ArticlePDF Available

Mapping forest functional type in a forest-shrubland ecotone using SPOT imagery and predictive habitat distribution modelling

Authors:

Abstract and Figures

The availability of land cover data at local scales is an important component in forest management and monitoring efforts. Regional land cover data seldom provide detailed information needed to support local management needs. Here we present a transferable framework to model forest cover by major plant functional type using aerial photos, multi-date Système Pour l’Observation de la Terre (SPOT) imagery, and topographic variables. We developed probability of occurrence models for deciduous broad-leaved forest and needle-leaved evergreen forest using logistic regression in the southern portion of the Wyoming Basin Ecoregion. The model outputs were combined into a synthesis map depicting deciduous and coniferous forest cover type. We evaluated the models and synthesis map using a field-validated, independent data source. Results showed strong relationships between forest cover and model variables, and the synthesis map was accurate with an overall correct classification rate of 0.87 and Cohen’s kappa value of 0.81. The results suggest our method adequately captures the functional type, size, and distribution pattern of forest cover in a spatially heterogeneous landscape.
Content may be subject to copyright.
Mapping Forest Functional Type in a Forest-Shrubland Ecotone using
SPOT Imagery and Predictive Habitat Distribution Modeling
TIMOTHY J. ASSALa, b*, PATRICK J. ANDERSONa, and JASON
SIBOLDc,
aU.S. Geological Survey (USGS), Fort Collins Science Center, 2150 Centre Avenue,
Fort Collins, CO 80526,
bGraduate Degree Program in Ecology, Colorado State University, 1401 Campus
Delivery, Fort Collins, CO 80523, USA
cDepartment of Anthropology, Colorado State University, 1787 Campus Delivery, Fort
Collins, CO 80523, USA
*Corresponding author at: USGS, Fort Collins Science Center, Fort Collins, CO
80526, USA. Tel.: +1 970 226 9134.
E-mail address: assalt@usgs.gov (T.J. Assal).
Pre-print of published version.
Reference:
Assal, T., Anderson, P., Sibold, J., 2015. Mapping forest functional type in a forest-
shrubland ecotone using SPOT imagery and predictive habitat distribution modelling.
Remote Sens. Letters. 6, 755–764. doi:10.1080/2150704X.2015.1072289
http://www.tandfonline.com/doi/full/10.1080/2150704X.2015.1072289
Disclaimer:
The PDF document is a copy of the final version of this manuscript that was
subsequently accepted by the journal for publication. The paper has been through peer
review, but has not been subject to any additional copy-editing or journal specific
formatting (so will look different from the final version of record, which may be
accessed following the DOI above depending on your access situation)
Mapping Forest Functional Type in a Forest-Shrubland Ecotone using
SPOT Imagery and Predictive Habitat Distribution Modeling
The availability of land cover data at local scales is an important component in
forest management and monitoring efforts. Regional land cover data seldom
provide detailed information needed to support local management needs. Here
we present a transferable framework to model forest cover by major plant
functional type using aerial photos, multi-date Système Pour l'Observation de la
Terre (SPOT) imagery, and topographic variables. We developed probability of
occurrence models for deciduous broad-leaved forest and needle-leaved
evergreen forest using logistic regression in the southern portion of the Wyoming
Basin Ecoregion. The model outputs were combined into a synthesis map
depicting deciduous and coniferous forest cover type. We evaluated the models
and synthesis map using a field validated, independent data source. Results
showed strong relationships between forest cover and model variables and the
synthesis map was accurate with an overall correct classification rate of 0.87 and
Cohen’s kappa value of 0.81. The results suggest our method adequately captures
the functional type, size and distribution pattern of forest cover in a spatially
heterogeneous landscape.
Keywords: Rocky Mountain forests, spatial heterogeneity, SPOT imagery, predictive habitat
distribution models, logistic regression modeling, Populus tremuloides
1. Introduction
Land cover data provides the foundation for a wide variety of geographical analysis and
science applications. There have been several national and regional land cover mapping
initiatives over the last two decades, most notably the National Land Cover Database
(NLCD) (Jin et al. 2013), LANDFIRE (Rollins 2009), and ReGAP (Davidson et al.
2009). These products have provided tremendous utility in studies documenting land
cover change (Radeloff et al. 2005), effects of climate change (Wylie et al. 2014),
vegetation change (Bradley and Fleishman 2008) and conservation planning (Stoms
2000). However, in some heterogeneous ecosystems, they lack the spatial resolution
needed to adequately characterize the extent and juxtaposition of land cover. In the
Wyoming Basin ecoregion, small areas of forest are found within sagebrush shrubland
at higher elevations which are difficult to adequately characterize using regional land
cover data.
Plant functional types (PFTs) are groups of species that share similar structural,
physiological and phenological traits (Barbour et al. 1999). PFTs provide a framework
to consider how species utilize resource availability and respond to environmental
change and management. At the stand scale, multispectral remote sensing can be used to
delineate the relationship between vegetation structure and physiology of PFTs, linking
biophysical properties to ecological theory (Ustin and Gamon 2010). Predictive habitat
distribution modeling offers the potential to assess the current extent of species as well
as effects of global climate change and other change agents. However, a challenge is to
better link remote sensing data to underlying ecological relationships and describe the
distribution of species along environmental gradients (Zimmermann et al. 2007).
Incorporating remote sensing variables in the modeling process allows us to take full
advantage of continuous gradients to delineate biophysical properties of vegetation such
as leaf shape, structure, longevity and chlorophyll content (Jones and Vaughan 2010).
We explored the potential of fine-scale remotely sensed spectral data in
predictive habitat distribution modeling of forest cover type across a forest-shrubland
ecotone. The two PFTs of interest in the study area are deciduous broad-leaved forest
(referred to as deciduous forest) and montane needle-leaved evergreen forest (referred
to as coniferous forest). We hypothesized that the delineation between forest functional
type would be aided with the addition of multitemporal remote sensing predictors due to
differences in phenology between deciduous and coniferous species (Bergen and
Dronova 2007; Zimmermann et al. 2007). The major goal of the study was to develop
an operational mapping framework using aerial photos coupled with fine-scale satellite
imagery to efficiently model dominant forest cover. The specific objectives were to: 1)
develop probability of occurrence models for deciduous and coniferous forest; and 2) to
combine model outputs into a field-validated synthesis map depicting forest cover type.
2. Study Area
The study area, managed largely by the U.S. Bureau of Land Management, is located in
the southern part of the Wyoming Basin ecoregion, spanning parts of southwestern
Wyoming, northwestern Colorado, and northeastern Utah (Figure 1). Several prominent
ridges form a transition zone between basins and mountainous areas (Knight 1994),
where several species of trees exist at the xeric fringes of their respective ranges.
Forests are dominated by either aspen (Populus tremuloides) or several coniferous
species, namely subalpine fir (Abies lasiocarpa), Douglas-fir (Pseudotsuga menziesii),
and lodgepole pine (Pinus contorta), that occur as relatively small patches on moist
sites in a matrix of mountain sagebrush (Artemisia tridendata spp. vaseyana) or mixed-
species shrublands. Scattered juniper (Juniperus communis var. depressa) and limber
pine (Pinus flexilis) woodlands, distinct from the montane conifer forest, are found on
rocky slopes at lower elevations and small patches of manzanita (Arctostaphylos patula)
are found in the southern part of the study area. The area has a midlatitude steppe
climate with a substantial portion of the annual precipitation occurring as snow.
