variable code description - lemma
TRANSCRIPT
Spatially explicit information regarding a wide range of forest attributes is often required for land
management decision-support, inventory and monitoring, and policy formulation. Nearest neighbor (NN)
imputation maps are widely used in the Pacific Northwest region of the US, when spatially explicit
information regarding a wide range of forest attributes are required at regional and watershed scales.
However, regional NN imputation maps have generally not been suitable for fine-scale (i.e. stand level)
decision-making for activities such as fuel reduction and forest restoration treatments, which would benefit
greatly from NN imputation maps developed for both regional and stand-level applications.
NN imputation maps are predicated on relating forest attributes from field plots to spatial predictors such as
climate, topography, satellite imagery, and LiDAR data. Potential sources of error in NN imputation maps
include the incorrect extraction of remotely sensed spatial predictors for a given field plot, which can result
from poor positional accuracy of field plots or poor geo-referencing of spatial predictors such as satellite
imagery. The extent to which positional accuracy of these two data sources (plots and satellite imagery)
influences NN imputation map accuracy is poorly understood, but is increasingly important if NN
imputation maps are developed for stand level applications. To address these questions we wanted to:
1. Quantify the positional error of federal forest inventory plots and Landsat TM imagery within a 500,000
ha area of Eastern Oregon Cascades, USA.
2. Quantify individual and combined influences of positional error from plot locations and Landsat TM
imagery on NN imputation map accuracy for a variety of forest composition and structural attributes.
3. Determine if the influence of field plot and Landsat TM positional error on NN imputation accuracy is
greater when NN imputation maps are developed using field data of smaller spatial grain.
Influence of inventory plot and Landsat imagery positional accuracies on nearest-neighbor (NN) imputation maps
of vegetation composition and structure. Harold Zald, College of Forestry, Oregon State University, Corvallis, OR USA
Janet Ohmann, Pacific Northwest Research Station, USDA Forest Service, Corvallis, OR USA
Heather Roberts, College of Forestry, Oregon State University, Corvallis, OR USA
Robert McGaughey, Pacific Northwest Research Station, USDA Forest Service, Seattle, WA USA
Matt Gregory, College of Forestry, Oregon State University, Corvallis, OR USA
Robert Kennedy, Department of Earth and Environment, Boston University, Boston, MA USA
Introduction
Objectives
Study Area
Plot & Imagery Accuracy Methods
Plot & Imagery Positional Accuracy
Results: Species Composition
Conclusions
Acknowledgements
References
Data Sources
The study area encompasses 539,269 ha (1,332,561 ac) of the
Eastern Oregon Cascades, covering most of the Deschutes
National Forest (gray area in map).
Large-scale gradients of topography, climate, and parent material
influence vegetation composition and productivity in the study
area. Western proportions of the study area are characterized by
high elevations, high precipitation, and moderated temperature
ranges; resulting in montane and subalpine conifer forests. The
eastern portion of the study area is lower in elevation, and has a
more arid and continental climate; resulting in lower productivity
drier conifer forests.
In contrast to large-scale species composition gradients, variability
in forest structure is more strongly influenced by smaller-scale
disturbances such as wildfire, insect outbreaks, and timber
harvests.
232 plots from the USDA Forest Service Forest
Inventory and Analysis Program (FIA) and The
USDA Forest Service Region 6 Current Vegetation
Survey (R6). Plots measured between 2004-2009.
Vegetation attributes (% cover, DBH, TPH, Snags,
CWD, etc.) calculated at plot-level and subplot1.
Official coordinates collected using recreational
grade GPS at time of field measurement. Sub-meter
coordinates collected using a Trimble WAAS
enabled and post-processed DGPS during 2009-
2010.
Spatial predictors are grids with 30m cell.
Climate data comes from PRISM (Daly et al. 2008)
30 year normals (1971-2001).
Topography and LiDAR vegetation structure from
LiDAR data collected in 2009 and 2010.
Topographic spatial predictors calculated from
LiDAR 30m bare earth model. Vegetation structure
spatial predictors calculated from LiDAR point
cloud using the grid metrics function in FUSION.
