International Journal of Remote Sensing
Vol. 27, No. 21, 10 November 2006, 4731–4749
Estimating area errors for fine-scale feature-based ecological mapping
E. C. ELLIS*{ and H. WANG{{
{Department of Geography and Environmental Systems, University of Maryland,
Baltimore County, 1000 Hilltop Circle, Baltimore, MD 21250 USA
{Current address: Environmental Sciences institute, Florida A&M University Science
Research Center 305 D, 1520 S. Bronough St., Tallahassee, FL 32307 USA
(Received 15 July 2005; in final form 29 March 2006 )
High spatial resolution feature-based approaches are especially useful for
ecological mapping in densely populated landscapes. This paper evaluates errors
in estimating ecological map class areas from fine-scale current (,2002) and
historical (,1945) feature-based ecological mapping by a set of trained
interpreters across densely populated rural sites in China based on field-validated
interpretation of high spatial resolution ((1 m) imagery. Median overall map
accuracy, corrected for chance, was greater than 85% for mapping by trained
interpreters, with greater accuracy for current versus historical mapping. An
error model based on feature perimeter proved as reliable in predicting 90%
confidence intervals for map class areas as did models derived from the
conventional error matrix. A conservative error model combining these
approaches was developed and tested for statistical reliability in predicting
confidence intervals for ecological map class areas from fine-scale feature-based
mapping by a set of trained interpreters across rural China, providing a practical
basis for statistically reliable ecological change detection in densely populated
landscapes.
1.
Introduction
Long-term ecological changes in densely populated landscapes are responsible for a
growing share of global and regional environmental change (Foley et al. 2003,
DeFries et al. 2004, Ellis 2004). Although changes in resource management and
ecosystem processes are a critical part of this (Vitousek et al. 1997, Houghton 2003),
long-term ecological changes in densely populated landscapes are dominated by
fine-scale changes in landscape structure (,30 m) caused by the creation,
transformation, and abandonment of anthropogenic features with distinct
boundaries, such as buildings, roads and small agricultural plots (Forster 1985,
Lo and Shipman 1990, Ellis et al. 2000b). For this reason, fine-scale feature-based
mapping is an especially useful approach in measuring long-term ecological changes
within densely populated landscapes (Ellis et al. 2000b, August et al. 2002, Thomas
et al. 2003, Ellis 2004, Lu et al. 2004, Ellis et al. 2006).
Feature-based environmental mapping has a long history rooted in aerial
photograph interpretation (Troll 1939) and this approach is expanding in response
to the wide availability of high spatial resolution satellite imagery and recent
advances in geographic information systems (GIS), image segmentation and
*Corresponding author. Email: ece@umbc.edu
International Journal of Remote Sensing
ISSN 0143-1161 print/ISSN 1366-5901 online # 2006 Taylor & Francis
http://www.tandf.co.uk/journals
DOI: 10.1080/01431160600735632
4732
E. C. Ellis and H. Wang
automated feature-extraction (Jensen and Cowen 1999, Hill et al. 2002, Thomas et
al. 2003, Lu et al. 2004, Wang and Ellis 2005b). Even with these advances, fine-scale
feature-based mapping remains a relatively resource-intensive process, as are
environmental sampling and land management surveying, and all of these methods
combined together are most powerful when measuring long-term ecological changes
within intensively managed landscapes (Ellis et al. 2000a, Johnes and Butterfield
2002, Ellis 2004). Therefore, regional scale application of these methods is usually
based on regionally stratified sampling designs, such as area frame sampling, that
can limit fine-scale mapping and other resource-intensive methods to a smaller set of
regionally-representative sample strata or sample cells (Gallego et al. 1994, Gallego
2000, Ellis 2004). By this approach, fine-scale measurements within a set of sample
cells, such as crop area or fertilizer use, can be used to make regional estimates of
these parameters by adjusting for sample cell relationships with regional data
(Gallego et al. 1994, Gallego 2000, Ellis 2004).
Fine-scale ecological changes within sample cells can be mapped by comparing
current and historical feature-based maps using GIS, as long as maps are prepared
to the same standards and the spatial scale of mapped features is substantially
greater than the scale of misregistration errors between maps (Verbyla and Boles
2000, Lu et al. 2004, Wang and Ellis 2005a). When historical imagery must be geocorrected with limited information however, misregistration errors may reach 10 m
or more, causing ‘false-change’ errors when smaller features are misaligned between
maps, potentially obscuring ecologically significant changes (Townshend et al. 1992,
Wang and Ellis 2005b, 2005a). Fortunately, false change errors caused by
misregistration are not a significant source of error in map class area estimates
for entire sample cells, if sample cells are large enough (Wang and Ellis 2005a). For
this reason, fine-scale ecological changes across mapped sample cells can be detected
by testing whether area changes are greater than errors in estimating map class areas
within each time period (Lu et al. 2004).
Conventional error-matrix approaches for estimating errors in map class areas are
well established (Congalton and Mead 1983, Story and Congalton 1986, Congalton
and Green 1999, Foody 2002) as are methods specific to feature-based mapping
(Green and Hartley 2000). In the present study, we apply both error matrix and
feature-based approaches to estimate errors in map class areas from current (,2002)
and historical (,1945) ecological feature-based maps by trained interpreters across
China’s densely populated rural landscapes as part of a regional study of long-term
ecological change (Ellis 2004).
Disagreement between trained interpreters is a major source of error in thematic
maps prepared by image interpretation, even when maps are field validated (Cherrill
and McClean 1999, Green and Hartley 2000, Powell et al. 2004). This study assesses
this source of error by comparing maps prepared by a set of trained interpreters
across a sample cell representing the most challenging mapping conditions across
the sites in our study (Ellis 2004). Although errors in feature-based mapping include
position, shape and classification (Green and Hartley 2000), analysis of ecotope
maps (table 4 in Wang and Ellis 2005a) demonstrated that positional error is a
negligible source of error in map class areas estimated as a percentage of entire
500 m6500 m sample cells, given that false change error was near zero for
cells .4636463 m at the maximum coregistration error and number of map classes
observed per cell (15 m, assuming maximum opposing errors at both times; 38
classes, in Jintang; Wang and Ellis 2005b). We therefore focused our analysis on
Estimating feature-based area error
4733
errors in map class areas caused by disagreement in feature mapping (shape-related)
and feature classification between trained interpreters.