Multiple state and federal agencies, along with the Wyoming Landscape Conservation
Initiative (wlci.gov), have identified the region as a priority area for conservation given
the important habitat it provides for many wildlife species. Drought-related mortality of
aspen is a concern in western North America (Worrall et al. 2008), and lack of aspen
regeneration due to high rates of herbivory is a concern locally. Active management
seeks to address these concerns and locally accurate maps of forest cover type are
critical to support conservation and monitoring efforts.
3. Methods
3.1 Explanatory Variables
We identified a contiguous area (1,088 km2) greater than 2,300 m in elevation known to
encompass the forest communities of the study area (Figure 1). We explored the
relationship of topographic and multi-date remotely sensed variables to forest presence
that have had utility in other SDMs (Turner et al. 2003; Zimmermann et al. 2007;
Jarnevich et al. 2014; Engler et al. 2013). Topographic variables were derived from a 10
m National Elevation Data set and remotely sensed variables were derived from terrain
corrected Level 1 T SPOT 5 HRG satellite imagery, acquired at no cost (USGS 2014)
(Table 1). We obtained two cloud-free dates during leaf-on (07 September 2010) and
leaf-off (19 October 2010) conditions from two SPOT scenes (KJ grid 555-267; 555-
286). Each of the four images were geometrically registered to National Agriculture
Imagery Program (NAIP) aerial photos using 20-25 ground control points with a root
mean square error of less than 0.5 pixel. A top-of-atmosphere correction was applied to
each image to account for differences in sensor and viewing angle (Wulder et al. 2006;
Vogelmann et al. 2012; Sankey, Moffet, and Weber 2008).
Remotely sensed spectral bands and derived vegetation indices often exhibit high
levels of collinearity (Engler et al. 2013). Multi-collinearity among all potential
explanatory variables was assessed prior to model calibration using the Pearson’s
correlation coefficient. Variables with a correlation coefficient greater than 0.8 or less
than -0.8 were removed from consideration within the same model (Jarnevich et al.
2014). All of the analysis was conducted using the R statistical package (R
Development Core Team 2013).
3.2 Sample Data
ReGAP land cover data was reclassified into eight land cover categories (including
deciduous and coniferous forest), and we used a stratified random selection procedure to
ensure an unbiased distribution of sample plots (10 m x 10 m, congruent with a SPOT
pixel) across land cover types. Our objective was to capture fine-scale patterns in the
study area, while minimizing the impact of spatial dependency between observations.
We used a total of 545 plots, with a minimum distance of 250 m (25 pixels) between
each plot, to develop presence and absence records. We interpreted recent aerial
photographs (natural color and infrared NAIP) and classified each plot as deciduous
forest, coniferous forest, or non-forest. Mixed forest is not found at broad scales in the
study area, and it is difficult to reliably classify a 10 m NAIP plot as mixed forest.
Absence records for deciduous forest included both non-forest and coniferous forest
plots, whereas absence records for coniferous forest included non-forest and deciduous
forest. We extracted the values of the 19 predictor variables (Table 1) at each sample
location. The respective leaf-on and leaf-off periods for each band exhibited high
collinearity. Leaf-on bands 1, 2, and 4 were also highly correlated along with leaf-on
NDVI. We opted to use leaf-on bands over leaf-off, although leaf-off data was
incorporated into ΔNDVI (Table 1). We retained band 1 over band 2 since information
from band 2 is incorporated into the NDVI variable. This selection process resulted in
13 variables for consideration (Table 1), including longitude and latitude to account for
spatial autocorrelation (Knapp et al. 2003; Hu and Lo 2007).
3.3 Data Analysis
Logistic regression is a widely used method to predict the probability of a
dichotomous variable (i.e., presence, absence of a forest cover type) that has been used
in species distribution modeling (SDM) (Engler et al. 2013; Jarnevich et al. 2014;
Stohlgren et al. 2010) and other ecological studies (Turner et al. 2003; Wulder et al.
2006; Dubovyk et al. 2013). We used a multivariate generalized linear model (GLM,
binomial distribution, logit link function) to create independent models of deciduous
(DECID) and coniferous (CONIF) forest cover in the study area. We modeled each
cover type independently to maximize information contained in the continuous gradient
of biophysical characteristics of these systems. The full dataset (n = 545) was used for
model calibration. To simplify the interpretation of the logistic model, we converted the
regression coefficients into odds ratios, then calculated the percent change in odds
(Wulder et al. 2006). This approach identifies the percent change in the probability of a
pixel containing deciduous or coniferous forest relative to changes in independent
variables.
We tested several models for each type of forest cover, using different combinations
of predictor variables below the acceptable collinearity threshold. For each model a
standard stepwise selection by Akaike’s Information Criterion (AIC) was used to select
the best subset of independent variables and we calculated variance inflation factors
(VIF) to ensure all model variables had a value below 5 (Dubovyk et al. 2013). In
logistic regression, spatial autocorrelation violates the assumption that observations are
independent and can cause unreliable estimates of the model parameters (Hu and Lo
2007). We evaluated spatial autocorrelation using the Moran’s I statistic on each model
using Pearson residuals which are comparable to residuals of linear regression models.
The neighborhood structure of the spatial weights matrix was defined using inverse
distance (Assal, Sibold, and Reich 2014).
The two selected models produced a continuous surface with values between zero
and one corresponding to the probability of a pixel containing either deciduous or
coniferous forest. We used an independent data set (n=321) representing coniferous,
deciduous and non-forest observations to evaluate each model. The reference data set
was compiled from observations of related studies that were visited in the field between
2010 and 2013. We randomly selected 100 presence and 100 absence points of the
respective forest cover type for each model to calculate the receiver operator
characteristic area under the curve (AUC). For each model, we selected a threshold
where the sensitivity was equal to the specificity (Liu et al. 2005) (i.e. the number of
false positives were equal to the number of false negatives) to convert each model
output into a binary map of presence and absence.
3.4 Synthesis Map
The binary maps from the two models were combined into a synthesis map of
deciduous and coniferous forest cover. If a pixel was predicted to contain both forest
types, the values from each model above the presence threshold were linearly rescaled
from 0 to 1. The cover type with the highest occurrence probability was then assigned to
the pixel (Engler et al. 2013). Within each cover type, pixels were assigned to neighbors
in all eight directions (‘queen’s move’) to identify contiguous forest patches. Patches of
three pixels and greater were retained in order to minimize small, likely incorrect
classified areas. We used the full, validation data set (n=321) to build a confusion
matrix to calculate the overall classification accuracy and Cohen’s kappa coefficient on
the synthesis map.
4. Results
Forest cover type was best predicted by a combination of topographic and spectral
variables, and the two models included several common variables selected through the
model fitting process (Table 2). The residuals of both models exhibited very weak or no
spatial autocorrelation (DECID; Moran’s I=0.005, p=0.052; CONIF; Moran’s I=0.001,
p=0.38). Aspen, the only deciduous forest type present, have a clonal growth form
which produces clustered patches of deciduous forest in the study area. Both of the
models had high accuracy with AUC values of 0.92 for DECID and 0.99 for CONIF.
