ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing

journal homepage: www.elsevier.com/locate/isprsjprs

Monitoring Andean high altitude wetlands in central Chile with seasonal

optical data: A comparison between Worldview-2 and Sentinel-2
imagery
Rocío A. Araya-López a,b, Javier Lopatin a,⇑, Fabian E. Fassnacht a, H. Jaime Hernández b
a
Institute of Geography and Geoecology, Karlsruhe Institute of Technology (KIT), Kaiserstraße 12, 76131 Karlsruhe, Germany
b
Laboratory of Geomatics and Landscape Ecology, Faculty of Forest and Nature Conservation, University of Chile, 11315 Santa Rosa, Santiago, Chile

a r t i c l e i n f o a b s t r a c t

Article history: In the Maipo watershed, situated in central Chile, mining activities are impacting high altitude Andean
Received 30 September 2017 wetlands through the consumption and exploitation of water and land. As wetlands are vulnerable
Received in revised form 24 February 2018 and particularly susceptible to changes of water supply, alterations and modifications in the hydrological
Accepted 4 April 2018
regime have direct effects on their ecophysiological condition and vegetation cover. The aim of this study
Available online xxxx
was to evaluate the potential of Worldview-2 and Sentinel-2 sensors to identify and map Andean
wetlands through the use of the one-class classifier Bias support vector machines (BSVM), and then to
Keywords:
estimate soil moisture content of the identified wetlands during snow-free summer using partial least
One-class classifier
Soil moisture
square regression.
Bias support vector machines (BSVM) The results obtained in this research showed that the combination of remote sensing data and a small
Partial least square regression (PLSR) sample of ground reference measurements enables to map Andean high altitude wetlands with high
Wet meadows accuracies. BSVM was capable to classify the meadow areas with an overall accuracy of over
78% for
both sensors. Our results also indicate that it is feasible to map surface soil moisture with optical remote
sensing data and simple regression approaches in the examined environment. Surface soil moisture esti-
mates reached r2 values of up to 0.58, and normalized mean square errors of 19% using Sentinel-2 data,
while Worldview-2 estimates resulted in non-satisfying results. The presented approach is particularly
valuable for monitoring high-mountain wetland areas with limited accessibility such as in the Andes.
Ó 2018 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier
B.V. All rights reserved.

1. Introduction Andean wetlands are particularly susceptible to changes of water

supply and corresponding alterations and modifications in the
Andean high altitude wetlands (‘vegas’ or wet meadows) are hydrological regime which were found to have direct effects on
usually located in humid areas of the Andes in north-central Chile vegetation cover (Ahumada et al., 2011; Ahumada and Faúndez,
(Squeo et al., 2006a, 2006b). These wetlands are important key 2009). In order to understand these ecosystems better, as well as
components of the Andean ecosystems, providing several ecosys- for conservation planning and an efficient management of
tem services like water supply and habitat for wildlife and live- resources, there is a strong need for a spatially explicit and up-
stock (Ahumada and Faúndez, 2009; Contreras, 2007; García and to-date inventory of Andean wetland ecosystems. This need is con-
Otto, 2015). However, Andean high altitude wetlands have suffered trasted by a lack of baseline datasets and limited knowledge on
a strong degradation and transformation in the last decades, espe- these habitats (Muñoz-Schick et al., 1997).
cially as a consequence of agriculture and mining activities Remote sensing (RS) data has proven to be a useful source of
(Cepeda-Pizzaro and Pola, 2013; Squeo et al., 2006a,b,c). information for mapping and monitoring wetland ecosystem at dif-
In the Maipo watershed located in central Chile, mining activity ferent temporal and spatial scales (e.g. Mahdianpari et al., 2017;
is the main factor impacting Andean wetlands through the con- Betbeder et al., 2015; Adam et al., 2010; Ozesmi and Bauer, 2002;
sumption and exploitation of water and land (GORE-RMS, 2013). Rundquist et al., 2001), and has been considered a valuable tool
for supporting their conservation (Baker et al., 2006; Davranche
⇑ Corresponding author. et al., 2013; MacAlister and Mahaxay, 2009; Whiteside and
E-mail address: javier.lopatin@kit.edu (J. Lopatin). Bartolo, 2015). RS allows obtaining objective information in remote

https://doi.org/10.1016/j.isprsjprs.2018.04.001
0924-2716/Ó 2018 International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). Published by Elsevier B.V. All rights reserved.

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
2 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