Choice of LiDAR-derived vegetation predictors
based on prior work relating LiDAR metrics to
forest structure in the Pacific Northwest region
(Hudak et al. 2008, Falkowski et al. 2010, Kane et
al. 2010)
Landsat TM images developed using the
LandTrendr algorithms (Kennedy et al. 2007,
2010), a trajectory-based change detection method
using a time-series of yearly Landsat TM images
that are cloud-free, geometrically corrected, and
radiometrically normalized.
Spatial predictors from LandTrendr include: fitted
tasseled cap brightness, greenness, and wetness
from 2004-2009 matching plot measurement years;
and 1984-2008 magnitude, duration, and years
since disturbance.
Subplot 1
FIA and R6
R6 Subplots
(7.32m radius)
FIA Subplots
(7.32m radius)
30m Landsat cell
Daly C., Halbleib M., Smith JI, Gibson WP, Doggett MK, Taylor GH, Curtis J, and PA Pasteris. 2008. Physiographically-sensitive mapping of temperature and
precipitation across the conterminous United States. International Journal of Climatology, 28: 2031-2064.
Falkowski MJ, Hudak AT, Crookston NL, Gessler PE, Uebler EH, and AMS Smith. 2010. Landscape-scale parameterization of a tree-level forest growth model: a k-
nearest neighbor imputation approach incorporating LiDAR data. Canadian Journal of Forest Research, 40: 184-199.
Hudak AT, Crookston NL, Evans JS, Hall DE, and MJ Falkowski. 2008. Nearest neighbor imputation of species-level, plot-scale structure attributes from LiDAR
data. Remote Sensing of Environment, 112: 2232-2245.
Kane VR, McGaughey RJ, Bakker JD, Gersonde RF, Lutz JA, and JF Franklin. 2010. Comparisons between field- and LiDAR-based measures of stand structural
complexity. Canadian Journal of Forest Research, 40: 761-773.
Kennedy RE, and WB Cohen. 2003. Automated designation of tie-points for image-to-image coregistration. International Journal of Remote Sensing, 24: 3467-
3490.
Kennedy RE, Cohen WB, and TA Schroeder. 2007. Trajectory-based change detection for automated characterization of forest disturbance dynamics. Remote
Sensing of Environment, 110: 370-386.
Kennedy RE, Yang Z, and WB Cohen. 2010. Detecting trends in forest disturbance and recover using yearly Landsat times series: 1. LandTrendr – Temporal
segmentation algorithms. Remote Sensing of Environment, 114: 2897-2810.
Ohmann, JL, MJ Gregory. 2002. Predictive mapping of forest composition and structure with direct gradient analysis and nearest-neighbor imputation in coastal
Oregon, USA. Canadian Journal of Forest Research, 32:725-741.
Pierce KB, Lookingbill T, and D Urban. 2005. A simple method for estimating potential relative radiation (PRR) for landscape-vegetation analysis. Landscape
Ecology, 20: 137-147.
Riemann R, Wilson BT, Lister A, and S Parks. 2010. An effective assessment protocol for continuous geospatial datasets of forest characteristics using USFS Forest
Inventory and Analysis (FIA) data. Remote Sensing of Environment, 114: 2337-2352
NN Imputation Methods
Official plot coordinates compared to sub-meter plot coordinates collected by
sub-meter GPS RMSE and distribution positional errors calculated.
2010 Landsat TM tasseled cap brightness image was referenced to a LiDAR-
derived canopy cover above 2m tall using ITPFind (Kennedy and Cohen
2003), an automated area-based method for identifying image tie-points.
Target and referenced image tie-point locations were compared to quantify
target image positional accuracy, Landsat TM geo-rectified using a 1st order
polynomial and nearest neighbor resampling in ERDAS IMAGINE.
Gradient Nearest Neighbor (GNN) method
of imputation mapping (Ohmann and
Gregory 2002), in which relationships
between forest attributes and spatial
predictors is quantified via Canonical
Correspondence Analysis
8 models comparing spatial grain of field
data, positional accuracy of plots and
imagery:
Observed versus predicted compared for a 3 x 3 pixel block
(plot-level), and single pixel (subplot 1 – level).