The goal of this assessment was to develop a model for predicting statistically
reliable error intervals for map class area estimates from sample cell maps across the
set of trained interpreters, with statistically reliable error intervals defined as those
enclosing the correct value in at least 90% of estimates (Schenker and Gentleman
2001). Based on this definition, models for predicting map class area errors were
developed and tested for their statistical reliability and overall error rate at the site
of model development and at an independent check site, to establish a general model
for predicting errors in class area estimates across interpreters.
2.
2.1
Methods
Study sites
A set of 60 500 m6500 m square sample cells was selected for ecological mapping
and other measurements by a regional sampling design, with 12 cells allocated
within each of five 100 km2 field sites representing environmentally distinct regions
across densely populated rural China (Ellis 2004). Preliminary mapping across sites
confirmed that our site in densely populated eastern Jintang County, Sichuan
Province in the Sichuan Basin Hilly Region of China was the most challenging for
mapping based on its greater complexity of landscape management and terrain and
lowest image quality across sites in China (Ellis 2004). The region has a subtropical
monsoon climate and intensively managed hilly terrain composed of semi-natural
bench plateaus with upland agriculture divided by steep slopes (both vegetated and
barren), with rice paddies and small ponds ringed by housing and narrow roads in
the elongated, gradually terraced, valleys in between. Sample cells at this site had the
largest number of map classes, the most difficult terrain, dispersed built-up features
obscured by tree and bamboo cover, complex patterns of land and vegetation
management and relatively poor quality historical aerial imagery (Ellis 2004, Wang
and Ellis 2005b).
We therefore selected a sample cell representing these challenging conditions at
the Jintang site for detailed investigation of interpreter error (centre5104.770uE,
30.549uN). This cell included more than 90% of the feature classes observed across
the 12 sample cells mapped at this site, and ranged in elevation from 396 to 433 m
with a population of ,800 persons km22 in 2002 and ,530 persons km22 in ,1945.
A check site in southern Yiyang County, Hunan Province, China was selected for
error model testing (400 m6400 m cell, centre5112.450uE, 28.385uN). This cell had
a similar climate to the Jintang site, but with the lower population density (2002
,470, 1945 ,180 persons km22) and less intensive land management typical of
China’s Subtropical Hilly Region, with gentle hillslopes covered by forestry and
perennial crops (tea) with gently terraced rice paddies ringed by houses and roads in
between (Ellis 2004).
2.2
Imagery
IKONOS 4 band pan-sharpened 1 m resolution GEO imagery (www.spaceimaging.com) was acquired 22 December 2001 across a 7614.25 km scene (100 km2) in
eastern Jintang County, and orthorectified using a digital elevation model and 25
ground control points obtained by submeter accuracy global positioning systems
(GPS) using the high resolution satellite model (i.e. rigorous model) of PCI
4734
E. C. Ellis and H. Wang
Geomatics orthoEngine version 8.2 (PCI Geomatics, Richmond Hill, Ontario,
Canada), yielding a positional error of 4.0 m CE90 (circular error 90) as detailed in
Wang and Ellis (2005b), meeting MS IIRS (MultiSpectral Imagery interpretability
Rating Scale) Level 5 image quality standards (Imagery Resolution Assessments and
Reporting Standards (IRARS) Committee 1995). Black and white aerial photographs for the site on 25 June 1944 were obtained from the US National Archives
and Records Administration (NARA; RG-373, www.archives.gov) and orthorectified to a positional error of 6.5 m (CE90) using image tie points from orthorectified
IKONOS imagery (Wang and Ellis 2005b), meeting NIIRS (National Imagery
interpretability Rating Scale) Level 3 image quality standards (Imagery Resolution
assessments and Reporting Standards (IRARS) Committee 1996). IKONOS
imagery for the Yiyang site was acquired on 1 January 2002 and orthorectified by
a similar approach (Wang and Ellis 2005b).
2.3
Feature mapping and classification
Ecologically distinct landscape features (ecotopes) were mapped across sample cells
by field-validated interpretation of high spatial resolution imagery ((1 m) by
trained interpreters using a standardized ecological mapping and classification
system designed to delineate stable land management and vegetation features
observable at ground level at the time of image acquisition by both ecologists and
local land managers (anthropogenic ecotope mapping; Ellis et al. 2006). Ecotope
features were mapped by a scale-explicit sequential mapping strategy, beginning
with linear features (>2 m in width and area >25 m2, length >46width; examples
are roads and ditches), followed by hard areal features (>5 m in width and area
>25 m2 with clear edges and homogeneous interiors; examples are buildings and
water bodies), with the remaining area divided into soft features (>5 m in width
and area >100 m2 with fuzzy edges and variable cores; examples are crop plots
and vegetation patches). All ecotope features were corrected by field validation to
conform to stable (potentially observable for >2 y) land-management boundaries at
ground level in the field by the interpreter and local land managers in cases where
vegetation cover, shadow, or off-nadir imagery confused land-use boundaries in
imagery (Ellis et al. 2006). This procedure enabled the production of comparable
ecotope maps of the same site from high resolution imagery acquired in different
seasons (leaf on versus leaf off), and by different sensors, as long as imagery met
basic quality standards as described in Ellis et al. (2006).
After initial mapping, features were classified using a four level a priori ecological
classification hierarchy, FORMRUSERCOVERRGROUP + TYPE, combining
simple land form, use and cover classes (FORM, USE, COVER) with a set of more
detailed feature management and vegetation classes (GROUPs) stratified into
TYPEs. Ecotope classes are created by combining all four classification levels within
each feature. For example, a forest of closed canopy regrowth deciduous trees
(GROUP + TYPE5dt02) on a gentle slope (FORM5SL5Sloping) managed for
harvest (USE5T5forestry) with Perennial COVER (P) is classified as the ecotope
‘SLTPdt02’ (FORM + USE + COVER + GROUP + TYPE).
Initial maps of sample cells were prepared by trained interpreters by direct
interpretation of polygon features from imagery in a GIS (Arcinfo 8.3,
Environmental Systems Research Institute, Redlands, California) followed by
feature verification and correction in the field assisted by local land managers and
1 : 1200 scale image and feature maps. Historical maps were groundtruthed with the
Estimating feature-based area error
4735
aid of two local elders per sample cell, aged >16 at the time of image acquisition,
by a combination of interviews assisted by 1:1200 scale image maps and by
visiting all confusing areas in the field together with elders. This interpretation
and groundtruthing sequence was repeated twice by the same interpreter at each
sample cell and then reviewed by another trained interpreter to ensure compliance
with the standard mapping and classification rules, which included correction for
height distortion, shadowing and tree cover over buildings, roads, and water bodies
(Ellis et al. 2006). Final ecotope maps were checked against an ecotope classification
code database to check for and correct invalid code combinations and feature
topology was checked and corrected to ensure continuous classified polygon feature
layers and an overall map area error tolerance of + /20.05% of sample cell area.
