Laird, N. M., & Rubin, D. B. (1977). Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statistical Society: Series B, 39, 1–38. Dennison, P. E., Brunelle, A. R., & Carter, V. A. (2010). Assessing canopy mortality during a mountain pine beetle outbreak using GeoEye-1 high spatial resolution satellite data. Remote Sensing of Environment, 114, 2431–2435. DeRose, R. J., Long, J. N., & Ramsey, R. D. (2011). Combining dendrochronological data and the disturbance index to assess Engelmann spruce mortality caused by a spruce beetle outbreak in southern Utah, USA. Remote Sensing of Environment, 115, 2342–2349. Dobbertin, M., & Brang, P. (2001). Crown defolitation improves tree mortality models. Forest Ecology and Management, 141, 271–284. Eitel, J. U. H., Vierling, L. A., Litvak, M. E., Long, D. S., Schulthess, U., Ager, A. A., et al. (2011). Broadband, red-edge information from satellites improves early stress detection in a New Mexico conifer woodland. Remote Sensing of Environment, 115, 3640–3646. Floyd, M. L., Clifford, M., Cobb, N. S., Hanna, D., Delph, R., Ford, P., et al. (2009). Relationship of stand characteristics to drought-induced mortality in three Southwestern piñon-juniper woodlands. Ecological Applications, 19, 1223–1230. Franklin, J. F., Shugart, H. H., & Harmon, M. E. (1987). Tree death as an ecological processes. BioScience, 37, 550–556. Frolking, S., Palace, M. W., Clark, D. B., Chambers, J. Q., Shugart, H. H., & Hurtt, G. C. (2009). Forest disturbance and recovery: A general review in the context of spaceborne remote sensing of impacts on aboveground biomass and canopy structure. Journal of Geophysical Research, 114, G00E02. http://dx.doi.org/10.1029/2008JG000911. Ganey, J. L., & Vojta, S. C. (2012). Trends in snag populations in drought-stressed mixed-conifer and ponderosa pine forests (1997–2007). International Journal of Forestry Research, 2012, 1–8. http://dx.doi.org/10.1155/2012/529197. Garbarino, M., Mondino, E. B., Lingua, E., Nagel, T. A., Dukić, V., Govedar, Z., et al. (2012). Gap disturbances and generation patterns in a Bosnian old-growth forest: a multispectral remote sensing and ground-based approach. Annals of Forest Science, 69, 617–625. Goodwin, N., Coops, N. C., & Stone, C. (2005). Assessing plantation canopy condition from airborne imagery using spectral mixture analysis and fractional abundances. International Journal of Applied Earth Observation and Geoinformation, 7, 11–28. Hanson, P. J., & Weltzin, J. F. (2000). Drought disturbance from climate change: Response of United States forests. The Science of the Total Environment, 262, 205–220. Harvey, N. R., Theiler, J., Brumby, S. P., Perkins, S., Szymanski, J. J., Bloch, J. J., et al. (2002). Comparison of GENIE and conventional supervised classifiers for multispectral image feature extraction. IEEE Transactions on Geoscience and Remote Sensing, 40, 393–404. Hicke, J. A., & Logan, J. (2009). Mapping whitebark pine mortality caused by a mountain pine beetle outbreak with high spatial resolution satellite imagery. International Journal of Remote Sensing, 17, 4427–4441. Hilker, T., Coops, N. C., Coggins, S. B., Wulder, M. A., Brownc, M., Black, T. A., et al. (2009). Detection of foliage conditions and disturbance from multi-angular high spectral resolution remote sensing. Remote Sensing of Environment, 113, 421–434.