Accuracy of species predictions assessed with Kappa
statistic. Accuracy of structure predictions assessed using
selected protocols outlined by Riemann et al. (2010) at plot
and subplot scales:
1. Kolmogorov-Smirnov (KS) statistic calculated to quantify
maximum distance between observed and predicted
empirical cumulative distribution functions (ecdf)
2. Systematic difference (ACSYS) and unsystematic
difference (ACUNS) calculated to compare estimated bias
and random error of predictions between models.
KS, ACSYS, and ACUNS values presented are averages for the
following 7 selected structural attributes:
• Density shade intolerant species (TPHINTOL_ GE_3)
• Density shade tolerant species (TPHTOL_ GE_3)
• Density small trees 3-25 cm dbh (TPH_3_25)
• Density large conifers ≥ 75 cm dbh (TPHC_GE_75)
• Density of snags ≥ 25 cm dbh (STPH_GE_25)
• Density large snags ≥ 75 cm dbh (STPH_GE_75)
• Down wood volume ≥ 12 cm dia (DVPH_GE_12)
1:1 line
(ACSYS)
Systematic difference
measure of
“estimated bias”
(ACUNS)
Unsystematic difference
measure of random error
Geometric mean
functional relationship
(GMFR)
Observed values
Pre
dic
ted v
alues
Sub-meter DGPS accuracy for the 232 plots was 0.91m (95% CI
0.81 – 1.01m).
On average, accuracy of old official plot coordinates was poor
(RMSE_total = 24.69m)
39%, 12%, and 4% of official plot coordinates had positional error
greater than 15, 30, and 60 m respectively
Based on over 27,000 RTK survey points, LiDAR positional
accuracy was very high (RSME total = 0.04m)
On average, positional accuracy of 358 tie points was very good
(RSME = 7.6m).
9% of image tie-points had positional error greater than ½ width of
a Landsat TM pixel (> 15m), No image tie-points had positional
error greater 30m.
From plot and image positional accuracy results above, we would
expect plot positional errors to be more important for NN map
accuracy than positional errors in the Landsat TM imagery.
Model Field Data Plot Positional Accuracy Imagery Positional Accuracy
1 Plt-OldGPS-OldGeoRef Plot-level Old Official GPS Coordinates Old Original Geo-referenced
2 Plt-SubMeterGPS-OldGeoRef Plot-level New Sub-meter DGPS New Geo-rectified against LiDAR
3 Plt-OldGPS-NewGeoRef Plot-level Old Official GPS Coordinates Old Original Geo-referenced
4 Plt-SubMeterGPS-NewGeoRef Plot-level New Sub-meter DGPS New Geo-rectified against LiDAR
5 Subplt1-OldGPS-OldGeoRef Subplot 1 Old Official GPS Coordinates Old Original Geo-referenced
6 Subplt1-SubMeterGPS-OldGeoRef Subplot 1 New Sub-meter DGPS New Geo-rectified against LiDAR
7 Subplt1-OldGPS-NewGeoRef Subplot 1 Old Official GPS Coordinates Old Original Geo-referenced
8 Subplt1-SubMeterGPS-NewGeoRef Subplot 1 New Sub-meter DGPS New Geo-rectified against LiDAR
Important to note only a limited number and type of accuracy diagnostics used. For example, Kappa
statistic does not provide components of map error needed to assess potential limitations.
For this data set, finer spatial grain field data and increased positional accuracy of spatial
predictors did not improve accuracy of NN maps of species composition and structure.
Despite plot (and to a lesser extent Landsat TM imagery) positional errors that often were ≥ ½
pixel width of spatial predictors, positional accuracy consistently did not influence NN prediction
accuracy
Potential explanations for these findings include:
Poor sample size (232 plots)
Hybrid species-structure response matrix used in gradient analysis may confound relationships
between field data and spatial predictors operating separately for composition and structure at
large (climate, topography) and small (LiDAR, forest disturbance history) spatial scales.
Tuning separate models for composition and structure may be needed to maximize predictive
accuracy.
Higher positional accuracy may not matter if within context of forest spatial heterogeneity
operating at larger scales (i.e. larger gradients and patch sizes of species and structure).
Subplot may be too small of a sample unit, resulting in high measurement variability and
inadequate species/structural representation for relating to spatial predictors
Accuracy of species predictions varied by species, spatial grain of field plots, positional accuracy of
plots, and positional accuracy of Landsat TM imagery. Accuracy was especially poor for PIMO, which
had the lowest prevalence of the selected species.