The percentage changes in the odds ratio for model variables are shown in Table 2. The
DECID model indicated that the presence of deciduous forest is mainly associated with
high values in the NIR band (band3.leaf-on), north facing slopes, and high values of
ΔNDVI. Deciduous forest is more likely to be found in areas higher in elevation,
particularly moderate elevations (classes 3 and 2). The model also indicated that
deciduous forest was less likely associated with higher values in the green band
(band1.leaf-on) and higher TPI values, found along ridge lines and hilltops. The CONIF
model indicated presence of coniferous forest is mainly associated with high values in
NDVI (NDVI.leaf-on), TPI and elevation and less likely associated with high values in
the NIR band (band3.leaf-on) and areas that experience a higher HLI.
A comparison of the output maps, derived from the binary models, revealed high
separation between the two models. Less than 1% of all pixels in the study area were
predicted to contain both deciduous and coniferous forest. The comparison indicated
satisfactory agreement with an overall classification accuracy of 87% (Table 3) and a
Cohen’s kappa coefficient (Cohen 1960) of 0.81. Our synthesis map identified 61.7 km2
of forest (Figure 2). Deciduous forest accounts for 44% (27.2 km2) of total forest cover,
while the remaining 56% (34.5 km2) is coniferous forest. There are over 7,000 patches
of deciduous forest compared to less than 2,400 coniferous forest patches, and the mean
patch size is much smaller for deciduous (0.004 km2) compared to coniferous forest
(0.015 km2).
5. Discussion
The spatial resolution (10 m) of the SPOT imagery appeared particularly appropriate for
identifying the extent and pattern of forest cover in this highly heterogeneous ecosystem
(Figure 3). The models performed well and had little overlap between forest functional
types. Furthermore, we considered spatial autocorrelation in our framework which is
often overlooked in SDMs. The inclusion of latitude and longitude in the model, as well
as treating elevation as an indicator variable, accounted for spatial autocorrelation in the
models. Consideration of plant physiology and species traits illuminates the ecological
context of biophysical variables that were captured with leaf-on and leaf-off SPOT
imagery. Due to differences in seasonal phenology, deciduous forests are more likely to
be associated with areas that have a large ΔNDVI between leaf-on and leaf-off periods.
Aspen leaves absorb more radiation (and reflect less) in the green region of the
electromagnetic spectrum compared to sagebrush and grassland plants (Jones and
Vaughan 2010). Therefore, deciduous forest is associated with lower values in the green
band (band1.leaf-on) compared to shrubland and grasslands. Coniferous forests have
higher NDVI values than grassland and shrubland, especially late in the growing season
when the leaf-on image was acquired. Although deciduous forests have a higher NDVI
than coniferous forests, NDVI was not as important in DECID, which also included
ΔNDVI and the NIR band (band3.leaf-on). Coniferous forests do not reflect as much
radiation as aspen forest, grassland, and sagebrush and, therefore, are more likely to be
associated with lower values in the NIR (band3.leaf-on).
In our synthesis map, the non-forest class had the lowest producer’s accuracy (the
percentage of reference observations correctly mapped), due to Type 1 errors (i.e. false-
positives) in the deciduous forest class (Table 3). The DECID model incorrectly
classified some small areas of manzanita, non-forested riparian areas, narrow, linear
bands of deciduous shrubs and one large herbaceous wetland. There is a trade-off
between capturing small patches of deciduous forest and incorrectly classifying some
small areas with similar spectral values to deciduous forest. In future work, our
methodology could be improved with additional information that discriminates between
herbaceous wetlands and deciduous forest. The deciduous forest class had the lowest
users accuracy (the percentage of map locations correctly identified) due to several
type II errors (i.e. false-negatives). The higher rate of misclassification might be
explained by the open canopy of aspen forests compared to closed, dense coniferous
forest. Furthermore, some aspen stands observed in the field had a substantial amount of
canopy dieback and tree mortality. Aspen forests with highly reduced leaf area at the
time the satellite image was acquired have lower reflectance values and therefore tend
to be mapped as non-forest. Although this can be a drawback, it highlights the value of
our methodology to capture changes in aspen forest if fine-scale imagery is available
over an appropriate time period. The CONIF model performed very strongly as needle-
leaved species have a closed canopy that forms dense stands. A few exceptions include
several conifer stands dominated by Douglas-fir that experienced high mortality in
recent years due to infestation of the Douglas-fir beetle (Dendroctonus pseudotsugae).
6. Conclusions
We have presented a framework that incorporates aerial photos and satellite imagery to
model dominant forest cover at local scales across a forest-shrubland ecotone. Our
modeling process offers a powerful alternative to traditional image classification and
our synthesis map provides managers with an important tool to support conservation
and monitoring efforts across management unit boundaries. Our study highlights the
advantages of using physiologically relevant remote sensing products in predictive
modeling and addresses an important research need (e.g. high-resolution remote sensing
of aspen distribution, (Kulakowski, Kaye, and Kashian 2013)). We conclude that our
approach is suitable to characterize the extent and juxtaposition of forest cover in a
highly heterogeneous ecosystem. Furthermore, our framework utilizes open access
aerial photos and satellite data. In this way it is transferable to highly heterogeneous
ecosystems to develop critical baseline tree cover data that can be updated at regular
intervals to monitor the effects of disturbance and long-term ecosystem dynamics.
Acknowledgements
This research was supported by the USGS Fort Collins Science Center and the Wyoming
Landscape Conservation Initiative. Logistical support was provided by the Wyoming Game and
Fish Department and Bureau of Land Management Rock Springs Field Office. We gratefully
acknowledge Geneva Chong and two anonymous reviewers for their constructive comments.
We thank Lucy Burris, Brian Cade, Alexandra Urza, Marie Dematatis, and Darlene Kilpatrick
for discussion on research direction and/or assistance in collection of field data.
Any use of trade, firm, or product names is for descriptive purposes only and does not imply
endorsement by the U.S. Government.
Table 1. Description of the explanatory variables considered in the analysis.
Variable
Description
Spectral variables
Band1.leaf-on*
Green band (0.50 µm - 0.59 µm)
Band2.leaf-on
Red band (0.61 µm - 0.68 µm)
Band3.leaf-on*
Near infrared band (0.79 µm - 0.89 µm)
Band4.leaf-on*
Short wave infrared band (1.58 µm - 1.75 µm)
Normalized difference vegetation index
(NDVI.leaf-on*)
NDVI = (B3 reflectance B2 reflectance )/(B3
reflectance + B2 reflectance) (Rousse et al., 1974)
Band1.leaf-off
Green band (0.50 µm - 0.59 µm)
Band2.leaf-off
Red band (0.61 µm - 0.68 µm)
Band3.leaf-off
Near infrared band (0.79 µm - 0.89 µm)
Band4.leaf-off
Short wave infrared band (1.58 µm - 1.75 µm)
Normalized difference vegetation index
(NDVI.leaf-off)
NDVI = (B3 reflectance B2 reflectance )/(B3
reflectance + B2 reflectance) (Rousse et al., 1974)
ΔNDVI*
ΔNDVI = NDVIleaf-on NDVIleaf-off
Topographic variables
Elevation*
Derived from National elevation data set
Slope*
Derived from National elevation data set
North exposure*
Cosine transformation of aspect
East exposure*
Sine transformation of aspect
Heat load index (HLI)*
Potential direct incident radiation (McCune & Keon,
2002; equation 3)
Topographic position index (TPI)*
A measure of slope position and landform type with
respect to adjacent grid cells
Longitude*
Longitude at cell centroid
Latitude*
Latitude at cell centroid
Notes: Leaf-on variables were acquired from the 07 September 2010 SPOT image; leaf-
off variables were acquired from the 19 October 2010 SPOT image. All variables have a
spatial resolution of 10 m. The native resolution for band 4 is 20 m, but it was
resampled to 10 m using a nearest neighbour transformation.