and isolated areas such as the high altitude Andean basins. Accurate during the first step by applying empirical regression algorithms.
and up-to-date mapping, inventory and assessment of wet mead- Such an approach may become possible because in the Andes wet-
ows is extremely important to assess changes in wetlands’ vegeta- lands the SM gradient is usually related to a vegetation cover gra-
tion composition, cover and water content due to natural or dient (Squeo et al., 2006a) which should be detectable by optical
anthropogenic disturbances over time (Ozesmi and Bauer, 2002; remote sensing data (Gao et al., 2013). The combination of both
Rundquist et al., 2001; Wang et al., 2015). approaches could be the basis for a monitoring system that is able
Soil moisture (SM) is one of the key parameters related to wet- to detect early responses of Andean wetlands to natural or anthro-
land ecology that can be estimated by RS data. SM is usually esti- pogenic disturbances.
mated from data delivered by passive microwaves sensors such
as SMOS (e.g., Kerr et al., 2012), sometimes disaggregated to higher 2. Materials and methods
resolution using thermal data and vegetation indices (Merlin et al.,
2012b; Molero et al., 2016, Wang et al., 2016). Another alternative 2.1. Study area
is the use of active radar systems (e.g. Zribi et al., 2005; Zribi and
Decharme, 2008). However, approaches using passive microwave The study area is situated in the high-alpine areas of the Maipo
sensors mostly deliver SM information at coarse resolution watershed in the Metropolitana region of Chile (Fig. 1A and B).
(
100–4000 m), and the success of active radars with suitable spa- Fieldwork was conducted on the Estero La Engorda sub basin
tial resolution (nowadays for example globally available through (Fig. 1C; 6259600 S407100 W) which covers an area of approxi-
Sentinel-1) depends greatly on the topography and the applied fre- mately 60,000 m2. This area is mainly characterized by arid
quency. In mountainous areas with high terrain roughness a Mediterranean climate, with precipitations occurring mainly as
proper calibration of active radar data can be challenging snow during the winter. Summers are dry, with occasional rain-
(Luckman, 1998; Pasolli et al., 2012) and due to the oblique acqui- storms. Currently, the wetlands located in La Engorda are being
sition mode of active radar sensors notable terrain shadows can disturbed through the construction and installation of weirs and
occur. Concerning suitable frequencies for estimating soil mois- temporary roads supporting a hydroelectric installation nearby.
ture, particularly low microwave frequencies (i.e. P- to L-band) Wet meadows are dominated by species with rhizome growth
have demonstrated to be highly related to soil moisture (Pasolli form and small grasses (less than 40 cm in height), with dominant
et al., 2012). However, C-band data – as provided by Sentinel-1 – species from the genera Carex and Scirpus. This type of vegetation is
has generally been found to be less correlated with soil moisture characterized as zonal (local) vegetation and it is determined by
due to the challenging separation of the moisture signal from inter- factors like precipitation, altitude and slope (Ahumada and
fering signals caused by the roughness of vegetation and the soil Faúndez, 2009) but also depends on soil properties and humidity.
(Zribi et al., 2005). Recent studies therefore combined Sentinel-1 Locally occurring soils can be divided into three types: (1) continu-
data with time series of optical vegetation indices (e.g. NDVI or ously flooded histosols (presenting low bulk densities), (2) entisols
EVI) and extensive field work to improve soil moisture estimates in the transitional wetness areas, with higher organic matter and
by isolating the vegetation effect in the radar backscatter (e.g. periodic flooding (intermedium bulk densities), and (3) entisols in
Paloscia et al., 2013; Gao et al., 2017; Alexakis et al., 2017). While the outer areas, which are rarely flooded (high bulk densities).
these procedures hold potential, they are associated with an
increased processing effort related to the data fusion and might 2.2. Ground reference data
still not be fully suitable in areas with rough topography due to
the mentioned complications associated with the processing of During the field survey, soil moisture was measured at 30
active radar data under such conditions. Based also on the earlier stratified random points samples during the summer period of
reported potential of short-wave infrared (SWIR) information to 2015–2016, with repetitive measurements taken in December,
estimate soil moisture (e.g. Sadeghi et al., 2017) we here examine January, February and March. The time period was limited to those
the sole use of optical Sentinel-2 data as a potentially straightfor- four months to minimize the influence of snow and short-term
ward alternative to estimate soil moisture content within the effects of snow melting.
Andean highlands. The high spatial resolution of Sentinel-2 may At each sampling point, soil moisture measurements were
enable the estimation of SM at local scales, which is key when obtained from soil samples taken by a cylinder with a diameter
mapping non-extensive isolated wetlands. The Andean wetlands of 52 mm (first 15 cm of soil). The soil samples were stored in plas-
are located in the highlands of the Andes and usually do not occupy tic bags inside coolers to avoid moisture loss by evapotranspira-
large extensions of land as they depend on special micro-relief tion. In the laboratory, we followed the procedure of Sadzawka
characteristics which only occur locally and are hence a vivid et al. (2006) to estimate the gravimetric soil moisture (g/g), where
example of such small and isolated wetland habitats. the soil samples were dried at 105 ± 5 °C for 48 h. Table 1 depicts
The aim of this study was to analyze the potential of high spa- the descriptive statistics of each field campaign.
tial resolution optical data, i.e. WorldView-2 and Sentinel-2, for
mapping and monitoring Andean wetlands of central Chile. To 2.3. Remote sensing data
achieve this, we propose the use of a one-class-classifier to map
the occurrences of Andean wetlands using sparse ground data. We acquired WorldView-2 (WV-2) images from December
Such an approach has been applied earlier for other ecosystems 2015 and March 2016, scheduled in correspondence with the field
(e.g. Mack et al., 2016; Stenzel et al., 2017) and we deem this campaigns (Table 2). WorldView-2 collects multispectral informa-
approach particularly valuable in our study area which is geo- tion in 8 bands, covering the spectral ranges between visible (VIS)
graphically isolated and where field data is hard to acquire. With and the near-infrared (NIR; Table 2). The data were delivered in
the suggested work-flow we aim to create baseline data to comple- radiance units, and were radiometrically corrected to reflectance
ment the Chilean Wetland National Inventory, which is known to at the top of the atmosphere (TOA) using ENVIÒ (Exelis Visual
underestimate the extent of Andean wetlands (MMA-Centro de Information Solutions, Boulder, CO, version 4.7) and the method
Ecología Aplicada, 2011) and hence hampers their proper conser- proposed by Updike and Comp (2010). Then, atmospheric correc-
vation and management. tions were carried out for both WV-2 images by applying the dark
Secondly, we tested the suitability of the two satellite systems object subtraction method (Chavez, 1988) to convert the TOA
to estimate SM within the areas identified as Andean wetlands reflectance to bottom of the atmosphere reflectance (BOA).

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx 3

Fig. 1. Study area in Estero la Engorda, Región Metropolitana, Chile. Areas delineated with black lines correspond to the locations of Andean wetlands as listed in the national
wetland inventory of the year 2011 (MMA-Centro de Ecología Aplicada, 2011). Dark dots show the location of the soil moisture sampling locations during the summer season
and the black line crossing the main wetland shows the location of the constructed road.

Table 1
Descriptive statistics of the soil moisture samples (g/g) collected during the summer season.

2015 2016
December January February March
Minimum 0.03 0.01 0.01 0.01
Median 0.27 0.34 0.27 0.23
Mean 1.02 1.11 1.06 1.08
Maximum 8.41 6.72 8.65 11.45
Coefficient of variation (%) 181 163 182 215

