Professional Documents
Culture Documents
net/publication/279231172
CITATIONS READS
25 1,558
3 authors:
Neal Scott
Queen's University
93 PUBLICATIONS 4,663 CITATIONS
SEE PROFILE
Some of the authors of this publication are also working on these related projects:
Biophysical Remote Sensing and Terrestrial CO2 Exchange at Cape Bounty, Melville Island View project
All content following this page was uploaded by Meg Southee on 25 February 2022.
December 2012
Volume 78, Number 12
PE&RS
Application of Lidar Terrain Surfaces for Soil
Moisture Modeling
Florence Margaret Southee, Paul M. Treitz, and Neal A. Scott
Figure 1. Location of high-precision lidar flightlines and sample plots in the Romeo Malette Forest.
(c) (d)
and a 100 percent value represents a peak in the DEM. where a ⫽ the local upslope area draining through each cell,
The PEI was applied to the DEMs using a round filter with and tanb ⫽ the local slope gradient.
different neighborhood sizes (Table 1). An effort was made The TWI was processed in Whitebox version 1.0.7
to select filter sizes that were roughly equivalent to the size (John Lindsay (2010); University of Guelph, Guelph,
of the plots; however, plot level precision was not directly Ontario) on the raw and IRA-treated DEMs using a multiple
achievable at the larger resolutions due to the larger grid flow direction algorithm (i.e., FD8) (Freeman, 1991; Quinn
resolution. The PEI was processed on the raw DEMs with et al., 1991). The FD8 flow algorithm allows water to flow
culverts in Whitebox version 1.0.7 (John Lindsay (2010); into multiple neighboring cells based on the convex or
University of Guelph, Guelph, Ontario). concave shape of the landscape. It assigns flow to all
The CHM represents the actual vegetation height in downstream neighbors of a given cell based on the degree
meters. To generate the CHM the raw DEM was subtracted of slope gradient. Wolock and McCabe (1995) observed
from the DSM to determine the maximum height of the forest that the TWI patterns derived from the multiple flow
canopy at each cell within the lidar flightlines. direction algorithms are smoother and more suited for
The TWI represents an index of wetness (i.e., surface modeling the spatial distribution of soil moisture at
saturation) at each cell. The scale of values in the derived different resolutions, while single flow algorithms (i.e., D8)
TWI surfaces range from zero to ten, where zero indicates a produce TWI values with higher variance and skew. Thus,
dry area and ten represents a saturated area. The TWI was the FD8 multiple flow algorithm was selected as it allows
calculated based on the following formula derived by Beven for more realistic water flow in the upper areas of a
and Kirkby (1979): catchment where flow convergence and divergence have
more influence on overland flow due to shape of the
TWI ⫽ ln 1a / tanb2 (4) landscape.
DEM Cell Size (m) DEM Cell Area (m2) Neighborhood Filter Size Neighborhood Filter Area (m2)
2m 2 * 2 ⫽ 4 m2 11 * 11 22 * 22 ⫽ 484 m2
5m 5 * 5 ⫽ 25 m2 5 * 5 25 * 25 ⫽ 625 m2
10 m 10 * 10 ⫽ 100 m2 3 * 3 30 * 30 ⫽ 900 m2
20 m 20 * 20 ⫽ 400 m2 3 * 3 60 * 60 ⫽ 3,600 m2
The PEI, CHM, and TWI values were extracted into tabular resolutions; however, regardless of the use of transforma-
format at each DEM resolution based on the cells correspon- tions, the PEI data failed the Brown-Forsythe test for homo-
ding to each plot. The average TWI and CHM value was geneity of variances. The same pattern was evident for the
extracted per plot, while bilinear interpolation was used to 10 m and 20 m CHM data and the 2 m TWI data (Table 2).