Prior to mapping across sites, all four interpreters were trained at two different
sites across China by the repeated blind comparison of their maps with standardized
reference maps, to calibrate results across interpreters. The Jintang sample cell was
mapped by the full set of trained interpreters on site from 3–10 July 2003 (current
map) and from 10–16 July 2003 (historical map), along with two additional ‘trainee’
interpreters (trained at one site or less). The Yiyang sample cell was mapped from
10–17 November 2002 by the same interpreters. Interpreters were permitted to
discuss the mapping process with each other, but were not permitted to view others’
maps until all were completed.
2.4
Accuracy assessment
Map accuracy was assessed by comparing current and historical maps by three
trained interpreters with a ‘gold-standard’ reference map of the sample cell for each
time period prepared by field validation and correction of the map deemed most
accurate across 4 interpreters (Powell et al. 2004, Ellis et al. 2006); the interpreter
producing this base reference map was subsequently excluded from analysis of
interpreter error. Use of a gold-standard reference map corrected across interpreters
is an optimal strategy for thematic map accuracy assessment, producing more
reliable results than those from any one interpreter, even though the reference map
must still include some error (Powell et al. 2004).
Data for map accuracy assessment were obtained for each sample cell by placing a
3 m triangular grid of points over the maps of each interpreter and the reference
map using GIS, and then determining the classification of each point by each
interpreter and the reference (the triangular grid had n531 837 points in the Jintang
sample cell, and had n520 064 points in the Yiyang cell). The triangular grid point
sample was a better representation of map class areas than a random point sample
of the same size, as the proportion of the sample cell covered by each ecotope class
was found to be nearly identical between the triangular grid point estimate and the
areas determined from the original classified ecotope polygons. (In Jintang, the sum
of absolute ecotope area differences between the polygon data and the triangular
grid sample was 0.4%, but was 1.8% for the random sample.) The classification data
obtained from the grid point sample were entered into a conventional error matrix
to calculate overall map accuracy (correct/total) and class user’s accuracy (correct/
mapped) by pairwise comparison of each interpreter with the reference map
(Congalton and Green 1999). Cohen’s Kappa (k) and Andrés and Marzo’s Delta (D;
2004) measures of agreement corrected for random chance were calculated using the
algorithm of Andrés and Marzo (2004; http://www.ugr.es/,bioest/delta.htm). k and
4736
E. C. Ellis and H. Wang
D are related statistics, with values .0.75 indicating strong agreement above
chance (Andrés and Marzo 2004, Norusis 2004b).
2.5
Area measurement errors
Three independent methods were used to estimate errors in map class areas. Two of
these, class user’s error (CUE) and class area error (CAE), were calculated from
conventional error matrix data obtained as above and map class areas calculated as
the sum of polygon areas, respectively. CUE, or ‘error of commission’, includes
errors in position and was calculated as 1 – class user’s accuracy from the row
marginals of the error matrix for each interpreter (Story and Congalton 1986). CAE,
the ‘non-site-specific error’ of Congalton and Green (1999), excludes positional
error, and was calculated as the difference between interpreter (CAobs) and reference
(CAref) map class areas, divided by interpreter area (CAobs):
CAE~
CAref {CAobs
CAobs
ð1Þ
A third, feature-based method for estimating area errors, feature area error (FAE),
was derived from analysis of a limited set of individual ecotope features (j) mapped
by each interpreter (i), as the absolute difference in area between corresponding
interpreter (FA(obs)) and reference (FA(ref)) features
FAEij ~ FAðref Þ{FAðobsÞi
ð2Þ
in theory, FAE for each map class (FAEk) could be calculated as the sum of FAEij
across all of the features in the class divided by the class area, thereby obtaining
error estimates comparable with CUE or CAE. This is not possible however,
because interpreters do not always recognize exactly the same features within each
sample cell. We therefore measured FAEij for a subset of current and historical
features reproduced by all interpreters and analysed these for relationships with
feature classification, area, perimeter and other factors to determine whether models
based on these factors might facilitate prediction of FAE across map classes.
Features selected for FAE analysis varied across the range of feature sizes and
included only those with perimeters mostly within the sample cell (.75% mapped
versus clipped edges). For current maps, FAEij was calculated from five independent
mappings of 35 features including 16 hard and 19 soft features (3 trained + 2 trainee
interpreters, total5175 features). FAEij for historical maps was calculated from
three independent mappings of 14 features, including 7 hard and 7 soft features
(three trained interpreters, total542 features).
2.6
Error estimators for map class areas
Errors in map class area estimates from the three interpreters of the Jintang sample
cell were tested for relationships with map class (ecotope, USE), feature type (hard
versus soft), feature area and feature perimeter to determine whether all or some of
these factors were useful predictors of error in map class area estimates. Ecotope
classes smaller than 0.25% of sample cell area were excluded from this and other
analyses because these smaller ecotope classes were both statistically unreliable
(more than half had ,70% user’s accuracy; Foody 2002) and unimportant: in total
these never accounted for .2.4% of any sample cell area. We tested for relationships
Estimating feature-based area error
4737
between FAEij and feature area and perimeter using linear regression, and then
combined these variables together with USE class and hard vs. soft feature type in a
Univariate general linear model (GLM) to test for ecotope FAEij differences and for
the factors linked to these differences using data for individual features averaged
across three interpreters (Norusis 2004b). This approach produced a complete
model of FAEij for each time period, including all variables as fixed factors, along
with their interactions. To test for differences among ecotope CUE and CAE
estimates, linear mixed models (LMM; Norusis 2004b) were used, with ecotope as
fixed factor, because this test is more reliable than GLM when observations are not
fully independent with equal variance, as was observed for ecotope CUE and CAE
estimates three interpreters per time period, 20 current ecotopes, 10 historical
ecotopes; Levene’s test P,0.05; (Snedecor and Cochran 1980).
Two models for predicting error intervals for class area estimates were evaluated
in terms of the total amount of error they produced and their statistical reliability.