In general, accuracy of species predictions was greater at larger spatial grain (i.e. full plots vs.
subplots). Co-mission errors were greater for predictions based on full plots.
For many species, prediction accuracy either declines or is not significantly different with increased
positional accuracy of plots or Landsat TM imagery. likely the result of fine-scale spatial predictors
being over-weighted in gradient analysis when species composition is more strongly controlled by
large-scale climatic and topography gradients.
Funding provided by the USDA Forest Service Pacific Northwest Research Station, and the Western Wildland Environmental Threat Assessment Center. Special
thanks to Mike Simpson and staff at the Deschutes National for providing sub-meter DGPS coordinates for FIA and R6 plots, Justin Braaten and Zhiqiang Yang
for assistance with LandTrendr fitted tassel cap imagery and disturbance maps, and Emilie Henderson with R code for GNN and GNN map accuracy assessments.
Observed
data ecdf F
Attribute values
Per
cent
of
dat
aset
0 1
00
Predicted
data ecdf G
Results: Vegetation Structure
Accuracy of spatial predictions calculated by comparing observed and predicted forest attributes using
a modified leave-one-out cross validation approach.
Mod1
Mod2
Mod3
Mod4
Mod5
Mod6
Mod7
Mod8
Predicted distribution of Ponderosa pine (PIPO)
Plt-OldGPS-OldGeoRef Subplt1-OldGPS-OldGeoRef
PIPO present PIPO absent
Open water Ice fields
Predicted density of small trees (TPH_3_25)
Plt-NewGPS-NewGeoRef Subplt1-NewGPS-NewGeoRef
Open water Ice fields
Models with larger spatial grain (i.e. field plots) have smaller maximum distances between modeled
and reference ecdfs (lower KS values), and have higher unsystematic agreement (ACUNS colored
triangles). There was little difference in systematic (ACUNS colored squares) between models.
Variable
subset
Code Description
Climate ANNPRE Mean annual precipiation (natural logarithm, mm)
ANNTMP Mean annual temperature (°C)
AUGMAXT Mean maximum temperature of hottest month (August) (°C)
CONTPRE Precentage of annual precipitation falling during the growing
season (June - August)
DECMINT Mean minimum temperature of coldest month (December)
(°C)
SMRTP Growing season moisture stress, the ratio of mean temperature
(°C) to precipiation (natural logarithm, mm), May -
September
Topography ELEV Elevation (m)
ASP Cosine transformation of aspects (degrees)
PRR Cumulative potention relative radiation during the growing
season (Pierce et al. 2005)
SLOPE Slope (%)
TPI150 Topographic position index, calculated as the difference
between a cell's elevation and the mean elevation of cells
within a 150-m radius window
TPI450 Topographic position index, calculated as above but with a
450-m radius window
Parent Material ASH Total ash deposition (feet), primarily from Mt. Mazama in the
eastern Cascades (unpubl. data from Mike Simpson)
GLACIAL Ice transported deposits (categorical)
PYROCLST Pyroclastic deposits (categorical)
PUMICE Pumice desposits (categorical)
Landsat TM TC1 Brightness axis from tasseled cap transformation
TC2 Greenness axis from tasseled cap transformation
TC3 Wetness axis from tasseled cap transformation
YSD Years since disturbance, from multitemporal Landat TM
analysis (Kennedy and Cohen 2011)
MAG Magnitude of disturbance (change in canopy cover) (%), from
multitemporal Landsat TM analysis (Kennedy and Cohen
2011)
DUR Duration of disturbance (years), from multitemporal Landsat
TM analysis (Kennedy and Cohen)
LiDAR VEGP95 Height (m) of 95th percentile of vegetation height returns
VEGMEAN Height (m) of mean of vegetation height returns
COVER Canopy cover above 2-m (percent), calculated as the
proportion of first returns greater than a lower height limit of
2-m above ground in the digital terrian model
VEGSD Standard deviation of vegetation height (m)
Location EAST Universal Transverse Mercator easting (m)
NORTH Universal Transverse Mercator northing (m)
LandTrendr method
D
DKS = max|F(x) – G(x)|
Image tie-points
Plot and Subplot Footprints
List of Spatial Predictors