*Indicates variable was utilized in the modeling process.
Table 2. Predictor variables used in the logistic regression models.
Explanatory variable
DECID model
CONIF model
Coefficient
Change in
odds (%)
Coefficient
Change in
Spectral variables
NDVI.leaf-on
-
-
1.9164*
ΔNDVI
0.3327*
39
-0.2914
Band1.leaf-on
-4.261*
-99
-
Band3.leaf-on
3.068*
2050
-3.7485*
Topographic variables
TPI
-0.5016*
-39
0.8194*
HLI
-
-
-0.7466*
North exposure
0.8654*
138
-0.6963
Longitude
-6.86e-05*
-
-
Latitude
-1.24e-04*
-
-
Elevation**
Class 1 (low)
-
-
-
Class 2 (low to moderate)
5.053*
15549
0.3221
Class 3 (moderate to high)
5.591*
26700
1.7885
Class 4 (High)
4.409*
-8119
3.5633*
Notes: Estimates of the model parameters are listed for each model accordingly. *p-
values are significant at 0.05 or lower. **Elevation is treated as an indicator variable;
therefore the percent change in odds for each class can only be compared to the
reference elevation class (Class 1).
Table 3. Confusion matrix for the synthesis forest cover map.
Field data
Classified as:
Producer’s
accuracy
(%)
Non-forest
Deciduous
forest
Coniferous
forest
Non-forest
95
24
3
Deciduous forest
6
86
5
Conifer forest
0
2
100
User’s accuracy (%)
94.1
76.8
92.6
Note: bold values in the matrix diagonal highlight the correctly predicted samples.
Figure 1. Location of study area. Note: approximately 50 km2 of non-forest were
omitted from the southeast corner of the study area due to the extent of the SPOT
image.
Figure 2. Results of the model outputs combined into the synthesis map: (a) Little
Mountain, (b) Middle Mountain and portions of Diamond Peak, (c) Pine Mountain and
(d) Cold Spring Mountain. Note: each map panel is displayed at the same scale.
Figure 3. Comparison of forest type maps derived from each data source of a
representative area of the landscape on Little Mountain: (a) 2009 color-infrared aerial
photo (National Agriculture Imagery Program) where dark red/black hues indicate
coniferous forest, red hues indicate deciduous forest, grey/light red/blue hues represent
non-forest and (b) USGS synthesis map. Note: each map panel is displayed at the same
scale and extent.
References
Assal, T.J., J. Sibold, and R. Reich. 2014. “Modeling a Historical Mountain Pine Beetle
Outbreak Using Landsat MSS and Multiple Lines of Evidence.” Remote Sensing of
Environment 155 (December). Elsevier B.V.: 275–88.
doi:10.1016/j.rse.2014.09.002.
Barbour, M.G., J.H. Burk, W.D. Pitts, F.S. Gilliam, and M.W. Schwartz. 1999.
Terrestrial Plant Ecology. 3rd ed. New York, New York: Benjamin Cummings.
Bergen, K.M., and I. Dronova. 2007. “Observing Succession on Aspen-Dominated
Landscapes Using a Remote Sensing-Ecosystem Approach.” Landscape Ecology
22 (9): 1395–1410. doi:10.1007/s10980-007-9119-1.
Bradley, B.A., and E. Fleishman. 2008. “Relationships between Expanding Pinyon–
juniper Cover and Topography in the Central Great Basin, Nevada.” Journal of
Biogeography 35 (5): 951–64. doi:10.1111/j.1365-2699.2007.01847.x.
Cohen, J. 1960. “A Coefficient of Agreement for Nominal Scales.” Educational and
Psychological Measurement 20: 37–46.
Davidson, A., J. Aycrigg, E. Grossmann, J. Kagan, S. Lennartz, S. McDonough, T.
Miewald, et al. 2009. “Digital Land Cover Map for the Northwestern United
States.” Moscow, ID: Northwest Gap Analysis Project: USGS GAP Analysis
Program.
Dubovyk, O., G. Menz, C. Conrad, E. Kan, M. Machwitz, and A. Khamzina. 2013.
“Spatio-Temporal Analyses of Cropland Degradation in the Irrigated Lowlands of
Uzbekistan Using Remote-Sensing and Logistic Regression Modeling.”
Environmental Monitoring and Assessment 185 (6): 4775–90. doi:10.1007/s10661-
012-2904-6.
Engler, R., L.T. Waser, N.E. Zimmermann, M. Schaub, S. Berdos, C. Ginzler and A.
Psomas. 2013. “Combining Ensemble Modeling and Remote Sensing for Mapping
Individual Tree Species at High Spatial Resolution.” Forest Ecology and
Management 310: 64–73.
Hu, Z., and C. Lo. 2007. “Modeling Urban Growth in Atlanta Using Logistic
Regression.” Computers, Environment and Urban Systems 31 (6): 667–88.
doi:10.1016/j.compenvurbsys.2006.11.001.
Jarnevich, C.S., W.E. Esaias, P.L.A. Ma, J.T. Morisette, J.E. Nickeson, T.J. Stohlgren,
T.R. Holcombe, J.M. Nightingale, R.E. Wolfe, and B. Tan. 2014. “Regional
Distribution Models with Lack of Proximate Predictors: Africanized Honeybees
Expanding North.” Edited by Mark Robertson. Diversity and Distributions 20 (2):
193–201. doi:10.1111/ddi.12143.
Jin, S., L. Yang, P. Danielson, C. Homer, J. Fry, and G. Xian. 2013. “A Comprehensive
Change Detection Method for Updating the National Land Cover Database to circa
2011.” Remote Sensing of Environment 132: 159–75.
Jones, H.G., and R.H. Vaughan. 2010. Remote Sensing of Vegetation: Principles,
Techniques, and Applications. Oxford, UK: Oxford University Press.
Knapp, R.A., K.R. Matthews, H.K. Preisler, and R. Jellison. 2003. “Developing
Probabilistic Models to Predict Amphibian Site Occupancy in a Patchy
Landscape.” Ecological Applications 13 (4): 1069–82.
Knight, D.H. 1994. Mountains and Plains: The Ecology of Wyoming Landscapes. New
Haven, Connecticut: Yale University.
Kulakowski, Dominik, Margot W. Kaye, and Daniel M. Kashian. 2013. “Long-Term
Aspen Cover Change in the Western US.” Forest Ecology and Management 299
(July). Elsevier B.V.: 52–59. doi:10.1016/j.foreco.2013.01.004.