Statistical reliability was quantified as the probability that the rate of successful
error predictions produced by an estimator (intervals covering the correct value) was
at least as great as expected for an estimator that is 90% successful in error
prediction (90% reliable). The first and simplest model was to predict class area error
as a direct percentage of ecotope class area (class area error5class area6error).
CUE and CAE already describe class area errors as a percentage of class area, so
general estimates of these errors across ecotope classes and interpreters (CUEoverall ,
CAEoverall) were made by bootstrapping their 95% confidence upper limit (pooled
CUE and CAE were not normally distributed; current n560520 ecotopes63
interpreters, CUE517.7%, CAE529.2%; historical n530510 ecotopes63 interpreters, CUE531.0%, CAE533.8%; 3000 parametric bootstrap runs; Efron and
Tibshirani 1986).
A second error model based on feature perimeter was derived from the
relationship between feature perimeter and FAEij, based on the concept of ‘epsilon
bands’ of error around mapped features (reviewed by Green and Hartley 2000). In
this model, the mapped perimeter of each ecotope feature (FPj; the portion of
feature perimeter not cut by sample cell boundaries) was multiplied by an error
factor (FAEoverall, in m) to estimate FAEj for each mapped feature. FAEj was then
divided by feature area (FAj) and added across all of the features in each map class
(k) to estimate map class FAE as a percentage of map class area (FAEk):
FAEk ~
X FPj |FAEoverall
J
FAj
ð3Þ
FAEoverall was estimated as the 95% confidence upper limit of the slope of the FAEij
to FPj relationship calculated by linear regression (current50.93 m,
historical53.57 m, regression R2.0.75, P,0.001 for both times).
90% confidence intervals for map class areas were calculated for the three error
estimators (CUEoverall, CAEoverall, FAEoverall), by multiplying each error estimator
by Z0.05 (1.645; 90% confidence for the normal distribution), given that ecotopelevel error across interpreters (CUE and CAE) approximated a normal distribution
(Shapiro-Wilk test P.0.05 for 56 out of 60 current + historical estimates; Norusis
2004a). The statistical reliability of these error intervals was then tested by
calculating them across the ecotope area estimates of the three interpreters of each
sample cell (Jintang current n560, historical n530), tabulating the number of
successful error predictions (error intervals containing the reference value), and
4738
E. C. Ellis and H. Wang
testing their resulting error prediction success rate against the hypothesis of 90%
prediction success using the binomial test (Taylor 1997). This test compared the
error prediction success rate with the chance of observing less than or equal to this
number of successes given the number of trials and the hypothesis that the error
estimator is 90% reliable (error prediction is successful in >90% of predictions).
When this binomial test gives a less than 10% chance of the observed success rate,
the hypothesis of 90% reliable error prediction is rejected with 90% confidence.
When the test yields a >10% chance of the observed success rate, a 90% reliable
error model is not rejected, and when the binomial test yields a >90% chance of
the observed success rate, then it is >90% probable that the error model is .90%
reliable.
To investigate relationships between error prediction success and total error
introduced by area-based (e.g. CUE, CAE) versus perimeter-based error models
(e.g. FAE), we calculated error intervals for ecotope area estimates by the three
interpreters of the Jintang sample cell across a range of values for area and perimeter
error. At each increment of error predicted by each model (area, perimeter), ecotope
error intervals for the three interpreters were tested for their error prediction success
rate and sum of errors across ecotope classes. Based on results of this analysis, areaand perimeter-based models were then tested for their prediction success rate and
overall error across three interpreters at the Yiyang sample cell (16 current ecotope
classes).
3.
Results and discussion
3.1
Mapping accuracy
Overall accuracy of current (2002) and historical (1944) ecological maps of the
Jintang sample cell (figure 1) was .85% across interpreters at all classification levels,
with median accuracy .90%, meeting widely accepted map accuracy standards
(tables 1 and 2; Foody 2002). Map accuracy corrected for chance using Cohen’s k
was greater for current maps (.85%) than for historical maps (.75%), and was even
higher using the related and more robust D statistic (.86% for current versus .83%
for historical maps; Andrés and Marzo 2004). Overall map accuracy for individual
classification levels (USE and GROUP) was greater than that of combined classes,
such as ecotope and USE + COVER, confirming that the accuracy of area estimates
can be increased by lowering their thematic resolution (Petit and Lambin 2001,
Smith et al. 2003; median overall accuracy for current GROUP maps 592.6%,
USE + COVER 591.6%). As expected, with less ecotope classes (16) and less
intensive land management, ecotope map accuracy at the Yiyang sample cell was
even higher than in Jintang, with overall accuracy ranging from 87% to 93% across
three interpreters, and with k estimates ranging from 84% to 91% (2002 map; overall
D ranged from 86–92%).
3.2
Error in map class areas
Map class accuracy and map class area were related, but not in a simple way. At all
levels of classification, classes with larger areas tended toward greater accuracy
(D and class user’s accuracy 512CUE), even though some small classes had
surprisingly high accuracy (mostly hard features such as houses), and some large
classes had surprisingly low accuracy (mostly soft features with Disturbed, Forestry,
or Fallow USE; tables 1 and 2). As a result, user’s accuracy was only weakly related
Estimating feature-based area error
4739
Figure 1. Imagery and ecotope reference maps of the Jintang 5006500 sample cell (UTM
projection). (a) 1944 aerial photograph and (b) historical ecotope map. (c) 2001 IKONOS
imagery and (d) current ecotope map. Ecotope features are symbolized with USE class
symbols overlaid by COVER class symbols, as described in the legend. A box highlights area
expanded in figure 3.