Sentinel-2 (S-2) Level-2A imageries (BOA reflectance) were

Table 2 acquired through the ‘‘Copernicus Scientific Data Hub” webpage.
Satellite sensor characteristics and image acquisition dates. We acquired imagery for the four dates most closely matching
Sensor type Acquisition date Band Spectral range used
the dates of the field campaigns (Table 2). S-2 has 13 bands ranging
number (resolution [m]) from the VIS to the short-wave infrared (SWIR) with different spa-
tial and spectral resolutions. We used only the 10 and 20 m bands
Worldview-2 20 December 2015 1 Coastal: 400–450 nm (2)
20 March 2016 2 Blue: 450–510 nm (2) due to the small distance between field sampling points (Table 2).
3 Green: 510–581 nm (2) The bands at 20 m were resampled to 10 m using the Sentinel-2
4 Yellow: 585–625 nm (2) Toolbox of SNAP (Sentinel Application Platforms, version 5.0). Fur-
5 Red: 630–690 nm (2)
thermore, geometric corrections were performed in ENVIÒ using
6 Red edge: 705–745 nm (2)
7 NIR 1: 770–895 nm (2)
the same method described above.
8 NIR 2: 860–1040 nm (2) Finally, all images from both sensors were brightness normal-
Sentinel-2 23 December 2015 2 Blue: 398–594 nm (10)
ized using the method of Feilhauer et al. (2010) (implemented in
25 January 2016 3 Green: 515–605 nm (10) R) to maximize spectral differences of vegetated areas.
11 February 2016 4 Red: 626–702 nm (10)
23 March 2016 5 Red edge: 695–710 nm (20)
6 NIR 1: 730–740 nm (20)
2.4. Mapping Andean high wetlands
7 NIR 2: 768–796 nm (20)
8 NIR 3: 690–980 nm (10) 2.4.1. Classification of Andean wetlands
8a NIR 4: 848–881 nm (20) To map the current extent of Andean wetlands we applied the
11 SWIR 1: 1470–1756 nm (20)
one-class classifier biased support vector machines (BSVM; Liu
12 SWIR 2: 1960–2444 nm (20)
et al., 2003). BSVM is an adaption of the classical formulation of

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
4 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

Support vector machines (SVM) to handle unlabelled data (Liu homogenization of the spatial and spectral resolution was per-
et al., 2003). In contrast to the supervised classifier, the training formed, where: (1) the resolution of WV-2 was resampled (‘bilin-
data of the one-class classifier only contains labeled samples from ear’) to 10 m to match S-2, and (2) only 6 bands of S-2 (the ones
the class of interest (positive class), which reduces the amount of closest to the wavelength of WV-2) were used to match the spec-
required in situ information significantly (Li et al., 2011; Song tral characteristics of WV-2.
et al., 2016). Instead of labeled negative samples, BSVM uses unla- We used partial least square regression (PLSR; Wold et al.,
beled samples with unknown class label as negative training sam- 2001) trained with the reflectance values and the log transformed
ples. As it is not feasible to use all the unlabeled pixels of an image, field-measured SM estimates (due to the strong left-skewness of
a randomly selected subset is used instead (‘background’ class; Li the measurements; Table 1) to predict SM across all areas identi-
and Guo, 2014; Mack et al., 2014). This algorithm has been used fied as wetlands during the preceding mapping procedure (see Sec-
successfully to map grasslands and wetland ecosystems using tion 2.4.3). PLSR is a well-known algorithm that summarizes the
remote sensing data before (e.g. Mack et al., 2016; Stenzel et al., predictors’ information into fewer less correlated components
2017). (Wold et al., 2001).
We used the Chilean wetland inventory of the year 2011 To assess potential overfitting in the individual PLSR models
(MMA-Centro de Ecología Aplicada, 2011) to obtain 50 positive and to obtain distributions of the model accuracies, we imple-
samples for training the BSVM. The positions of the training sam- mented an iterative validation procedure based on a bootstrap
ples were randomly selected from inside the areas indicated as split. As described in Roberts et al. (2017), treating multiple tempo-
wetland in the inventory data. 100 more random pixels were ral observations collected at the same location as independent
selected from the whole study area extent to provide samples for samples during cross-validation may lead to overly optimistic
the ‘background’ class. This process based on a completely random model results. We hence temporally blocked our samples during
selection, hence, the background class is likely to contain some pix- the iterative validation procedure. This means that the up to four
els from the positive class. The obtained classification maps were field measured soil moisture values collected at a single location
validated using the 30 field-visited plot locations as positive sam- can always only either be all four in the training data or all four
ples and an additional set of 100 random samples for the back- in the validation data. This resulted in the following implementa-
ground class. We ran eight classifications (four datasets and two tion: in each iteration, 30 samples were randomly drawn with
dates). As datasets we used S-2 data at 10 m pixel size (10 bands), replacement from the 30 samples (i.e. the physical plots) available
S-2 data at 10 m pixel size (6 bands), WV-2 data at 2 m pixel size (8 for one time period (month) to build a training dataset. During this
bands), WV-2 data at 10 m pixel size (6 bands) (see Section 2.4.2 procedure, approximately a 36.8% of the samples were not selected
for further information); as dates we selected December and March at all (Kohavi, 1995). These samples were subsequently used as
as for these two months datasets from both sensors were available. holdout samples for the independent validation. The same combi-
We applied BSVM with a radial basis kernel (e.g. Mack et al., nation of training and validation points were then also sampled
2016; Stenzel et al., 2017) and the parameters ‘sigma’, ‘Cv’ and from the remaining time periods to complete the dataset while
‘Cb’ were tuned by a 10-fold cross-validation to find the best model ensuring that all samples from an individual (physical) plot are
based on the training data. ‘Sigma’ is a function related to the ker- always either exclusively in the training set or exclusively in the
nel, ‘Cv’ is the penalty associated to classification errors and ‘Cb’ is validation set. We then used the training dataset to train PLSR
the relative cost of the errors inside the ‘background’ class. To models which were internally optimized by a backward-selection
internally select the best model, we considered two accuracy met- of predictors (Martens and Martens, 2000). The optimal number
rics often used in combination with one-class-classifiers: (a) the of components was automatically determined by comparing the
probability of positive prediction (PPP), and (b) the true positive internal leave-one-out cross-validated root mean square error
rate (TPR; Mack et al., 2014). The model which maximized the (RMSEval) of models with different numbers of components.
TPR while minimized the PPP was selected. After finding the optimal model based on the training data, the
As output BSVM delivers a continuous value for each classified trained model was applied to the hold out sample to calculate the
pixel that describes the distance from the hyperplane. The hyper- coefficients of determination (r2; calculated as the squared
plane defines the mathematical rule created by the SVM algorithm Pearson’s correlation coefficient), the normalized root mean square
according to which unknown samples are allocated to either the error (nRMSE) and the bias between predicted and observed SM of
positive or the negative class. Positive distances from the hyper- the holdout samples.
qffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi nRMSE was calculated
 as
plane indicate a class-membership with the target class (in our case Pn  2ffi
nRMSE ¼ 1
n j¼1 y j  b
y j =½ maxð SM Þ  min ðSM Þ  100, where
wetlands) (see Mack, 2017 for an illustrative example). To separate
the positive class from the negative class (or ‘background’) we n is the number of samples and SM is soil moisture. Likewise, the
applied a distance threshold > 0.5. This threshold slightly deviates bias of prediction was measured as one minus the slope of a regres-
from the standard threshold of 0, but was found to deliver more bal- sion without intercept of the predicted versus observed values
anced results in the given case. The validity of the threshold was (Lopatin et al., 2016). The R-package ‘autopls’ was applied for the
visually confirmed using the analytical graphs provided by the ‘one- PLSR analysis (Schmidtlein et al., 2012).
Class’ package (see Appendix A; Mack et al., 2014). We tested other regression algorithms, such as random forest,
support vector machines and K-nearest neighbors, that did not
2.4.2. Predicting soil moisture result in an improvement of the model, hence their results are
Soil moisture (SM) was modelled using WV-2 and S-2 imagery not presented in the manuscript.
separately. In case of WV-2, only the field data obtained in Decem- Finally, we applied a one-sided bootstrapping test (Lopatin
ber and March were used and combined with the reflectances of et al., 2016, 2017) to check for significant differences (a = 0.05; in
the two WV-2 images extracted at the plot locations (two times terms of R2 and nRMSE) in the model results obtained with
30 samples). For S-2, two datasets were used: (1) Reflectance val- WV-2 and S-2 data. We tested for significant differences between
ues and field measurements collected in December and March the classification results of the models trained with different
(two times 30 samples to compare with WV-2), and (2) Reflectance number of observations (S-2 models using 120 (samples from all
values and field data for all time periods (December, January, dates) and 60 observations (samples from December and
February and March, that is four times 30 samples). Furthermore, March), as well as for the sensors (homogenized models of WV-2
to compare also the radiometric characteristic of both sensors, a and S-2).