Liu, C., P.M. Berry, T.P. Dawson, and R.G. Pearson. 2005. “Selecting Thresholds of
Occurrence in the Prediction of Species Distributions.” Ecography 28 (December
2004): 385–93.
McCune, B., and D. Keon. 2002. “Equations for Potential Annual Direct Incident
Radiation and Heat Load.” Journal of Vegetation Science 13 (4): 603–6.
R Development Core Team. 2013. “R: A Language and Environment for Statistical
Computing.” Vienna, Austria: R Foundation for Statistical Computing.
http://www.r-project.org/.
Radeloff, V.C., R.B. Hammer, S.I. Stewart, J.S. Fried, S.S. Holcomb, and J.F.
McKeefry. 2005. “The Wildland-Urban Interface in the United States.” Ecological
Applications 15 (3): 799–805.
Rollins, M.G. 2009. “LANDFIRE: A Nationally Consistent Vegetation, Wildland Fire,
and Fuel Assessment.” International Journal of Wildland Fire 18 (3): 235–49.
Rousse, J.W., R.H. Hass, J.A. Schell, D.W. Deering, and J.C. Harlan. 1974. Monitoring
the Vernal Advancement of Retrogradation of Natural Vegetation. Type III, Final
Report. Greenbelt, Maryland.
Sankey, Temuulen Tsagaan, Corey Moffet, and Keith Weber. 2008. “Postfire Recovery
of Sagebrush Communities: Assessment Using Spot-5 and Very Large-Scale
Aerial Imagery.” Rangeland Ecology & Management 61 (6): 598–604.
doi:10.2111/08-079.1.
Stohlgren, T.J., P. Ma, S. Kumar, M. Rocca, J.T. Morisette, C.S. Jarnevich, and N.
Benson. 2010. “Ensemble Habitat Mapping of Invasive Plant Species.” Risk
Analysis 30 (2): 224–35. doi:10.1111/j.1539-6924.2009.01343.x.
Stoms, D.M. 2000. “GAP Management Status and Regional Indicators of Threats to
Biodiversity.” Landscape Ecology 15: 21–33.
Turner, M.G., W.H. Romme, R.A. Reed, and G.A. Tuskan. 2003. “Post-Fire Aspen
Seedling Recruitment across the Yellowstone ( USA ) Landscape.Landscape
Ecology 18: 127–40.
USGS. 2014. “EarthExplorer Archive.” http://edcsns17.cr.usgs.gov/NewEarthExplorer/.
Ustin, S.L., and J.A. Gamon. 2010. “Remote Sensing of Plant Functional Types.” New
Phytologist 186: 795–816. doi:10.1111/j.1469-8137.2010.03284.x.
Vogelmann, J.E., G. Xian, C. Homer, and B. Tolk. 2012. “Monitoring Gradual
Ecosystem Change Using Landsat Time Series Analyses: Case Studies in Selected
Forest and Rangeland Ecosystems.” Remote Sensing of Environment 122 (July).
Elsevier B.V.: 92–105. doi:10.1016/j.rse.2011.06.027.
Worrall, J.J., L. Egeland, T. Eager, R.A. Mask, E.W. Johnson, P.A. Kemp, and W.D.
Shepperd. 2008. “Rapid Mortality of Populus Tremuloides in Southwestern
Colorado, USA.” Forest Ecology and Management 255 (3-4): 686–96.
doi:10.1016/j.foreco.2007.09.071.
Wulder, M.A., J.C. White, B.J. Bentz, M.F. Alvarez, and N.C. Coops. 2006.
“Estimating the Probability of Mountain Pine Beetle Red-Attack Damage.” Remote
Sensing of Environment 101: 150–66.
Wylie, B., M. Rigge, B. Brisco, K. Murnaghan, J. Rover, and J. Long. 2014. “Effects of
Disturbance and Climate Change on Ecosystem Performance in the Yukon River
Basin Boreal Forest.” Remote Sensing 6 (10): 9145–69. doi:10.3390/rs6109145.
Zimmermann, N.E., T.C. Edwards, G.G. Moisen, T.S. Frescino, and J.A. Blackard.
2007. “Remote Sensing-Based Predictors Improve Distribution Models of Rare,
Early Successional and Broadleaf Tree Species in Utah.” The Journal of Applied
Ecology 44 (5): 1057–67. doi:10.1111/j.1365-2664.2007.01348.x.
... Species distribution models (SDMs) relate species distribution data to the environmental characteristics of an area and have been utilized in diverse environments for a wide [18][19][20][21]. SDMs can be used to estimate the potential distribution of a species, then to predict the extent of suitable habitat over a given area [22,23]. ...
... SDMs use environmental data, including both abiotic and biotic variables, to map and predict species occurrence across a landscape. Incorporating remote sensing data into the process, known as predictive habitat distribution modelling [19,26], allows for a more dynamic model because it captures vegetation phenology [27]. Advances in multispectral satellite capabilities provide the opportunity to identify a more complete spectral signature, or response, of a species. ...
... Contrast stretches were applied to the NAIP imagery to highlight the unique texture and pattern of AWC compared to the surrounding vegetation. This photo interpretation technique has been used previously to develop presence/absence data in SDMs [18][19][20][21]. ...
Article
Full-text available
Atlantic White Cedar (Chamaecyparis thyoides) (AWC) anchors a globally threatened ecosystem that is being impacted by climate change, as these trees are vulnerable to hurricane events, sea-level rises, and increasing salinity at the forest–marsh ecotone. In this study, we determined the current amount and distribution of AWC in an area that is experiencing sea-level rises that are higher than the global average rate. We used a combination of a field investigation and aerial photo interpretation to identify known locations of AWC, then integrated Sentinel-1 and 2A satellite data with abiotic variables into a species distribution model. We developed a spectral signature of AWC to aid in our understanding of phenology differences from nearby species groups. The selected model had an out-of-bag error of 7.2%, and 8 of the 11 variables retained in the final model were derived from remotely sensed data, highlighting the importance of including temporal data to exploit divergent phenology. Model predictions were strong in live AWC stands and, accurately, did not predict live AWC in stands that experienced high levels of mortality after Hurricane Sandy. The model presented in this study provides high utility for AWC management and tracking mortality dynamics within stands after disturbances such as hurricanes.
... This challenge is highlighted throughout Paper I and also in Paper II after adding multiple predictors including CLGs. This is in agreement with numerous other studies (Álvarez-Martínez et al. 2018;Assal, Anderson, and Sibold 2015;Chen et al. 2015;Barry and Elith 2006), which express similar requirements. At the time of writing the individual papers, many desired predictor data were unavailable due to inadequate or incomplete nation-wide coverage. ...
... As of today, categorical vegetation maps with types defined 63 by the species composition are largely based on field surveys, such as the British national 64 vegetation survey (Mucina et al. 1993;Rodwell 2018), the EU habitats survey (MNHN-EEA 65 2014), the LUCAS survey (Eurostat 2003) and the Norwegian AR18x18 survey (Strand 66 2013). In contrast to remote sensing-based categories defined by physiognomy and separated 67 by spectral signatures (Assal, Anderson, and Sibold 2015), field-based vegetation-type 68 mapping is time-consuming and expensive . 69 ...