Holden, Z. A., Morgan, P., Smith, A. M. S., & Vierling, L. A. (2010). Beyond Landsat: a comparison of four satellite sensors for detecting burn severity in ponderosa pine forests of the Gila Wilderness, NM, USA. International Journal of Wildland Fire, 19, 449–458. Holland, J. H. (1975). Adaptation in natural and artificial systems. Ann Arbor: The University of Michigan Press. Huang, C. Y., Asner, G. P., Barger, N. N., Neff, J. C., & Floyd, M. L. (2010a). Regional aboveground live carbon losses due to drought-induced tree dieback in piñon-juniper ecosystems. Remote Sensing of Environment, 114, 1471–1479. Huang, C., Goward, S. N., Masek, J. G., Thomas, N., Zhu, Z., & Vogelmann, J. E. (2010b). An automated approach for reconstructing recent forest disturbance history using dense Landsat time series stacks. Remote Sensing of Environment, 114, 183–198. Hurst, J. M., Stewart, G. H., Perry, G. L. W., Wiser, S. K., & Norton, D. A. (2012). Determinants of tree mortality in mixed old-growth Nothofagus forest. Forest Ecology and Management, 270, 189–199. Ismail, R., Mutanga, O., & Bob, U. (2007). Forest health and vitality: The detection and monitoring of Pinus patula trees infected by Sirex noctilio using digital multispectral imagery. Southern Hemisphere Forestry Journal, 69, 39–47. Kane, J. M., Meinhardt, K. A., Chang, T., Cardall, B. L., Michalet, R., & Whitham, T. G. (2011). Drought-induced mortality of a foundation species (Juniperus monosperma) promotes positive afterlife effects in understory vegetation. Plant Ecology, 212, 733–741. Kelly, M., Shaari, D., Guo, Q., & Liu, D. (2004). A comparison of standard and hybrid classifier methods for mapping hardwood mortality in areas affected by “Sudden Oak Death”. Photogrammetric Engineering and Remote Sensing, 70, 1229–1239. Kennedy, R. E., Yang, Z., & Cohen, W. B. (2010). Detecting trends in forest disturbance and recovery using yearly Landsat time series: 1. LandTrendr – Temporal segmentation algorithms. Remote Sensing of Environment, 114, 2897–2910. Laarmann, D., Korjus, H., Sims, A., Stanturf, J. A., Kiviste, A., & Köster, K. (2009). Analysis of forest naturalness and tree mortality patterns in Estonia. Forest Ecology and Management, 285S, S187–S195. Laben, C. A., & Brower, B. V. 2000. Process for enhancing the spatial resolution of multispectral imagery using pan-sharpening. Technical Report. U.S. Patent # 6,011,875, Eastman Kodak Company. McDowell, N. G., Allen, C., & Marshall, L. (2010). Growth, carbon isotope discrimination, and mortality across a ponderosa pine elevation transect. Global Change Biology, 16, 399–415. McDowell, N. G., Beerling, D. J., Breshears, D. D., Fisher, R. A., Raffa, K. F., & Stitt, M. (2011). The interdependence of mechanisms underlying climate-driven vegetation mortality. Trends in Ecology & Evolution, 26, 523–532. McDowell, N. G., Pockman, W. T., Allen, C. D., Breshears, D. D., Cobb, N., Kolb, T., et al. (2008). Mechanisms of plant survival and mortality during drought: why do some plants survive while others succumb to drought? New Phytologist, 178, 719–739. Meddens, A. J. H., Hicke, J. A., & Vierling, L. A. (2011). Evaluating the potential of multispectral imagery to map multiple stages of tree mortality. Remote Sensing of Environment, 115, 1632–1642. Metsaranta, J. M., Lieffers, V. J., & Wein, R. W. (2008). Dendrochronological reconstruction of jack pine snag and downed log dynamics in Saskatchewan and Manitoba, Canada. Forest Ecology and Management, 255, 1262–1270. Mildrexler, D. J., Zhao, M., & Running, S. W. (2009). Testing a MODIS Global Disturbance Index across North America. Remote Sensing of Environment, 113, 2103–2117. Oumar, Z., & Mutanga, O. (2011). The potential of remote sensing technology for the detection and mapping of Thaumastocoris peregrines in plantation forests. Southern Forests, 73, 23–31. Pasher, J., & King, D. J. (2009). Mapping dead wood distribution in a temperate hardwood forest using high resolution airborne imagery. Forest Ecology and Management, 258, 1536–1548. Perkins, S., Harvey, N., Brumby, S., & Lacker, K. (2001). Support vector machines for broad area feature extraction in remotely