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx 5

2.4.3. Mapping soil moisture closest to the road intervention are highly saturated with water
Predictive maps of SM were calculated for the beginning and while the exterior areas show less soil moisture. Moreover, areas
end of the summer season (December and March) using the best depicting low to medium SM values were more stable during boot-
obtained model and the corresponding satellite scenes. We strapping (CV < 20%), while wetter areas were highly variable (CV >
restricted the SM predictions to areas that were identified as wet- 20%).
lands in the preceding mapping procedure.
To also account for the variability in SM predictions caused by
the composition of the training samples during the model building 4. Discussion
phase, we created a predictive map for each of the 500 iterations
(see Section 2.4.2). We then calculated the median and the coeffi- 4.1. Classification of Andean wetlands
cient of variation (CV, between 0 and 100, calculated as
CV p ¼ ðSDjp=meanðSM Þo Þ  100, where CVp is the pixel coefficient The results showed that it is feasible to accurately map the
of variation, SDp is the pixel’s standard deviation and mean(SM)O presence and extent of Andean wetlands by optical RS with few
is the mean SM value of all reference measurements) of SM values field efforts. The BSVM classifications delivered intermediate to
for each pixel. High values of CV indicate less stable SM predic- high accuracies for both sensors. For the WV-2 models, we
tions. As the model was trained using log transformed values of observed a tendency to overestimate the wetland area (false posi-
SM, the predicted maps were back-transformed (using the exp- tives of
48% for the December image). A visual comparison with
function) to obtain SM estimates in g/g. the official current wetland inventory of Chile (compare black
polygons in Fig. 2 with the areas identified as wetlands) also sug-
gests a general overestimation of wetland areas. However, we
3. Results
think that this error is likely to at least partly be explainable by
the incompleteness of the reference dataset which is known to
3.1. Identification of Andean wetlands
underestimate the number and the extent of Andean wetlands
(MMA-Centro de Ecología Aplicada, 2011; see Section 4.3 for an
Fig. 2 illustrates the areas identified as wetlands by the BSVM.
explanation of how these areas were mapped in the reference data-
The continuous values indicate the distance from the separating
set). On the other hand, such errors can also be attributed to the
hyperplane and can be interpreted as a measure of reliability of
structure of the so called ‘background’ class, where a random selec-
the classification, where higher values indicate a clearer member-
tion of pixels (including from the positive class) is performed. This
ship with the positive (wetland) class. The confusion matrices
process has been reported to lead to commission errors in earlier
(Table 3) show a general tendency for higher performances in
studies. For example, Skowronek et al. (2017) reported commission
March than in December. For March, S-2 obtained an overall accu-
errors of 34% when mapping the invasive species Phalaris aquatica.
racy (OA) of 85%, a producer accuracy (PA) of
80% and a user
Moreover, Stenzel et al. (2014) mapped alkaline fens and raised
accuracy (UA) of
92%, while WV-2 depicted lower accuracies
bogs with similar commission errors, while Mack et al. (2016)
(OA
85%, PA
77% and UA
64%). For the homogenized datasets
obtained high levels of overestimation of raised bogs (
75%). This
the differences in performances were less pronounced for Decem-
underlines the importance of the ‘background’ class when working
ber as compared to the original datasets (S-2 with an OA of 88 and
with the BSVM classifier. Recently, Mack et al. (2016) developed an
WV-2 with 83). For March, the performance difference according to
iterative process to ‘clean’ the ‘background’ class as much as possi-
the OA did not change in comparison to the original datasets.
ble from pixels of the positive class. This procedure improved their
From all pixels classified as wetlands (i.e.
241 ha), only
48%
classification results for mapping raised bogs by
26%. In our case,
and
75% obtained hyperplane distances over 0.5 for WV-2 and S-
the separation of the positive class (i.e. wetlands) from the rest of
2, respectively. WV-2 registers a decrease of the wetland area for
the landscape was deemed sufficiently accurate (i.e. overall accura-
March in comparison to December (from 42 ha to 33 ha), while
cies over
78% in all cases) without any previous iterative clearing
on the contrary S-2 accounted a slight increase of the wetland area
process; we reached an overall accuracy of over 90% using the S-2
from December to March (from 40 ha to 42 ha; Table 3).
data in March. We explain this with a clear spectral separability of
the wetlands from the remaining landscape. Not performing the
3.2. Soil moisture prediction iterative step saves a considerable amount of processing time
(Mack et al., 2016), which is an important factor to account for
In Table 4 the results of all tested models using PLSR and the when monitoring large areas.
iterative validation are presented. Generally, models using S-2 One advantage of BSVM is that the classification maps (Fig. 2)
imagery showed higher performances. The best model was can be displayed as continuous values depicting the distances to
obtained by including all 120 samples from the four field measure- the SVM hyperplane. Contrary to other binary classifiers that deli-
ment dates and images (median bootstrap r2 of 0.58 and nRMSE of ver result with sharp boundaries, BSVM hence allows for the detec-
18.96% using 4 PLSR components). The accuracy of the WV-2 tion of continuous patterns in the landscape, which agree more
model increased when enlarging the pixel size (to 10 m to match with the graduate gradients of water, nutrient and plant covers
S-2). Nevertheless, when matching the two sensor characteristics in nature (Reschke and Hüttich, 2014; Tiner, 2015).
(i.e., both with 10 m pixel and 6 bands with similar wavelengths), Here, we used single satellite scenes during all wetland clas-
S-2 models still obtained significantly (a = 0.05) higher fits than sifications. Similar approaches were presented by e.g. Lee and
the WV-2 models (median bootstrap r2 of 0.46 and nRMSE of Yeh (2009) and McCarthy et al. (2015), who used optical RS of
20.24% using 3 PLSR components). The whole bootstrapped distri- single satellite scenes to classify mangroves and coastal marshes,
butions of accuracies using the best model (i.e. S-2 with 120 obser- respectively. Contrary, most of the previous studies focusing on
vations) are presented in Fig. 3. The model shows that small and wetland mapping used phenological information to enhance
intermediate SM values tend to be slightly overestimated, while the classification (Adam et al., 2010; Davranche et al., 2010;
high SM values tend to be underestimated. Ozesmi and Bauer, 2002; Townsend et al., 2001). However, due
Finally, median and coefficient of variation (CV) prediction maps to the special geographical situation of our study sites, the phe-
for December and March are presented in Fig. 4. The patterns agree nology of the wetlands is mostly described by two stages: win-
with the field observations, where the central areas and areas ter, where most of the wetlands will be snow-covered and hence