to map class area either directly or after log-transformation of area (regression
R2,0.31, P,0.05). On the other hand, more than 90% of current ecotope classes
with satisfactory user’s accuracy (.70%; Foody 2002) had larger areas (.0.25% of
sample cell area), and more than half of ecotopes with poor user’s accuracy (,70%)
were smaller classes ((0.25% of sample cell area), similar to previous results
(Cherrill and McClean 1995). As these smaller ecotopes never covered .2.4% of any
sample cell across sites, we consider it both prudent and practical to flag all map
USE
Area (% cell)
D (%)
CUE (%)
CAE (%)
FAE (%)
Small-scale staple crops (Bench Plateau)
‘‘ (Sloping)
‘‘ (Summit)
‘‘ (Foot Slope)
Small-scale immature pear orchard (Bench Plateau)
Small-scale mature mandarin orange orchard (Bench Plateau)
‘‘ (Sloping)
65.1¡0.6
36.4¡0.8
10.7¡0.1
2.3¡0.1
1.4¡0.2
2.8¡0.1
11.0¡0.3
0.4¡0.3
93.2¡2.8
93.3¡2.5
95.8¡1.0
89.5¡0.5
84.6¡4.0
93.3¡10.3
90.4¡4.6
60.3¡44.0
3.4
4.4
3
7.7
3.6
6.9
10.7
21.4*
0.7
0.9
1.2
1.9
11.6
2
1.9
103.0*
6.5
5.7
3.3
8.5
8
8.7
10.5
13.3
Reservoir paddy (Foot Slope)
Rice paddy (Foot Slope)
‘‘ (Bench Plateau)
18.2¡0.7
12.7¡0.6
3.8¡0.2
1.8¡0.1
95.2¡2.0
95.0¡2.1
92.7¡2.2
89.6¡0.7
4.0
2.9
6.7
8.3
1.8
2.2
2.7
3.3
5
4.9
4.3
9.9
Regrowth open woody vegetation (Steep Slope)
Planted conifer forest (Steep Slope)
Regrowth open wooded brush (Steep Slope)
5.7¡1.2
0.4¡0.1
4.8¡1.2
0.5¡0.0
75.6¡9.0
77.4¡13.6
86.0¡5.3
71.0¡0.9
27.9
19.2*
24.9*
16.5*
12.6
10.3
14.6
2.8
16.8
16.5
16.7
18
Unpaved local roads
Single story attached housing
Multistory attached housing
Single story detached housing
5.5¡0.2
1.8¡0.0
1.9¡0.1
0.9¡0.1
0.4¡0.1
81.4¡5.0
68.3¡7.9
86.5¡7.0
84.2¡11.2
73.8¡15.3
18.1
29.3*
11.9
12.6
15.8*
3.0
1.5
1.8
9.7
13.7
29.1
56.7
14
10.8
19.6
Disturbed woody vegetation with debris (Steep Slope)
Disturbed trees with debris (Bench Plateau)
‘‘ (Steep Slope)
5.4¡0.5
1.7¡1.2
1.4¡0.4
1.2¡0.9
68.2¡2.9
37.7¡14.9
54.6¡17.2
81.2¡8.7
31.8
23.9*
27.9*
47.3*
5.8
127.9*
23
38.4
24.6
13.9
22.2
36.4
USE classification, overall k588.5¡3.3
99.9¡3.3
90.9¡2.1
16.9
20.4
16.2
Ecotope classification, overall k586.2¡2.9
96.0¡5.2
87.8¡2.6
15.2
18.7
15.1
8
5.7
8.5
Rainfed agriculture
Paddy
Forestry
Constructed
Disturbed
Sum of ecotope errors as % sample cell area
*
Significant ecotope fixed effect in LMM (a50.05).
E. C. Ellis and H. Wang
Ecotope
4740
Table 1. Current (2002) USE and ecotope class areas (reference¡max difference from reference), accuracy corrected for chance using Andrés and Marzo’s D
(median¡max difference from median), mean class user’s error (CUE) and class area error (CAE) across three interpreters, along with predicted feature area
error (FAE) as % class area for the Jintang sample cell (figure 1). USE classes are sorted by area and partitioned into ecotopes, with FORM class indicated in
parentheses where needed.
Table 2. Historical (1944) USE and ecotope class areas (reference¡maximum difference from reference), map class accuracy corrected for chance using
Andrés and Marzo’s D (median¡maximum difference from median), mean class users error (CUE) and class area error (CAE) across 3 interpreters, and
predicted feature area error (FAE) as percent of class area for the Jintang sample cell (figure 1).
Ecotope
Area (% cell)
D (%)
CUE (%)
CAE (%)
FAE (%)
90.6¡8.2
2.7
3.9
8.2
Rainfed agriculture
Small-scale staple crops (Bench Plateau)
77.5¡6.7
Paddy
Reservoir paddy (Foot Slope)
10.0¡2.1
94.9¡9.5
16.8
11.2
23.0
Broadleaf herbaceous regrowth vegetation (Summit)
Exposed rock (Bench Plateau)
‘‘ (Steep Slope)
Small pond
5.2¡1.3
2.3¡0.1
1.2¡0.2
0.5¡1.0
0.8¡0.3
84.9¡8.5
96.4¡1.5
77.7¡11.4
74.2¡11.5
82.3¡38.9
21.7
3.6
22.8
12.2
28.1
9.7
2.0
72.7*
11.9
12.8
37.6
17.6
37.1
45.3
50.3
Tree-covered regrowth grave
Disturbed trees with debris (bench plateau)
3.8¡1.2
2.2¡0.3
1.6¡0.5
68.6¡12.9
82.9¡1.6
50.1¡14.2
29.8
17.2*
48.4*
19.4
9.5
25.0
50.5
22.4
78.5
Regrowth conifer forest (Steep Slope)
2.6¡1.8
77.6¡16.6
41.7*
38.0
52.4
0.9¡0.5
86.6¡23.6
45.6
*
41.5
56.5
100.0¡13.6
99.5¡14.3
87.7¡4.4
87.6¡4.4
26.4
23.9
20.6
22.8
38.0
39.1
7.0
7.2
14.5
Fallow
Disturbed
fromestry
Constructed
Single-story attached housing
USE classification, overall k579.0¡5.7
Ecotope classification, overall k578.5¡5.6
Sum of ecotope errors as % sample cell area
*
Estimating feature-based area error
USE
Significant ecotope fixed effect in LMM (a50.05).
4741
4742
E. C. Ellis and H. Wang
Figure 2. Classification disagreement between 3 trained interpreters across the Jintang
sample cell (figure 1) in terms of (a) historical and (b) current ecotope map class (blank5all
agree). A box highlights area expanded in figure 3.
classes covering(0.25% of a sample cell as unreliable and to eliminate them from
further analysis, although up to half of these smaller estimates might in fact be
accurate. For example, the gold standard current reference map of the Jintang
sample cell had 32 ecotope classes, and interpreters used from 30 to 33 classes, but
all had the same 20 larger ecotope classes after eliminating ecotopes smaller than
0.25% of the sample cell.
Comparison of maps across trained interpreters confirmed that errors in feature
alignment and edge mapping were ubiquitous across current and historical maps of
the Jintang sample cell (figure 2, with detail in figure 3; Green and Hartley 2000). It
was also apparent that misclassification was far more common for smaller features
(higher perimeter/area), likely because these features attracted less attention from
interpreters and were often more difficult to classify than larger features (figure 2).