... To further reduce the biases and uncertainties of DMbased PFT profiles, we recommend exploring the use of variables derived from RS directly as predictors in DM. Previous studies have shown that RS-based predictors may enhance DM performance on different scales: on the vegetation-type level (Álvarez-Martínez et al., 2018), on the habitat-type level (Mücher et al., 2009) and on the PFT level (Assal et al., 2015). Further suggestions for improvement of the methods used in this study are found in Table 4. ...
Thesis
Full-text available
Reliable information about the land surface and vegetation is a prerequisite for managing protected areas, maintaining biodiversity or even predicting climate change. Yet, despite the enormous efforts in field-mapping campaigns, the coverage of vegetation maps in Norway is low. To supplement the coverage of vegetation maps, statistical methods were applied to predict the vegetation distribution. Information about environmental conditions regulating the vegetation distributions were extracted in the process. Moreover, the environmental information about the preferences of each vegetation type was used to improve the ecological components of climate models. This doctoral thesis (1) presents detailed predictions of vegetation-type distribution for the whole Norway; (2) shows that large-scale landscape patterns are helpful additional information for establishing good predictions; (3) tests ways of assembling wall-to-wall vegetation maps; and (4) improves how vegetation is represented in climate models. The doctoral work focuses on the Norwegian terrestrial areas, but the methodological findings are applicable to the high-latitude, boreal and alpine ecosystems.
... At local scales, monitoring is critical for managers to establish baseline conditions by which future changes, including change resulting from management actions and natural ecological processes, can be assessed (Manier et al., 2017). Regional land cover metrics and vegetation indices support habitat modeling at broad scales (Razenkova et al., 2020), yet often lack the detailed information needed to support local management needs (Assal et al., 2015). Repeated monitoring over long periods of time makes it possible to detect the direction and magnitude of change in resource condition (Manier et al., 2017). ...
... To create presence/absence data for species distribution modeling, we selected samples across NLCD land cover classes common in the study area: deciduous forest, coniferous forest, herbaceous, mixed forest, shrub/scrub, and woody wetlands/emergent vegetation. We used a stratified random selection to ensure an unbiased distribution of sample plots (30 m × 30 m, congruent with a Landsat pixel) across the land cover types for model input data (Assal et al., 2015). Our intent was to capture fine-scale patterns in the study area, including riparian areas that only occupy a small fraction of the landscape, while minimizing the impact of spatial dependency between observations. ...
... We assembled a suite of topographic and multi-date remotely sensed variables related to vegetation distribution (Assal et al., 2015;Engler et al., 2013;Zimmermann et al., 2007). Topographic variables were derived from the NED and remotely sensed variables were derived from the Landsat Operational Land Imager (OLI) using predominantly cloudfree imagery at three points during the growing season to capture early, high, and late growing season deciduous phenology (Assal et al., 2015). ...
Article
Full-text available
Riparian ecosystems provide critical habitat for many species, yet assessment of vegetation condition at local scales is difficult to measure when considering large areas over long time periods. We present a framework to map and monitor two deciduous cover types, upland and riparian, occupying a small fraction of an expansive, mountainous landscape in north-central Wyoming. Initially, we developed broad-scale predictions of predominant woody vegetation types by integrating Landsat data into species distribution models and combining subsequent outputs into a synthesis map. Then, we evaluated a 35-year Landsat time series (1985–2019) using the Mann-Kendall test to identify significant trends in the condition of upland and riparian deciduous vegetation and assessed the rate and direction of change using the Theil-Sen estimator. Finally, we used plot level data to assess the utility of the framework to detect bottom-up controls (ungulate browse pressure and management actions) on vegetation condition. The synthesis map had an overall correct classification rate of 87% and field data indicated deciduous vegetation within 45 m of coniferous forest faces increased pressure of conifer expansion. The trend assessment identified consistent patterns operating at the landscape scale across both upland and riparian deciduous vegetation; a predominant greening trend was observed for 12 years followed by a 9-year browning trend, before switching back to a greening trend for the last 13 years of the study. Our results indicate trends are driven by the climate of the measurement period at the landscape scale. Although we did not find conclusive evidence to establish a strong link between browse pressure and satellite data, we highlight examples where prevailing trends can be overridden by local disturbance or management intervention. This framework is transferable to other understudied riparian environments throughout western North America to provide insight on ecohydrological processes and assess global and local stressors across broad spatiotemporal scales.
... Then, the atmospheric correction was applied and the atmospheric differences caused by various Pseudo-Invariant Features (PIF) were overcome. The radiometric and atmospheric corrections were conducted employing Landsat and RStool packages available in the R library [41][42][43][44][45]. ...
... The LULC classification was performed in R [41] using Classes and Methods for Spatial Data (Sp), Raster Geospatial Data Abstraction Library (Rgdal), Raster, and Random forests packages [42][43][44][45]. The accuracy of image classification was evaluated by comparing the classified LULC maps with reference Google Earth images from 2013 -2016 of the study area obtained from Google Earth Engine (GEE) platform [27]. ...
... Where, NIR = Near Infrared Band value and R = Red Band value recorded by the Landsat ETM+ imageries [44]. The photosynthesis is the main function of plants, which is directly associated with electromagnetic energy [59,60]. ...
Preprint
Full-text available
Mining for resources extraction may lead to several geological and associated environmental changes due to ground movements, collision with mining cavities and deformation of aquifers. Geological changes may continue in a reclaimed mine area, and the deformed aquifers may entail a breakdown of substrates and an increase in ground water tables, which may cause surface area inundation. Consequently, a reclaimed mine area may experience surface area collapse, i.e. subsidence, and degradation of vegetation health. Thus, monitoring short-term landscape dynamics in a reclaimed mine area may provide important information on the long-term geological and environmental impacts of mining activities. We studied landscape dynamics in Kirchheller Heide, Germany, which experienced extensive soil movement due to longwall mining without stowing, using Landsat imageries between 2013 and 2016. A Random Forest image classification technique was applied to analyse land-use and land-cover dynamics and the growth of wetland areas was assessed using a Spectral Mixture Analysis (SMA). We also analyzed the changes in vegetation health using a Normalized Difference Vegetation Index (NDVI). We observed a 19.9% growth of wetland area within the four years with 87.2% of growth in the coverage of two major waterbodies in the reclaimed mine area. NDVI values indicate that 66.5% of the vegetation of the study area was degraded due to changes in ground water tables and surface flooding. Our results inform environmental management and mining reclamation authorities about the subsidence spots and priority mitigation areas from land surface and vegetation degradation in Kirchheller Heide.
... typical areas) of the study area from 1979 to 2018 acquired from the Google Earth Engine (GEE) platform and national land survey in China(Assal et al., 2015; Padmanaban et al., 2017a;Cao et al., 2018;Luo et al., 2020). We generated a set of 265 random points and extracted those values for four different study periods. ...