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
6 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

A December 2015 March 2016

1.4

1.3

1.2

1.1

1.0

0.9

0.8

0.7

0 500 m 0.6

0.5

B
1.4

1.3

1.2

1.1

1.0

0.9

0.8

0.7

0.6

0.5
Fig. 2. Occurrences of Andean wetlands in the La Engorda valley using WorldView-2 (A) and Sentinel-2 (B) imageries from the beginning (December) and ending (March) of
the summer. The given values indicate the distance from the separating hyperplane. Black lined polygons show the wetlands presented in the Wetland National Inventory of
2011 (MMA-Centro de Ecología Aplicada, 2011). Areas with a distance to the hyperplane < 0.5 were masked out.

Table 3
Classification results of the Biased Support Vector Machine classifications for WorldView-2 and Sentinel-2 for the original and homogenized information. We applied a
hyperplane distance threshold > 0.5 to separate the wetlands from the background during all classifications.

Sensor Dataset Spatial resolution (m) Number of bands UA PA OA Area (ha)

Original information
WorldView-2 December 2 6 52 73 78 42
March 74 83 89 33
Sentinel-2 December 10 10 74 83 89 39
March 92 80 94 42
Homogenized information
WorldView-2 December 10 6 63 63 83 36
March 71 73 87 38
Sentinel-2 December 10 6 71 83 88 36
March 83 83 92 42

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx 7

Table 4
Summary of soil moisture (g/g) predictions by WV-2 and S-2 with PLSR. Median bootstrapping values are presented.

Model Sensor Dataset Number of observation Spatial resolution (m) Number of bands r2 nRMSE (%) Bias
Original information
1 WV-2 Dec/Mar 60 2 6 0.34 24.36 0.40
2 S-2 Dec/Mar 60 10 10 0.51 21.02 0.32
3 S-2 Dec/Jan/Feb/Mar 120 10 10 0.58* 18.96* 0.26
Homogenized information
4 WV-2 Dec/Mar 60 10 6 0.44 23 0.34
5 S-2 Dec/Mar 60 10 6 0.46* 21.24* 0.34
*
Denote a significant improvement (a = 0.05) of the model accuracy.

r2 = 0.58, nRMSE = 18.96%, bias = 0.26

●
4
●●
●●
●●
●
●
●●
●●●
●
● ●
●●
● ●
●
●
●
●● ●●●
●●
● ●
●
●
● ●●●●
●●
●
● ●
●
●
●
●
● ●● ●
● ● ● ●
●●
●●
●
● ●
●
●
●
●● ● ●
● ●●
●●
●
● ●
●
●
●
●●●●
● ●
● ●
●
●●●
●
●
●●
● ●
●
●
●
●
2 ●
●
● ●
● ● ●
● ●
●
●
●●●●
● ●
●
●
●
●
●
● ●
●●
●
●● ●
● ● ● ● ● ●
●●●●
●
●
● ●
●
●
●
●
●
● ●
●●
●
●
● ●●●●● ● ● ●● ●● ●●●
● ●
●●
●
● ●
●
● ●
●
●●●●
●
● ●
● ●
●
●
● ●● ●
● ●
●●●●
●●
●●
●●
● ●
●
●
●● ●
●
●
●
● ●
●●
●●
●
● ●●●
● ●
● ●●
●
●●
● ●●●
● ●
● ●●●●●
● ●
● ●
● ●
●
● ●
●
●●●●
●
●●●
●●
● ●●
● ●
●●●
● ●
●●●
●
●
● ●
●
●
●
●
●●
●
●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●
● ●
●
●
●
●●
●
●
●
● ●
●
●
●
●
●
●
● ●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●● ●
●
●●● ●
●
●
●●
●
● ●●
●
● ●●
● ●
●
●
●●
●● ●
● ●
●●
●
● ●
● ● ●
●
●
● ●
●
● ●●
●
●
●●●● ● ●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
●
● ●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●●
●
● ●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●●
●●
● ●●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
● ● ●
●●
● ●
●
●
●
●
●
●
●
●
●
●
●●
●
●● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
● ●
●
●
●●●
●●
●●
●
●
●
●
●
●
●
●● ● ●
● ●
●● ● ● ● ●●● ● ●
●●●
● ●●
●
● ●
●
●● ●●●
●●
● ●
●
●●
●●●●
●●
●
●●
●●●
●●●●
●
●
●
●●●●
●
●●
●
●
●
●
●●
●●
●
●●
●●●●
●●
● ●
●●
● ●
●
●●
●
●
●●
● ●
●
●●●●●
● ●
Predicted [log(g/g)]