Together, these observations indicate that feature perimeter should be a strong
predictor of error in class area estimates (Green and Hartley 2000).
Indeed, feature perimeter was a stronger linear predictor of interpreter error in
feature area estimates (FAEij) than feature area itself (current linear regressions,
figure 4: perimeter R250.76, area R250.62; historical linear regressions, not shown:
perimeter R250.86, area R250.84; P,0.001 for all regressions). Although feature
area was nearly as strong a predictor of error as feature perimeter using power
regression, this relationship can be explained by the strong relationship between
feature area and feature perimeter (Figure 4(c), P,0.001 for both regressions).
Given that perimeter was the stronger predictor, and that feature area was a
nonsignificant predictor of FAE when area and perimeter were regressed together,
feature perimeter was the more powerful predictor of FAE (multiple linear
regression, current perimeter P,0.001, area P50.16, R250.77 overall P,0.001;
historical perimeter P50.07, area P50.25, R250.88 overall P,0.001). Surprisingly,
Estimating feature-based area error
4743
Figure 3. Detail of current feature mapping errors at the Jintang sample cell (figure 1). (a)
Reference feature boundaries over 2001 IKONOS image. (b) Feature boundaries for 5
interpreters (3 trained, 2 trainee) overlaid on interpreter classification disagreement (b).
FAE did not differ significantly between land USE class or feature types (hard vs.
soft), nor did these factors affect the relationship between FAE and feature
perimeter or area (P.0.2, figure 4). Therefore, the simple linear relationship
between FAEij and feature perimeter provides a straightforward model for
predicting errors in map class areas from feature perimeters by multiplying these
by the slope of the perimeter to FAE relationship (the line intercept was statistically
non-significant).
Error estimators based on the conventional error matrix produced similar, but
not identical results. CUE includes errors in feature position and estimates error as
the misclassified percentage of each class, while CAE is independent of feature
position and agreement with other classes (Congalton and Green 1999). As a result,
even though their means across ecotope classes were not significantly different, CUE
was not correlated with CAE and also varied less between classes than CAE (tables 1
and 2). As expected, CUE also showed a tendency toward higher error estimates for
ecotopes with smaller features (roads and detached houses) and features with less
discrete edges (Disturbed and Forestry land USE), while CAE differences between
ecotopes were without clear trend (tables 1 and 2).
3.3
Error estimators for ecological map class area estimates
Given the goal of producing statistically reliable error estimates for ecological map
class areas from sample cell maps by a set of trained interpreters across five rural
sites in China, we developed a practical strategy to predict these errors based on four
principles. First, to avoid false detection of changes and differences in class areas
(Type I error), larger, more conservative error intervals were estimated, allowing
4744
E. C. Ellis and H. Wang
Figure 4. Relationships of feature area error (FAE) with (a) feature area and (b) perimeter,
and (c) the relationship between feature area and perimeter for current maps of the Jintang
sample cell. USE class of features is indicated by symbols, solid lines are linear regression,
dashed lines are power regression.
that smaller differences might go undetected (Type II error; Schenker and
Gentleman 2001). Second, interpreter error was considered the dominant source
of error in ecological map class areas from fine-scale feature-based mapping
(Cherrill and McClean 1999). Third, a set of trained and calibrated interpreters
following a standardized scale-explicit ecological mapping and classification
procedure, such as the system used in this study, should produce consistent types
of interpreter error across sites (Cherrill and McClean 1999, Powell et al. 2004).
Finally, the assessment of interpreter error under the most challenging mapping
conditions should produce conservative estimates of interpreter error across sample
cells in general. Based on these principles, and the observation that misregistration
error was nonsignificant for ecotope area estimates across 5006500 m sample cells
(Wang and Ellis 2005a ), we developed a conservative error prediction model based
on analysis of interpreter error at a sample cell in the Jintang site, representing the
Estimating feature-based area error
4745
Figure 5. Statistical reliability of error intervals estimated for ecotope areas versus the error
produced by the estimator across three interpreters of the Jintang sample cell. Horizontal
control lines at P50.1 and 0.9 highlight minimum and greater than expected P for a 90%
reliable estimator. Large filled (current) and hollow (historical) symbols represent ecotopelevel estimates of CUE ( , #) and CAE (m, n; tables 1 and 2), small symbols of the same
type are CUE and CAE estimators pooled across ecotopes, along with FAE predicted from
feature perimeters (¤, e). Solid lines (heavy5current, light5historical) describe error
estimated as a factor of feature perimeter (i.e. FAE), dashed lines are error estimated as a
factor of feature area (e.g. CUE and CAE).
N
most challenging current and historical mapping conditions across the sites in this
study (Ellis 2004).
Figure 5 illustrates the statistical reliability and sum of errors introduced by
different models for estimating error intervals for ecological map class areas across
the three interpreters of the Jintang sample cell. This figure demonstrates that all
error estimators were acceptable as 90% reliable under both current and historical
mapping conditions, but that their reliability ranged from the minimum acceptable
to significantly greater than required of a 90% reliable error estimator. As expected,
CUE and CAE error intervals estimated for each ecotope class (tables 1 and 2) were
more powerful error predictors than CUE or CAE pooled across ecotope classes,
and produced less overall error than all other models (figure 5). However, FAE was
nearly as powerful an error predictor as ecotope-level CUE and CAE even though
FAE was predicted by a pooled error model across ecotope classes (equation (3);
figure 5; tables 1 and 2). Surprisingly, historical error estimates were usually more
reliable than current estimates, most likely owing to their wider error intervals and
smaller numbers of map classes; for every sample cell in this study, historical maps
had fewer ecotope classes than current maps.
Error prediction by a simple perimeter-based model (e.g. FAE, equation (3) in
section 2.6) was more powerful than prediction by a simple area-based model (error
as constant factor of class area, e.g. pooled CUE and CAE), producing far less
overall error at all levels of prediction reliability (figure 5). In contrast with areabased errors, perimeter-based errors varied substantially between ecotope classes,
with current FAE predictions ranging from ,40% to .6 times the overall map
error (table 1). Moreover, ecotope-level FAE predictions followed similar trends as
those observed in ecotope CUE (linear regression R250.61, P,0.001), with larger
4746
E. C. Ellis and H. Wang
errors for ecotopes with smaller features and for features with more complex edges
(tables 1 and 2; CAE and FAE were unrelated, P50.77). Taken together, these
results demonstrate that a simple perimeter-based error model can yield statistically
reliable error predictions that approximate observed CUE, a robust error estimator
based on the conventional error matrix (Congalton and Green 1999), while
introducing much less overall error than simple area-based models.