Article
Full-text available
The exploitation of mineral resources is crucial for cost-effective construction but has also led to severe damage to the ecological environment in mining areas. Therefore, it is particularly important to effectively monitor surface environmental problems in the mining subsidence area caused by the exploitation of mineral resources. Herein, the Huinong coal mining subsidence area, Shizuishan city, Ningxia, China, was taken as the study area. The remote sensing image features of various environmental elements were investigated through field investigations, the measured spectra, and image spectrum contrast analysis. On this basis, an object-oriented random forest classification method was used to classify images from different time phases and sources in coal mining subsidence areas. Next, the man–machine interactive interpretation was confirmed by referring to the pre-classification results. By overlaying the interpretation result map and analyzing the land-use class changes, the spatial–temporal evolution monitoring of the surface environment in the coal mining subsidence area from 1979 to 2018 was carried out. The results show that the surface environment in the coal mining subsidence area has undergone significant changes over the past 40 years, among which—from 1979 to 2003—the environment of the coal mining area was severely damaged by the intensive mining activities. The area of cultivated land and vegetation coverage decreased sharply, while the area of other land-use classes, such as coal heaps, water bodies, and coal gangue, exhibited a trend of rapid growth. From 2003 to 2018, after more than 10 years of mine geological environment renovation and management, the surface environment of the coal mining subsidence area greatly improved, among which the vegetation coverage has shown the fastest growth rate, while the area of coal gangue, badlands, and other land-use classes have significantly reduced. The hidden dangers of geological disasters have been drastically mitigated. In addition, the residential area continued to decrease in the early stages and then rebounded to a certain extent, indicating that urbanization was carried out at the same time as the ecological environment began to improve. The surface environment before and after the renovation is consistent with the results from remote sensing monitoring.
... Today, categorical vegetation maps with types defined by the species composition are largely based on field surveys, such as the British National Vegetation Classification (NVC) survey (Mucina et al. 1993;Rodwell 2018), the European Vegetation Survey (EVS) (MNHN-EEA 2014), the LUCAS survey (Eurostat 2003), and the Norwegian AR18x18 survey (Strand 2013). In contrast to remote sensing-based categories defined by physiognomy and separated by spectral signatures (Assal, Anderson, and Sibold 2015), field-based vegetationtype mapping is time-consuming and expensive (Ullerud, Bryn, and Klanderud 2016). Accordingly, only a fraction of the Earth's surface has been mapped by field survey methods (Alexander and Millington 2000), and alternative and more cost-efficient ways to produce wall-to-wall maps of vegetation types are needed to cover large areas. ...
Article
Full-text available
Distribution modeling methods are used to provide occurrence probability surfaces for modeled targets. While most often used for modeling species, distribution modeling methods can also be applied to vegetation types. However, surfaces provided by distribution modeling need to be transformed into classified wall-to-wall maps of vegetation types to be useful for practical purposes , such as nature management and environmental planning. The paper compares the performance of three methods for assembling predictions for multiple vegetation types, modeled individually, into a wall-to-wall map. The authors used grid-cell based probability surfaces from distribution models of 31 vegetation types to test the three assembly methods. The first, a probability-based method, selected for each grid cell the vegetation type with the highest predicted probability of occurrence in that cell. The second, a performance-based method, assigned the vegetation types, ordered from high to low model performance, to a fraction of the grid cells given by the vegetation type's prevalence in the study area. The third, a prevalence-based method, differed from the performance-based method by assigning vegetation types in the order from low to high prevalence. Thus the assembly methods worked in two principally different ways: the probability-based method assigned vegetation types to grid cells in a cell-by-cell manner, and both the performance-based method and prevalence-based method assigned them in a type-by-type manner. All methods were evaluated by use of reference data collected in the field, more or less independently of the data used to parameterize the vegetation-type models. Quantity, allocation, and total disagreement, as well as proportional dissimilarity metrics, were used for evaluation of assembly methods. Overlay analysis showed 38.1% agreement between all three assembly methods. The probability-based method had the lowest total disagreement with, and proportional dissimilarity from, the reference datasets, but the differences between the three methods were small. The three assembly methods differed strongly with respect to the distribution of the total disagreement on its quantity and allocation components: the cell-by-cell assignment method strongly favored allocation disagreement and the type-by-type methods strongly favored quantity disagreement. The probability-based method best reproduced the general pattern of variation across the study area, but at the cost of many rare vegetation types, which were left out of the assembled map. By contrast, the prevalence-based and performance-based methods represented vegetation types in accordance with nationwide area statistics. The results show that maps of vegetation types with wall-to-wall coverage can be assembled from individual distribution models with a quality acceptable for indicative purposes, but all the three tested methods currently also have shortcomings. The results also indicate specific points in the methodology for map assembly that may be improved. ARTICLE HISTORY
... The accuracy assessment is a widely applied technique to quantify how close the result of land cover classification is to the reference image. In this study, the accuracy was assessed by comparing the results of the obtained LULC map (classified image) with reference Google Earth images retrieved from the Google Earth Engine (GEE) [54] for the year 2019. From the classified LULC, seventy-five random point samples were produced for the study period. ...
Article
Full-text available
Citation: Charrua, A.B.; Padmanaban, R.; Cabral, P.; Bandeira, S.; Romeiras, M.M. Impacts of the Tropical Cyclone Idai in Mozambique: A Multi-Temporal Landsat Satellite Imagery Analysis. Remote Sens. 2021, 13, 201. https://doi.
... To further reduce the biases and uncertainties of DMbased PFT profiles, we recommend exploring the use of variables derived from RS directly as predictors in DM. Previous studies have shown that RS-based predictors may enhance DM performance on different scales: on the vegetation-type level (Álvarez-Martínez et al., 2018), on the habitat-type level (Mücher et al., 2009) and on the PFT level (Assal et al., 2015). Further suggestions for improvement of the methods used in this study are found in Table 4. ...
Article
Full-text available
Vegetation is an important component in global ecosystems, affecting the physical, hydrological and biogeochemical properties of the land surface. Accordingly, the way vegetation is parameterized strongly influences predictions of future climate by Earth system models. To capture future spatial and temporal changes in vegetation cover and its feedbacks to the climate system, dynamic global vegetation models (DGVMs) are included as important components of land surface models. Variation in the predicted vegetation cover from DGVMs therefore has large impacts on modelled radiative and non-radiative properties, especially over high-latitude regions. DGVMs are mostly evaluated by remotely sensed products and less often by other vegetation products or by in situ field observations. In this study, we evaluate the performance of three methods for spatial representation of present-day vegetation cover with respect to prediction of plant functional type (PFT) profiles – one based upon distribution models (DMs), one that uses a remote sensing (RS) dataset and a DGVM (CLM4.5BGCDV; Community Land Model 4.5 Bio-Geo-Chemical cycles and Dynamical Vegetation). While DGVMs predict PFT profiles based on physiological and ecological processes, a DM relies on statistical correlations between a set of predictors and the modelled target, and the RS dataset is based on classification of spectral reflectance patterns of satellite images. PFT profiles obtained from an independently collected field-based vegetation dataset from Norway were used for the evaluation. We found that RS-based PFT profiles matched the reference dataset best, closely followed by DM, whereas predictions from DGVMs often deviated strongly from the reference. DGVM predictions overestimated the area covered by boreal needleleaf evergreen trees and bare ground at the expense of boreal broadleaf deciduous trees and shrubs. Based on environmental predictors identified by DM as important, three new environmental variables (e.g. minimum temperature in May, snow water equivalent in October and precipitation seasonality) were selected as the threshold for the establishment of these high-latitude PFTs. We performed a series of sensitivity experiments to investigate if these thresholds improve the performance of the DGVM method. Based on our results, we suggest implementation of one of these novel PFT-specific thresholds (i.e. precipitation seasonality) in the DGVM method. The results highlight the potential of using PFT-specific thresholds obtained by DM in development of DGVMs in broader regions. Also, we emphasize the potential of establishing DMs as a reliable method for providing PFT distributions for evaluation of DGVMs alongside RS.