sensed images. Proceedings of SPIE, 4381, 286–295. Phillips, O. L., Aragão, L. E. O. C., Lewis, S. L., Fisher, J. B., Lloyd, J., Lόpez-González, G., et al. (2009). Drought sensitivity of the Amazon rainforest. Science, 323, 1344–1347. Plaut, J. E., Yepez, A., Hill, J., Pangle, R., Johnson, J., Pockman, W. T., et al. (2012). Hydraulic limits on water use under experimental drought in a piñon-juniper woodland. Plant, Cell & Environment, 35, 1601–1617. Rich, P. M., Breshears, D. D., & White, A. B. (2008). Phenology of mixed woodyherbaceous ecosystems following extreme events: Net and differential responses. Ecology, 89, 342–352. Rich, R. L., Frelich, L., Reich, P. B., & Bauer, M. E. (2010). Detecting wind disturbance severity and canopy heterogeneity in boreal forest by coupling high-spatial resolution satellite imagery and field data. Remote Sensing of Environment, 114, 299–308. Shaw, J. D., Steed, B. E., & DeBlander, L. T. (2005). Forest inventory and analysis (FIA) annual inventory answers the question: What is happening to pinyon-juniper woodlands? Journal of Forestry, 103, 280–285. U S Forest Service (2002). Forest insect and disease conditions in the southwestern region, 2002. United States Department of Agriculture (R3-03-01). U S Forest Service (2003). Forest insect and disease conditions in the southwestern region, 2003. United States Department of Agriculture (R3-04-02). U S Forest Service (2004). Forest insect and disease conditions in the southwestern region, 2004. United States Department of Agriculture (R3-05-0). U S Forest Service (2005). Forest insect and disease conditions in the southwestern region, 2005. United States Department of Agriculture (PR-R3-16-1). van der Molen, M. K., Dolman, A. J., Ciais, P., Eglin, T., Gobron, N., Law, B. E., et al. (2011). Drought and ecosystem carbon cycling. Agricultural and Forest Meteorology, 151, 765–773.

S.R. Garrity et al. / Remote Sensing of Environment 129 (2013) 54–65 van Mantgem, P. J., Stephenson, N. L., Byrne, J. C., Daniels, L. D., Franklin, J. F., Fulé, P. Z., et al. (2009). Widespread increase of tree mortality rates in the western United States. Science, 323, 521–524. Vogelmann, J. E., Tolk, B., & Zhu, Z. (2009). Monitoring forest changes in the southwestern United States using multitemporal Landsat data. Remote Sensing of Environment, 113, 1739–1748. Vogelmann, J. E., Xian, G., Homer, C., & Tolk, B. (2012). Monitoring gradual ecosystem change using Landsat time series analyses: case studies in selected forest and rangeland ecosystems. Remote Sensing of Environment, 122, 92–105. White, J. C., Wulder, M. A., Brooks, D., Reich, R., & Wheate, R. D. (2005). Detection of red attack stage mountain pine beetle infestation with high spatial resolution satellite imagery. Remote Sensing of Environment, 96, 340–351. Williams, A. P., Allen, C. D., Macalady, A. K., Griffin, D., Woodhouse, C. A., Meko, D. M., et al. (2012). Influence of temperature on regional forest-drought stress and tree mortality. Nature Climate Change. http://dx.doi.org/10.1038/nclimate1693. Wulder, M. A., Dymond, C. C., White, J. C., Leckie, D. G., & Carroll, A. L. (2006a). Surveying mountain pine beetle damage of forests: A review of remote sensing opportunities. Forest Ecology and Management, 221, 27–41.

65

Wulder, M. A., Ortlepp, S. M., White, J. C., Coops, N. C., & Coggins, S. B. (2009). Monitoring the impacts of mountain pine beetle mitigation. Forest Ecology and Management, 258, 1181–1187. Wulder, M. A., White, J. C., Bentz, B., Alvarez, M. F., & Coops, N. C. (2006b). Estimating the probability of mountain pine beetle red-attack damage. Remote Sensing of Environment, 101, 150–166. Wulder, M. A., White, J. C., Coggins, S., Ortlepp, S. M., Coops, N. C., Heath, J., et al. (2012). Digital high spatial resolution aerial imagery to support forest health monitoring: the mountain pine beetle context. Journal of Applied Remote Sensing, 6. http://dx.doi.org/ 10.1117/1.JRS.6.062527. Wulder, M. A., White, J. C., Coops, N. C., & Butson, C. R. (2008). Multi-temporal analysis of high spatial resolution imagery for disturbance monitoring. Remote Sensing of Environment, 112, 2729–2740.