Analysis of perimeter-based error demonstrated that highly reliable error intervals
with a >95% chance of .90% prediction success were produced by error factors of
>1.3 m for current maps and >2.7 m for historical maps, introducing overall errors
of about 11% for each time period (figure 5, refer to equation (3) in section 2.6). To
achieve the same reliability, area-based errors of 39% and 25% were required for
current and historical maps, respectively (figure 5). Given that for the same level of
prediction reliability, perimeter-based error estimates were sometimes much higher
than area-based estimates for the same map classes, a mixed error model utilizing
only the lower of the two error estimates for each map class should decrease overall
error without sacrificing statistical reliability. When this mixed error model was
applied to the Jintang sample cell using the highly reliable error factors noted above,
error was reduced in 6 of 60 current ecotope estimates, decreasing overall error from
11.8 to 11.1%, and also decreased error in 17 of 30 historical ecotope estimates while
lowering overall error from 11.0 to 9.5%, without any effect on prediction success
rates for either time period.
Conservative error models based on interpreter error under the most challenging
conditions across sites should be even more reliable when predicting errors at sites
with less challenging conditions. We tested this hypothesis by applying the highly
reliable error models from the Jintang sample cell to current ecotope maps of a
sample cell in Yiyang, Hunan by the same set of interpreters. From 16 ecotope
classes mapped by 3 interpreters (n548), the area-based error model yielded 46
successful predictions with a 39% sum of errors across the sample cell (96% chance
of .90% reliability), while the perimeter-based error model produced 45 successful
predictions and a 13.8% sum of errors (87% chance of .90% reliability). The mixed
error model was as reliable as the perimeter-based model, but lowered error in 7 of
48 of estimates, yielding a 12.6% sum of errors across the sample cell. These results
confirm that conservative error models based on interpreter error at the Jintang
sample cell performed equally well under substantially different mapping conditions
at a sample cell more than 800 km distant across rural China.
Statistically reliable ecological map class area estimates are a fundamental part of
long-term ecological change measurements in densely populated landscapes (Johnes
and Butterfield 2002, Ellis 2004). Though no error model can predict the full range
of errors across all mapping conditions, our analysis of interpreter error under the
most challenging mapping conditions across sites in rural China facilitated the
development of statistically reliable error interval prediction models for map class
area estimates across sample cells mapped by a set of trained and calibrated
interpreters using our standardized feature-based ecological mapping system.
Though our error models are specific to ecological mapping by trained interpreters
in rural China, it should be possible to apply a similar error modelling approach,
potentially incorporating feature type and image quality, to other standardized
feature-based mapping systems, including automated feature detection, providing a
practical basis for statistically reliable estimates of long-term ecological changes in
landscape structure across densely populated landscapes.
Estimating feature-based area error
4747
Acknowledgements
This material is based upon work supported by the US National Science
Foundation under Grant DEB-0075617 awarded to Erle C. Ellis in 2000, conducted
in collaboration with Professor Linzhang Yang of the Institute of Soil Science,
Chinese Academy of Sciences (CAS), Nanjing, China, Professor Hua Ouyang of the
Institute of Geographic Sciences and Natural Resources Research, CAS, Beijing,
China and Professor Xu Cheng of China Agricultural University, Beijing, China.
We are grateful to our local collaborators for field assistance and to our site
researchers Hongsheng Xiao, Kui Peng, Shoucheng Li, and Xinping Liu and to
Jonathan Dandois and Junxi Wu for mapping work across China, and to Dominic
Cilento for help with GIS analysis. Thanks to the National Archives and Records
Administration for historical aerial photographs in China. Any opinions, findings
and conclusions or recommendations expressed in this material are those of the
authors and do not necessarily reflect the views of the National Science Foundation.
References
ANDRÉS, A.M. and MARZO, P.F., 2004, Delta: a new measure of agreement between two
raters. British Journal of Mathematical and Statistical Psychology, 57, pp. 1–19.
AUGUST, P., IVERSON, L. and NUGRANAD, J., 2002, Human conversion of terrestrial habitats.
In Applying Landscape Ecology in Biological Conservation, K.J. Gutzwiller (Ed.) (New
York: Springer), pp. 198–234.
CHERRILL, A. and MCcLEAN, C., 1995, An investigation of uncertainty in field habitat
mapping and the implications from detecting land-cover change. Landscape Ecology,
10, pp. 5–21.
CHERRILL, A. and MCcLEAN, C., 1999, Between-observer variation in the application of a
standard method of habitat mapping by environmental consultants in the UK.
Journal of Applied Ecology, 36, pp. 989–1008.
CONGALTON, R.G. and GREEN, K., 1999, Assessing the Accuracy of Remotely Sensed Data
(Boca Raton, Fl: CRC Press).
CONGALTON, R.G. and MEAD, R.A., 1983, A quantitative method to test for consistency and
correctness in photointerpretation. Photogrammetric Engineering and Remote Sensing,
49, pp. 69–74.
DEFRIES, R.S., FOLEY, J.A. and ASNER, G.P., 2004, Land-use choices: balancing human
needs and ecosystem function. Frontiers in Ecology and the Environment, 2, pp.
249–257.
EFRON, B. and TIBSHIRANI, R., 1986, Bootstrap methods for standard errors, confidence
intervals, and other measures of statistical accuracy. Statistical Science, 1, pp. 54–77.
ELLIS, E.C., 2004, Long-term ecological changes in the densely populated rural landscapes
of China. In Ecosystems and Land Use Change, R.S. Defries, G.P. Asner and R.A.
Houghton (Eds) (Washington, DC: American Geophysical Union), pp. 303–320.
ELLIS, E.C., LI, R.G., YANG, L.Z. and CHENG, X., 2000a, Changes in village-scale nitrogen
storage in China’s Tai Lake region. Ecological Applications, 10, pp. 1074–1089.
ELLIS, E.C., LI, R.G., YANG, L.Z. and CHENG, X., 2000b, Long-term change in Village-scale
ecosystems in China using landscape and statistical methods. Ecological Applications,
10, pp. 1057–1073.
ELLIS, E.C., WANG, H., XIAO, H.S., PENG, K., LIU, X.P., LI, S.C., OUYANG, H., CHENG, X.
and YANG, L.Z., 2006, Measuring long-term ecological changes in densely populated
landscapes using current and historical high resolution imagery. Remote Sensing of
Environment, 100, pp. 457–473.