0 ●
●●
●
●
●
●
● ●
●● ●●
●
●
●
●
●●●
● ●
●●●
●
●●
●●●●●
●
●
●●
●●
●
●●
●●
●
●
●
●
●
●●
●●
●
●
●● ●
●
●
●●
●
●●
●
●●
●
●
●●
●
●
●●
●●
●●
●
●
●● ●
●
●
●
●●
●
●
●
●●●
● ●●
●●●
●●
●
●●
●●
●
● ●
●
●
●●●●
● ●
●
●
●
●●
●
●
● ● ●●●
●● ●
●●
●
●
●
●●
● ●
●
●●●
●
●●
●
●
●●
●●●
●
●●
●
●●
●● ●
●●
●
●
●●●
●
●
●
●
●●
●
●
●●
●
●●
●
●
●● ●
●
●
●●
●
●
●●
● ●
●
●●●
●●
●
●
●●
●
●●●
● ●
●
●
● ●●●
● ●
●
●
●●
●
●
●
●
● ●
●
●
●
● ●
●
●
●
●
●
●
●●
●
●
●●
● ●
●
●
●
●
●
●●
●
●
●●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●●●
●● ●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●●
●
●
●
●
●●
●
●
●
● ●●
●●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
● ●
●●● ●
●●
●
●
●●
●
● ●
●
● ●
●●
●
●
●
●
●●
●
●●
●
●
●
●●●●
●●
●
●
● ●
●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●
●
●●
●●
●
●
●
●●
●
●
●
●●
●●●
● ●
●●●
●●
●
●
●
●
●
●
●●
●
●
●●
●●
●
●
●●
●
●
●
●
● ●
●
●●
●
●●
●
●
●● ●
●●
●●
●
●
●
●●
●
●●
●
●
● ●
●
●
●
●
●●
● ● ●
● ●
●●
●
● ●
● ●●
●
●●
● ●●●
●
● ●
●●●
●
●●
●
●
●●
●
●●●
●●
●
●
●●
●● ●
●●●●
●●
●
●●
●
●●
●
●● ● ●
● ●●● ●
● ● ● ●
●● ●
●●●
●●●
●
●
●
●
●
●
●●
●
●●
●
● ●
●●
●
●
● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●●
●
●●●
●
●
●
●
●
●
●●●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●●
●
●●●
●
●
●
● ●●
●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●
●
●
●
●
●
●●
● ●●
●●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
● ●
●
●● ●● ●●
●
●●●
●●●● ● ● ●● ●●●● ●
●●
●
●●● ●
● ●● ● ●
●
●●● ●
● ●●●
●● ● ●●
●●● ●●
● ●
●● ●●
● ●●
● ●
● ●
●
●
●
●
●● ●
●
●
● ●●
●
●
● ●
●●
●
●
●
●
●●●
●●
●●●
●●
●
●●
●
●●
●
●
●●●●
●
●
●
●●
●●
●
●
●
●●
●
●●
●
●
●
●
●●
●
●●
●●
● ●
●
● ●●
●
●●
●●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
● ●
●●
●
●
●
●● ●
● ●
●●●
● ●
●●●●●
●
● ●● ● ●● ●
●
●
●
●
●
●●●
● ●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●●
● ●●
●
●
●
● ●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●
●
●
● ●●
●
●
●
●
●
●
● ●●●
●●
●
●
●
●
●
●
●
●
●
● ●●●
● ●●
●
●● ●●●
●
●
●
●●
●
●
●●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●●
●
●●
●●
●
●
●
●
●●●
●
●●
●
●
●
●●●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●●
●
●
● ●●●
●
●
●
●
● ●
● ●
● ●●
●
● ● ●
●
● ●
●
●●●●●●
●
● ●
●●
●●●
● ●
●●
●
● ●
●
●
●●
●
●
●
●●●
●
●
●
●
● ●
●
●
●
●
●●
●●
●●
●
●
●
●●
●
●●
●●
● ●
●
●●
●●
●
●
●
●●
●
● ●
●
● ●
●
●
●
● ●
●●
● ●
●
●
● ●
●
●●
●
●
●
●
●
●●
●●
●
●●
●
●
●
●●
●
● ●
●
● ●
●
● ●●
● ●●
●●
●
●
● ●
●
● ● ●
●
● ● ●
●●●
● ● ●
●
●
●●
●●●
●
●●
●
●
●
●●
●
●
●● ●
●
●
●
●
● ●
●
●
●
●●
●
●
●
●●
●
●●
● ●
●
●
●●
●
●
●●
●●
●
●●●
●●
●●●●
●●
●
●
●
●
●●
●
●● ●
●●●
●
●
●
● ●●
●
● ●
●●
● ●
●
●
●●
●
●
●
●
● ●
●
●●●
●●●●
●
●● ●● ●● ●
●●●●
●● ●● ● ●● ● ●
−2 ● ● ● ● ● ● ● ● ● ●● ●
● ●● ● ●● ●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●●●
●
●●
●
●
●
●
●
●
●
●●●● ●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●●
● ●●● ●●
●●
●
●
●
●●
●
● ●
●
● ●
●
●
●
●●
●
●
●●
●●
● ●
●
●●
●
●●
●
●
●●
●●
●
●
●
●
●●
●
●
●●
●
●●●
● ●
●
●
●●●
●●
●
●
●
●●
●
●●
●
●
●
●
●
●
● ●
●
●
●
●
●●
●
●●
●
●
●●
●
●
●
●
●●
●●●
●●●
●
●●
●
●
●
● ●
●
●
●
● ●
●
●●
● ●●
● ●●
●
●●●●
●
●●
●
● ● ● ● ●●
●
●
● ● ●
●
●●
●
●
●
●●
●●
●●
●●
●
●●
●
●●●
●
●
●●
●
●
●●
●
●
●
●
●
●●
● ●
●
● ●
●
●●
●
●●
●
●●
● ●
●
● ●
●
●●
●
●
●
●●●
●
●
●●
●
●●●
●
●●
●●
●
●●
●
●● ●
●
●
● ●
●
●
●
● ● ●
● ●
●●
●
●
●●●● ●
●
●
●
● ●
●
● ●
●
●●
●
●
●●
●●
● ●
●●
●●●
●●●
●
●●
●●
●
●●
●●
● ●
●●●●
● ●
●
●
● ●
●●
●●
●
●
●●●
●● ●●●● ●
●●
●
● ● ●●
● ●● ●● ●●●● ● ●●●
●
●
●
●
●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●●●
●
●
●
●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●
●
●
●●
●●
●
● ●
●
●●
●
●●
●
●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●●
●●
● ●
●
●
●
●●
●●
●●● ●●
●
● ●
● ●
● ●●
●●●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●●●
●
●●●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●●
●
●
●
●
●
●●● ●
●●
●
●
●
●
●
●
●
●
● ●
●
●
● ●●● ●●
● ●
● ●
●●
●
●●
●●●●
●
●●
●
●●
●●
●●●
●
●●
●●●
●●●●
● ●
●●●
●●
●
●
●●●●
●● ●
● ● ● ● ●
●
●
●
●
● ●
●●
●
●●●
●
●
●
●
● ●
●
●●
● ●
●
●
●●
●
●●●
●
●●
●
●●
●
● ●●●●●
●
●
●●●
●
●●
●
●
● ●● ●●●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●●
●
●
● ●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
●
● ●
●●
●●
●●
●
●
●
●
●
●
●
●
●
●●
●●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●
●
●
●
●
●
●
●●
●
●
●
●
●
●●●●●
●
●●●
●
●
●
●
●
●
●
●
● ●
●
●
●
● ●●
●
●
●
●●●
●
●
● ●
●
●
●
●● ●●
●
●
●● ●
●
● ●●
●
● ●
●
●
●●●
●
●●● ●
●
● ●
−4 ●
● ●●
●●
●
●
●
●
●●●
● ●
●
●●● ●
●
●● ● ●
●●●
● ●
●
●
● ●
●●
● ●● ●
●
●
●
●
●
● ●
●
●
●
●
● ●●
●
●
● ●
●
● ●
● ●
●
●
●
● ●
●●
●
●
●
● ●
●
●●●
● ●●
●
● ●
●
●●
●
● ●
●●
● ● ●
●
● ● ● ●
●●●● ●
●
● ●
●● ● ●
●
●
● ●
●
●
●●●
● ● ●
●
● ●
●
●
●●●●●
● ●
●
●●
● ●
●
● ●
●●● ●
● ●●● ●
●
●
● ●●
●
●●
●
●
●
●
●●
●
●●
●
● ● ●
● ● ●
●
●
●
●●●
●
●●●●●
●
●● ●
●
●
● ● ●
● ● ●
●
● ●●
●
●
●
●
●
●
●●●
●
●●
●
● ●
●
● ●●
●
●●●
●●●●
●
●
●
● ●
●
● ●
●
●●● ● ● 1:1 line
●●
● ●● ●●
●
●●
● ●● ● ●
●
●
● ● ●
●
● fitted line
−6 ●●
●
●●