extract the PEI value for each plot based on the four closest Simple and multiple linear regressions were used to
cells surrounding the centre of each plot. As such, the CHM determine the relationships between the lidar-derived
and TWI values are representative of the entire plot at each variables and the seasonal soil moisture data. Regression
scale (i.e., 400 m2), whereas the PEI values are scale dependant analyses were conducted using the 2 m, 5 m, 10 m, and
and calculated based on an area which ranges from 16 m2 to 20 m resolution data and the following lidar variables: TWI,
1,600 m2 around the plot center. Other terrain surfaces such as CHM, and PEI. The CHM and PEI values were regressed against
slope gradient, slope aspect, slope curvature, surface flow the seasonal moisture measurements at each spatial resolu-
direction, and surface flow accumulation were extracted from tion, while the TWI values were regressed against the
the DEM datasets; however, they were eliminated from further moisture data at each resolution to determine the best
analysis due to non-normal distributions and/or collinearity. output based on: (a) the two different hydrological process-
ing techniques used (i.e., raw data versus IRA-altered data),
Statistical Analysis and (b) the seasonal measurement of soil moisture.
All the soil moisture data and lidar-derived surface layers
were tested for normality using the Shapiro-Wilk test for
normality (n ⬍ 2000) and the Brown-Forsythe test for Results
homogeneity of variances (i.e., homoscedasticity). The
Brown-Forsythe test is considered to be more robust than Seasonal Trends in Soil Moisture
the Levene test when the treatment groups are unequal in Soil moisture measurements at each plot in the 0 to 40 cm
size (Brown and Forsythe, 1974). Repeated measures analy- layer varied significantly (p ⬍0.05) over the season
sis of variance (RMANOVA) was used to assess the differences (RMANOVA) (F[1.741, 71.386] ⫽ 4.774, p ⫽ 0.015), but
between early-, mid- and late-season soil moisture at each seasonal differences in the 0 to 15 cm layer measurements
plot and the statistical significance level was adjusted using were not significant (F[1.560, 63.970] ⫽ 3.124, p ⫽ 0.063).
the Bonferroni correction. Huyhn-Feldt estimates (epsilon ⬎0.75) were used to correct
The three moisture measurements (i.e., N, SE, and SW) the degrees of freedom as Mauchly’s test indicated that the
were pooled for each plot and a square root transformation assumption of sphericity was violated (chi-square ⫽ 8.532
was applied to the 0 to 15 cm and 0 to 40 cm soil moisture and 15.462, p ⬍0.05) for both the 0 to 15 cm and 0 to 40 cm
measurements to normalize the data. A constant value of moisture measurements throughout the field season. Post-
one was added to the original data for the 0 to 15 cm hoc tests conducted on the 0 to 40 cm seasonal moisture
measurements prior to transformation, to ensure that measurements revealed that soil moisture was significantly
consistent scaling was applied during the square root higher in the early season compared to late season
transformations. This process was not necessary for the 0 to (p ⫽ 0.024). However, neither the early season (p ⫽ 0.602)
40 cm data since the raw values were already greater than or late season (p ⫽ 0.098) differed significantly from the
one. Square root transformations were also applied to the mid-season conditions (Table 3).
5 m raw TWI, 20 m raw TWI, and 20 m IRA-altered TWI data
layers; while the 2 m raw TWI and IRA-altered TWI data TWI and Soil Moisture
layers were normalized using the Winsorizing technique. The regression analyses comparing TWI and soil moisture
The Winsorizing technique normalizes the data by changing indicate that the 2 m, 5 m, and 10 m data explain the most
the value of an extreme outlier to make it equivalent with variation in soil moisture at both depths during the mid-
the next closest value within the inter-quartile range (Dixon, season sampling period; whereas, the 20 m resolution
1960; Wainer, 1976). The 2 m data had an outlier at plot explains more variability in the 0 to 15 cm measurements
MW14, which is located very close to Opishing Lake, with during the early season and more variability in the 0 to
shallow soils and a high water table. Regression analysis 40 cm measurements during the late season (Table 4). The
performed with and without the outlier plot (MW14) 5 m IRA-altered TWI data provide the best overall correlations
indicated that both relationships were statistically significant at each time interval and depth, with the best correlations
(p ⬍0.05); however, the regression relationships were occurring during the mid-season measurements where
reduced when the outlier was removed. Thus, the Winsorizing 29.2 percent and 34.6 percent of the variability is explained
technique was used to treat the 2 m data and all plots were for the 0 to 40 cm and 0 to 15 cm moisture values,
included in subsequent analyses. respectively (p ⬍0.001) (Figure 3).