FOLEY, J.A., COSTA, M.H., DELIRE, C., RAMANKUTTY, N. and SNYDER, P., 2003, Green
surprise? How terrestrial ecosystems could affect Earth’s climate. Frontiers in Ecology
and the Environment, 1, pp. 38–44.
4748
E. C. Ellis and H. Wang
FOODY, G.M., 2002, Status of land cover classification accuracy assessment. Remote Sensing
of Environment, 80, pp. 185–201.
FORSTER, B.C., 1985, Examination of some problems and solutions in monitoring urban areas
from satellite platforms. International Journal of Remote Sensing, 6, pp. 139–151.
GALLEGO, F.J., 2000, Double sampling for area estimation and map accuracy assessment. In
Quantifying Spatial Uncertainty in Natural Resources: Theory and Applications from
GIS and Remote Sensing, H.T. Mowrer and R.G. Congalton (Eds) (Chelsea, Mich.:
Ann Arbor Press), pp. 65–77.
GALLEGO, F.J., DELINCE, G. and CARFUGNA, E., 1994, Two stage area frame on squared
segments for farm surveys. Survey Methodology, 20, pp. 107–115.
GREEN, D.R. and HARTLEY, S., 2000, Integrating photointerpretation and GIS for vegetation
mapping: some issues of error. In Vegetation Mapping: From Patch to Planet, R.W.
Alexander and A.C. Millington (Eds) (Chichester; New York: Wiley), pp. 103–134.
HILL, R.A., SMITH, G.M., FULLER, R.M. and VEITCH, N., 2002, Landscape modelling using
integrated airborne multi-spectral and laser scanning data. International Journal of
Remote Sensing, 23, pp. 2327–2334.
HOUGHTON, R.A., 2003, Why are estimates of the terrestrial carbon balance so different?
Global Change Biology, 9, pp. 500–509.
IMAGERY RESOLUTION ASSESSMENTS and REPORTING STANDARDS (IRARS) COMMITTEE 1995,
Multispectral Niirs Reference Guide, Imagery Resolution assessments and Reporting
Standards, (IRARS) Committee.
IMAGERY RESOLUTION ASSESSMENTS and REPORTING STANDARDS (IRARS) COMMITTEE 1996,
Civil Niirs Reference Guide, Imagery Resolution Assessments and Reporting Standards,
(IRARS) Committee.
JENSEN, J.R. and COWEN, D.C., 1999, Remote sensing of urban/suburban infrastructure and
socio-economic attributes. Photogrammetric Engineering and Remote Sensing, 65, pp.
611–622.
JOHNES, P.J. and BUTTERFIELD, D., 2002, Landscape, regional and global estimates of
nitrogen flux from land to sea: errors and uncertainties. Biogeochemistry, 57–58, pp.
429–476.
LO, C.P. and SHIPMAN, R.L., 1990, A GIS approach to land-use change dynamics detection.
Photogrammetric Engineering and Remote Sensing, 56, pp. 1483–1491.
LU, D., MAUSEL, P., BRONDÍZIO, and MORAN, E., 2004, Change detection techniques.
International Journal of Remote Sensing, 25, pp. 2365–2401.
NORUSIS, M., 2004a, SPSS 12.0 Guide to Data Analysis (Upper Saddle River, New Jersey:
Prentice Hall).
NORUSIS, M., 2004b, SPSS 12.0 Statistical Procedures Companion (Upper Saddle River, New
Jersey: Prentice Hall).
PETIT, C.C. and LAMBIN, E.F., 2001, Integration of multi-source remote sensing data for land
cover change detection. International Journal of Geographical Information Science, 15,
pp. 785–803.
POWELL, R.L., MATZKE, N., DE SOUZA J., C., CLARK, M., NUMATA, I., HESS, L.L. and
ROBERTS, D.A., 2004, Sources of error in accuracy assessment of thematic land-cover
maps in the Brazilian Amazon. Remote Sensing of Environment, 90, pp. 221–234.
SCHENKER, N. and GENTLEMAN, J.F., 2001, On judging the significance of differences by
examining the overlap between confidence intervals. American Statistician, 55, pp.
182–186.
SMITH, J.H., STEHMAN, S.V., WICKHAM, J.D. and YANG, L., 2003, Effects of landscape
characteristics on land-cover class accuracy. Remote Sensing of Environment, 84, pp.
342–349.
SNEDECOR, G.W. and COCHRAN, W.G., 1980, Statistical Methods, 7th edition (Ames, Iowa,
(USA: The Iowa State University Press).
STORY, M. and CONGALTON, R.G., 1986, Accuracy assessment: a user’s perspective.
Photogrammetric Engineering and Remote Sensing, 52, pp. 397–399.
Estimating feature-based area error
4749
TAYLOR, J.R., 1997, An Introduction to Error Analysis: the Study of Uncertainties in Physical
Measurements, 2nd edition (Mill Valley, California: University Science Books).
THOMAS, N., HENDRIX, C. and CONGALTON, R.G., 2003, A comparison of urban mapping
methods using high-resolution digital imagery. Photogrammetric Engineering and
Remote Sensing, 69, pp. 963–972.
TOWNSHEND, J.R.G., JUSTICE, C.O., GURNEY, C. and MCMANUS, J., 1992, The impact of
misregistration on change detection. IEEE Transactions on Geoscience and Remote
Sensing, 30, pp. 1054–1060.
TROLL, C., 1939, Luftbildplan Und Ökologische Bodenforschung (Aerial Photography and
Ecological Studies of the Earth). Zeitschrift Der Gesellschaft Für Erdkunde, pp.
241–298.
VERBYLA, D.L. and BOLES, S.H., 2000, Bias in land cover change estimates due to
misregistration. International Journal of Remote Sensing, 21, pp. 3553–3560.
VITOUSEK, P.M., MOONEY, H.A., LUBCHENCO, J. and MELILLO, J.M., 1997, Human
domination of Earth’s ecosystems. Science, 277, pp. 494–499.
WANG, H. and ELLIS, E.C., 2005a, Image misregistration error in change measurements.
Photogrammetric Engineering and Remote Sensing, 71, pp. 1037–1044.
WANG, H. and ELLIS, E.C., 2005b, Spatial accuracy of orthorectified Ikonos imagery and
historical aerial photographs across five sites in China. International Journal of
Remote Sensing, 26, pp. 1893–1911.