−6 −4 −2 0 2 4

Observed [log(g/g)]

1.5
0.8
50

1.0
0.6
nRMSE (%)

40
bias
r2

0.4 0.5
30

0.2
20 0.0

0.0
10

Fig. 3. Results from the iterative validation for the models using observations from all four time periods and Sentinel-2 data (120 samples). On the top, a scatterplot of
predicted against observed values of log transformed soil moisture is presented (reddish and bluish areas denote more and less overlapped points, respectively) while on the
bottom the distribution of the accuracies from the 500 iterations are displayed.

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
8 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

A December 2015 March 2016 g/g

1
0 500 m

B % 100

80

60

40

20

0
Fig. 4. Soil moisture prediction maps for the beginning (December) and end (March) of the summer season with Sentinel-2 imageries. In A, the mean value of the 500
iterations is showed for each pixel, while in B the coefficient of variation (%) per pixel is presented. Black lined polygons show the wetlands listed in the Wetland National
Inventory of 2011 (MMA-Centro de Ecología Aplicada, 2011). Predictions were carried out only within areas which were classified as wetland areas during the BSVM
classification.

not detectable and summer where the contrast between wet- component to take into account when classifying wetlands at a
lands and other land-cover classes is comparably high, indepen- local scale. Here, both sensors achieved high accuracies, which
dent from the date. We examined the simultaneous application indicates that for mapping Andean wetlands, a spatial resolution
of two Sentinel-2 scenes and found no relevant improvements of 10 m as provided by Sentinel-2 is adequate. Furthermore, we
and hence followed the principle of parsimony and present the also tested the effects of the homogenization of the sensors (i.e.
results obtained using individual images. Comparing the both at 10 m spatial resolution and 6 similar spectral bands) in
obtained classification accuracies for the two tested months, the classification procedure to test for possible effects of both spec-
we observed generally better results for March than December. tral and spatial resolution on the model accuracies. We could
We explain this with a reduced influence of snow melting effects observe a slight reduction of the performance difference between
which might still have had a certain influence in the December WV-2 and S-2 for the December images in the homogenized
images. We hence recommend March as a suitable month for images while there were no changes observable for March. Surpris-
identifying high Andean wetlands. ingly, even with the SWIR-bands missing, S-2 still performed nota-
Concerning spatial resolution, Frohn et al. (2012) and McCarthy bly better than the WV-2 datasets, probably due to their narrower
et al. (2015) agreed that high spatial resolution imagery is a key band width (higher spectral detail).