It should be noted that the late-season moisture meas-
urements and the 20 m CHM failed the test for normality, but TWI, PEI, and CHM and Soil Moisture
had Shapiro-Wilk test values close to one (Table 2). These Based on the results from the TWI regressions at each
data were still used in the analysis with the acknowledge- resolution, the three surface layers (TWI, CHM, and PEI) were
ment that they violate the assumption of normality, so the regressed against the seasonal moisture values that provided
regressions utilizing these variables should be interpreted the best correlations with TWI. The 2 m, 5 m, and 10 m data
with caution. Normality was achieved for the PEI data at all were regressed against the mid-season moisture, whereas the
TABLE 3. PAIRWISE COMPARISONS OF SEASONAL MOISTURE MEASUREMENTS resolution data provided better predictions for the 0 to 40 cm
moisture measurements, while the 2 m and 5 m data had
Mean better predictions for the 0 to 15 cm measurements.
Difference P-Value
TWI .154 .133 .010* .202 .182 .003* .199 .181 .002* .189 .170 .003* .186 .167 .003* .129 .109 .015*
IRA-altered TWI .221 .201^ .002* .322 .305^ ⬍.001* .308 .292^^ .000* .361 .346^^ ⬍.001* .287 .271^ ⬍.001* .270 .254^ ⬍.001*
10 m Resolution
TWI .168 .148^ .007* .170 .149^ .007* .231 .214^ .001* .188 .169^ .003* .212 .194^ .001* .125 .105^ .017*
IRA-altered TWI .138 .116 .016* .150 .128 .011* .210 .191 .002* .178 .159 .004* .183 .164 .003* .113 .092 .024*
20 m Resolution
TWI .110 .087 .032* .104 .082 .037* .130 .110 .015* .090 .069 .046* .160 .140 .006* .086 .064 .051
* * * * *
IRA-altered TWI .222 .203^ .002 .189 .168^ .004 .236 .218^ .001 .157 .137^ .007 .272 .255^ ⬍.001 .165 .145^ .006*
*
^ ⫽ best regression at each resolution per depth class by season, ^^ ⫽ best overall regression, ⫽ significant relationship (p⬍0.05)
D e c e m b e r 2 0 1 2 1247
(a) (b)
Figure 3. Regressions of 5 m IRA-altered TWI against Soil Moisture: (a) represents the 0 to 15 cm soil
moisture measurements, and (b) represents the 0 to 40 cm soil moisture measurements.
coefficients of determination between soil moisture and TWI weaker regressions observed during the significantly wetter
can range from 0.08 to 0.43 at 0 to 30 cm soil depth based early-season measurements (Table 4).