Article
Accurate information on the distribution of vegetation species is used as a proxy for the health of an ecosystem, a currency of international environmental treaties, and a necessary planning tool for forest preservation and rehabilitation, to name just a few of its applications. However, direct, extensive observation of vegetation across large geographic regions can be very expensive. The extensive coverage and high temporal resolution of remote sensing data collected by satellites like the European Space Agency’s Sentinel-2 system could be a critical component of a solution to this problem. We propose a hierarchical model for predicting vegetation cover that incorporates high resolution satellite imagery, landscape characteristics such as elevation and slope, and direct observation of vegetation cover. Besides providing model-based predictions of vegetation cover with accompanying uncertainty quantification, our proposed model offers inference about the effects of landscape characteristics on vegetation type. Implementation of the model is computationally challenging due to the volume and spatial extent of data involved. Thus, we propose an efficient, approximate method for model fitting that is able to make use of all available observations. We demonstrate our approach with an application to the distribution of three post-fire resprouting deciduous species in the Jemez Mountains of New Mexico.Supplementary materials accompanying this paper appear on-line.
Preprint
Full-text available
Vegetation is an important component in global ecosystems, affecting the physical, hydrological and biogeochemical properties of the land surface. Accordingly, the way vegetation is parameterised strongly influences predictions of future climate by Earth system models. To capture future spatial and temporal changes in vegetation cover and its feedbacks to the climate system, dynamic global vegetation models (DGVM) are included as important components of land surface models. Variation in the predicted vegetation cover from DGVMs therefore has large impacts on modelled radiative and non-radiative properties, especially over high-latitude regions. DGVMs are mostly evaluated by remotely sensed products, but rarely by other vegetation products or by in-situ field observations. In this study, we evaluate the performance of three methods for spatial representation of vegetation cover with respect to prediction of plant functional type (PFT) profiles – one based upon distribution models (DM), one that uses a remote sensing (RS) dataset and a DGVM (CLM4.5BGCDV). PFT profiles obtained from an independently collected vegetation data set from Norway were used for the evaluation. We found that RS-based PFT profiles matched the reference dataset best, closely followed by DM, whereas predictions from DGVM often deviated strongly from the reference. DGVM predictions overestimated the area covered by boreal needleleaf evergreen trees and bare ground at the expense of boreal broadleaf deciduous trees and shrubs. Based on environmental predictors identified by DM as important, we suggest implementation of three novel PFT-specific thresholds for establishment in the DGVM. We performed a series of sensitivity experiments to demonstrate that these thresholds improve the performance of the DGVM. The results highlight the potential of using PFT-specific thresholds obtained by DM in development and benchmarking of DGVMs for broader regions. Also, we emphasize the potential of establishing DM as a reliable method for providing PFT distributions for evaluation of DGVMs alongside RS.
Article
Full-text available
A warming climate influences boreal forest productivity, dynamics, and disturbance regimes. We used ecosystem models and 250 m satellite Normalized Difference Vegetation Index (NDVI) data averaged over the growing season (GSN) to model current, and estimate future, ecosystem performance. We modeled Expected Ecosystem Performance (EEP), or anticipated productivity, in undisturbed stands over the 2000–2008 period from a variety of abiotic data sources, using a rule-based piecewise regression tree. The EEP model was applied to a future climate ensemble A1B projection to quantify expected changes to mature boreal forest performance. Ecosystem Performance Anomalies (EPA), were identified as the residuals of the EEP and GSN relationship and represent performance departures from expected performance conditions. These performance data were used to monitor successional events following fire. Results suggested that maximum EPA occurs 30–40 years following fire, and deciduous stands generally have higher EPA than coniferous stands. Mean undisturbed EEP is projected to increase 5.6% by 2040 and 8.7% by 2070, suggesting an increased deciduous component in boreal forests. Our results contribute to the understanding of boreal forest successional dynamics and its response to climate change. This information enables informed decisions to prepare for, and adapt to, climate change in the Yukon River Basin forest.
Article
Full-text available
James J. a , Corresponding Author Contact Information , E-mail The Corresponding Author , Leanne Egeland a , Thomas Eager a , Roy A Concentrated patches of recent trembling (Populus tremuloides) mortality covered 56,091 ha of Colorado forests in 2006.
Article
Estimation of potential annual direct incident radiation has traditionally required numerical integration with simulation models. As an alternative, we present convenient equations for use in spreadsheet, GIS, and database applications. Input variables are latitude, slope, and aspect. The equations apply to 0-60degrees north latitude, slopes from 0-90degrees, and all aspects. By transforming aspect, the equations can also be applied as an index of heat load, symmetrical about a northeast to southwest axis.
Article
Abstract Aim Species distribution models have often been hampered by poor local species data, reliance on coarse-scale climate predictors and the assumption that species–environment relationships, even with non-proximate predictors, are consistent across geographical space. Yet locally accurate maps of invasive species, such as the Africanized honeybee (AHB) in North America, are needed to support conservation efforts. Current AHB range maps are relatively coarse and are inconsistent with observed data. Our aim was to improve distribution maps using more proximate predictors (phenology) and using regional models rather than one across the entire range of interest to explore potential differences in drivers. Location United States of America. Methods We provide a generalized framework for regional and local species distribution modelling with our more nuanced and spatially detailed forecast of potential AHB spread using multiple habitat modelling techniques and newly derived remotely sensed phenology layers. Results Variable importance did differ between the two regions for which we modelled AHB. Phenology metrics were important, especially in the south-east. Main conclusions Results demonstrate that incorporating a combination of both climate drivers and vegetation phenology information into models can be important for predicting the suitable habitat range of these pollinators. Regional models may provide evidence of differing drivers of distributions geographically. This framework may improve many local and regional species distribution modelling efforts.
Article
The ability to map vegetation and in particular individual trees is a key component in forest management and long-term forest monitoring. Here we present a novel approach for mapping individual tree species based on ensemble modeling, i.e. combining the projections of several modeling techniques in order to reduce uncertainty. Using statistical modeling in conjunction with high-resolution aerial imagery (50 cm spatial resolution) and topo-climatic variables (5 m spatial resolution), we map the distributions of six major tree species (3 broadleaf and 3 conifers) in a study area of North-Eastern Switzerland. We also compare the relative predictive power of both topo-climatic and remote-sensing variables for map-ping the spatial tree patterns and assess the importance of calibration data quality on model perfor-mance. We evaluate our projections using cross-validation as well as with independent data. Overall, the evaluations that we obtain for our vegetation maps are in line with, or higher than, those in similar studies. Depending on the considered tree species, 47.8–85.6% of our samples were correctly predicted, and we obtain an overall CCR (correct classification rate) of 0.72 and a Cohen's kappa of 0.65. Comparing the predictive power of the different modeling techniques, we find that ensemble modeling (i.e. combin-ing the projections of different individual modeling techniques) generally performs better than individual modeling techniques.