on wet and dry years, respectively (Schmidt and Persson, The IRA-altered TWI generally explained more of
2003). Similarly, Western et al. (1999) found that terrain the variability in soil moisture than the raw TWI values at the
indices derived from a 5 m DEM and soil moisture in the top 2 m, 5 m, and 20 m resolutions throughout the season. The
30 cm of the soil profile explained little of the variation 10 m data do not follow this trend, as they consistently have
during dry conditions and up to 43 percent of the variation slightly higher coefficients of determination when using the
during wet conditions over an entire year. Their results raw TWI values (Table 4). The inverse trend observed with
indicate that the TWI explains the most variation when the the 10 m data is of interest because it provides the best raw
soil is wet and the catchment has experienced significant predictions of 0 to 40 cm soil moisture when compared to
lateral redistribution across its surface; however, when the values at other resolutions. Based on this relationship, it
soil is extremely wet the correlation between TWI and soil could be assumed that the 10 m data would become the best
moisture became weaker as soil porosity is more important overall predictor when the IRA is applied. However, this
for controlling soil moisture during these conditions (West- spatial correspondence did not occur; but rather the 5 m data
ern et al., 1999). This explanation corresponds well to the that showed the most significant improvements in correla-
2 m Surface Layers
IRA-altered TWI .196 .177^ .002* .229 .211 .001*
IRA-altered TWI ⫹ PEI .201 .163 .009* .256 .220^ .002*
IRA-altered TWI ⫹ CHM .199 .161 .009* .239 .203 .003*
IRA-altered TWI ⫹ PEI ⫹ CHM .204 .145 .024* .262 .208 .006*
5 m Surface Layers
IRA-altered TWI .308 .292 ⬍.001* .361 .346 ⬍.001*
IRA-altered TWI ⫹ PEI .325 .293^^ ⬍.001* .394 .365^^ ⬍.001*
IRA-altered TWI ⫹ CHM .309 .276 ⬍.001* .371 .341 ⬍.001*
IRA-altered TWI ⫹ PEI ⫹ CHM .327 .278 .001* .406 .362 ⬍.001*
10 m Surface Layers
TWI .231 .214 .001* .188 .169 .003*
TWI ⫹ PEI .248 .212 .003* .200 .162 .009*
TWI ⫹ CHM .308 .275 ⬍.001* .287 .253 .001*
TWI ⫹ PEI ⫹ CHM .335 .287^ .001* .310 .259^ .002*.
20 m Surface Layers Late-Season Moisture (0 to 40cm) Late-Season Moisture (0 to 15cm)
IRA-altered TWI .272 .255 ⬍.001* .165 .145 .006*
IRA-altered TWI ⫹ PEI .318 .286^ ⬍.001* .227 .190^ .004*
IRA-altered TWI ⫹ CHM .272 .237 .001* .168 .129 .021*
IRA-altered TWI ⫹ PEI ⫹ CHM .319 .269 .001* .229 .172 .013*
Conclusions
Ecological land classification distinguishes forest types based References
on distinctions between soil moisture and nutrient regimes. Agnew, L.J., S. Lyon, P. Gerard-Marchant, V.B. Colllins, A.J. Lembo,
Since a better spatial understanding of the ecological vari- T.S. Steenhuis, and M.T. Walter, 2006. Identifying hydrologi-
ability in soil moisture can facilitate mapping, planning, and cally sensitive areas: Bridging the gap between science and
application, Journal of Environmental Management, 78:63–76.
forest management practices, lidar-derived indices, specifi-
cally the TWI, PEI, and CHM, were evaluated to determine how Axelsson, P., 1999. Processing of laser scanner data - Algorithms
and applications, ISPRS Journal of Photogrammetry and Remote
effectively they model soil moisture patterns. The TWI Sensing, 54:138–147.
represents surface saturation based on surface morphology of
Beven, K.J., and M.J. Kirkby, 1979. A physically based, variable
the DEM, the PEI represents relative landscape position, and contributing area model of basin hydrology, Hydrological
the CHM represents the height of the forest canopy at any Sciences Bulletin, 24(1):43–69.
given cell. The TWI and PEI contributed more significantly to Brown, M.B., and A.B. Forsythe, 1974. Robust tests for equality of
the regressions with soil moisture; however, due to the variances, Journal of the American Statistical Association,
complexity of modeling soil moisture patterns in heteroge- 69:364-367.
neous forested landscapes, the TWI performed with mixed Canadian Forest Service, 2010. The Atlas of Canada: Forest Danger
results but displayed some interesting trends with respect to Rating, 2009. URL: http://atlas.nrcan.gc.ca/auth/english/maps/
seasonal moisture variability, soil depth and DEM resolution. environment/naturalhazards/forest_fires/firedangerrating, Natural
It was observed that the 2 m, 5 m, and 10 m TWI Resources Canada, (last date accessed: 23 September 2012).
explained the most variation during the mid-season meas- Canadian Interagency Forest Fire Centre (CIFFC), 2010. 2009 Canada
urements, while the 20 m TWI data explained the most Report. URL: http://www.ciffc.ca/images/stories/pdf/2009_Canada_
variation during the early- and late-season measurements. Report.pdf, Canadian Interagency Forest Fire Centre, (last date
The best overall soil moisture predictions occurred at 5 m accessed: 23 September 2012).
spatial resolution using both the IRA-altered TWI and the PEI Dixon, W.J., 1960. Simplified estimation from censored normal
to predict soil moisture in the top 15 cm of the soil profile samples, The Annals of Mathematical Statistics, 31(2):385–391.
where they explained 36.5 percent of the variation in the Ecological Land Classification (ELC) Working Group, 2009. Ecosites
measured soil moisture during the mid-season. In general, it of Ontario, Ontario Ministry of Natural Resources, Ontario,
was observed that the 10 m and 20 m spatial resolution data Operational Draft: 29 April 20.
gave better predictions for the 0 to 40 cm moisture measure- Ecological Stratification Working Group, 1995. A National Ecologi-
ments, while the 2 m and 5 m data had better predictions cal Framework for Canada, Appendix 1: Narrative Descriptions
for the 0 to 15 cm measurements. This indicates that coarser of Terrestrial Ecozones and Ecoregions of Canada, Number 96:
Abitibi Plains, Agriculture and Agri-Food Canada, Research
resolutions may be more suited to modeling soil moisture Branch, Centre for Land and Biological Resources Research and
trends at greater depths, while finer resolutions may be Environment Canada, State of the Environment Directorate,
more adept at resolving trends at shallow depths. The Ecozone Analysis Branch, Ottawa-Hull, Ontario, pp. 67–68.
results also indicate that the IRA depression removal algo- Environment Canada, 2010. Monthly Data Report for 2009, Tim-
rithm significantly increased the prediction efficiency for mins, Ontario, URL: http://www.climate.weatheroffice.gc.ca/
the 2 m, 5 m, and 20 m resolution data when compared to climateData/monthlydata_e.html?timeframe53&Prov5XX&
the raw lidar DEM surfaces. These results support the use of StationID54180&Year52009&Month51&Day51, National
the IRA depression removal algorithm to improve the Climate Data and Information Archive, Environment Canada
quality of a DEM for terrain modeling, particularly when (last date accessed: 23 September 2012).
cell connectivity is important for the end uses of the DEM. Florinsky, I.V., R.G. Eilers, G.R. Manning, and L.G. Fuller, 2002.
Lidar data provide superior capacity for deriving high- Prediction of soil properties by digital terrain modelling,
resolution terrain variables that are related to hydrology, Environmental Modelling & Software, 17:295–311.
surface flow, soil moisture retention, and exposure. This Freeman, T.G., 1991. Calculating catchment area with divergent
capacity offers enhanced potential for modeling of soil flow based on a regular grid, Computers & Geosciences,
moisture, a variable that is significant within the ecosite 17(3):413–422.
classification framework at site and landscape scales as well Gessler, P.E., I.D. Moore, N.J. McKenzie, and P.J. Ryan, 1995. Soil-
as for the modeling of ecosystem productivity. Greater landscape modelling and spatial prediction of soil attributes,
International Journal of Geographical Information Science,
understanding of ecosite and site productivity potential will 9(4):421–432.
assist forest managers in their decision making related to
Gessler, P.E, O.A. Chawick, F. Chamran, L. Althouse, and K.
harvesting plans and silvicultural treatments. Although the Holmes, 2000. Modeling soil-landscape and ecosystem proper-
relationships observed in this study support these concepts, ties using terrain attributes, Soil Science Society of America
further exploration of the potential of these precise terrain Journal, 64:2046–2056.