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx 9

4.2. Prediction of soil moisture by optical sensors full range of SM values is sampled in a balanced way. Another
potential influential factor could be that the successful estimation
Soil moisture (SM) values of the field reference samples showed of SM values in Andean wetlands relates to the correlation
a high variability across small distances with hardly any clear gra- between SM and vegetation cover. In areas with high SM values,
dient. Furthermore, the samples were highly skewed (i.e. high this relationship may saturate or even be inverted as permanently
presence of small values and few large ones), probably due to the flooded areas might have a slightly reduced vegetation signal due
small sample size (Hills and Reynolds, 1969). This resulted in many to the increased influence of open surface water on the observed
samples within a low range of observed SM values with the poten- reflectance.
tial presence of outliers in water-saturated areas. In our case, the
elimination of outliers did not translate into any improvement of 4.3. Potential of the examined sensor systems for operational
the models, hence we kept all samples in the model. monitoring of Andean wetlands in the framework of the chilean
Soil moisture is also known to be influenced by seasonal effect. national wetland inventory
Hills and Reynolds (1969) pointed out four decades ago the impor-
tance of an adequate sample size and the selected sampling time Chile has an historical deficit of ecological information about its
point in the year. Extreme sampling periods (e.g. beginning and wetlands, with Andean high altitude wetlands being one of the
end of summer) will enhance the sample variability in environ- least studied wetland ecosystems (Cepeda-Pizzaro and Pola,
ments such as the Andean highlands (Western et al., 2003). 2013; Squeo et al., 2006a,b,c). Even though Chile joined the Ramsar
From all tested SM models, the model applying S-2 data with all convention already in 1981, the first general classification and
120 observations reached highest accuracies. Applying reference detection of national wetlands within the framework of the
data from all four field sampling dates (using temporal blocking National Wetland Inventory (MMA-Centro de Ecología Aplicada,
for the validation) instead of just the beginning and the end of 2011) was only conducted in 2011. This inventory was carried
the season improved significantly (a = 0.05) the accuracies. High using Landsat imagery, where water bodies, vegetated areas and
improvements were found for r2 (median value increase from vegetation with visible water were classified by using the normal-
0.51 to 0.58 in the iterative validation), while the difference in ized difference water index (NWI) and the normalized vegetation
terms of nRMSE was not pronounced. S-2 showed significantly bet- index (NDVI). Validation was carried out by means of two sources
ter performance than WV-2 for predicting SM (a = 0.05; Table 4). of information: (1) a water bodies layer elaborated by the military
We relate this to S-20 s two bands in the SWIR which were found geographical institute (Instituto Geográfico Militar (IGM)), with a
to have a high importance for estimating SM (see Fig. A2 for vari- scale of 1:50,000, and (2) with GIS information created by the Min-
able importance of final PLSR model). The importance of the SWIR istry of Environment in the northern area, by using Landsat ima-
region to estimate SM has been stated in many earlier studies (e.g. gery and field information. Hence, the current inventory
Chen et al., 2014; Harris et al., 2006; Kaleita et al., 2005; Wang methodology is not exhaustive and it does not use reliable valida-
et al., 2008). Surprisingly, even when using only bands in the VIS tion data, which may hamper a proper monitoring and conserva-
and NIR spectral regions, S-2 still obtained higher performances tion of these ecosystems.
than WV-2 (median r2’s of 0.44 and 0.46 and nRMSE of 23% and Here, we proposed the use of optical Sentinel-2 data as a reli-
21.24% for S-2 and WV-2, respectively). This is maybe due the able alternative for Andean high wetlands monitoring for several
higher spectral resolution of S-2 which features bands with a nar- reasons:
rower spectral range than WV-2, and hence might be able to depict
more detailed spectral information related to SM. 1. Sentinel-2 data is free of cost: Hence, the data can be used for
Interestingly, Table 4 shows that WV-2 based models improve long term monitoring in conjunction with few sampling efforts,
their performance when increasing the pixel size from 2 m to 10 which is an important issue for governmental management. The
m (i.e. r2 increases from 0.34 to 0.44 and nRMSE decreases from high temporal resolution of Sentinel-2 furthermore allows for
24.36% to 21.24%). This may be due to the fact that high resolution the implementation of an operational monitoring system. Ref-
imagery such as W-2 are more prone to suffer from spectral arte- erence data could be obtained with a sparse number of field
facts caused by shadowing, while lower resolution sensors such campaigns and additional high-resolution RGB images (e.g.
as S-2 smooth the spectral information and hence reduce such GoogleEarth) which are nowadays freely available.
artefacts. 2. Spatial resolution: Sentinel-2 has higher spatial resolution than
The bootstrap validation of the selected model showed approx- other freely available optical RS data with SWIR information,
imately normal distributions or r2, nRMSE and bias values (Fig. 3). such as Landsat and MODIS, and clearly outcompetes other sen-
The value ranges obtained from the iterative validation procedure sor systems used for SM estimations with regard to spatial
are rather large which indicates that the selection of training and detail (e.g. passive microwaves;
100–4000 m). For the Andean
validation sample has a certain influence on the obtained results. wetlands this is a key requirement due to their small size and
Most models were based on 3–4 PLS components which can be often patchy occurrences.
considered a reasonable number of variables given the 120 sam- 3. Spectral properties: Sentinel-2 was found to deliver sufficient
ples (assuming the validity of the 1:10 rule-of-thumb; e.g. Austin spectral details to adequately identify Andean wetlands from
and Steyerberg, 2015). all other land-cover classes in this study. The direct comparison
Fig. 3 shows a systematic tendency of the models to underesti- to the commercial WorldView-2 data underlined the very high
mate higher values of SM. This is a typical behavior in this kind of radiometric quality of the sensor system.
models, where optical sensors are not able to differentiate properly 4. Vertical acquisition: The vertical acquisition of optical data
gradients of high or saturated values of SM. Some earlier studies could be a major advantage in mountainous areas such as the
consequently masked these areas out for the analysis (e.g. Andes, were the oblique acquisition of e.g. radar imagery may
Sadeghi et al., 2017). Fig. 4 shows that the highest uncertainties be difficult due to the steep slopes and complicated terrain
in terms of a high coefficient of variation of SM predictions were which can cause shadowing effects.
observed in areas with high SM values and near to borders. This
is likely to be at least partly explainable with the highly skewed All these characteristics turn S-2 into a robust alternative for
reference data with a limited number of reference samples for high monitoring Andean wetlands. The presented classification
SM values. In future studies, it should hence be ensured that the approach may allow the establishment of new datasets to actualize

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
10 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

the National Wetland Inventory, while the monitoring of soil mois- ping the occurrence and extent of Andean wetlands and
ture gradients may furthermore support the early detection of moderate performance for estimating soil moisture content within
impacts caused by anthropogenic impacts or global warming. the wetlands. According to the aims of the study we can conclude:

1. The one-class classifier Bias Support Vector Machines was

5. Conclusions found to deliver accurate results for the delineation of Andean
wetlands when using single-date imagery and minimal field
Optical remote sensing was found to be a reliable source of data. Both WorldView-2 and Sentinel-2 imagery deliver similar
information to map Andean wetlands in complicated terrain. Mod- results demonstrating that the separation of the wetlands from
els based on Sentinel-2 data showed high performances for map- the landscape background is possible without information from

A December 2015 March 2016

0.4

0.4
0.3

0.3
0.2

0.2
Density

Density
0.1

0.1
0.0

0.0

−4 −3 −2 −1 0 1 2 −4 −3 −2 −1 0 1 2

predictive value predictive value

B
0.4

0.4
0.3

0.3
0.2

0.2
Density

Density
0.1

0.1
0.0

0.0

−4 −3 −2 −1 0 1 2 −2 −1 0 1

predictive value predictive value

Fig. A1. Analytical graph of the ‘oneClass’ R package to visually validate the selected threshold (0.5) of the BSVM classification. The gray, dark blue and light blue boxplots
summarize the predictions for the background, the training and the validation data, respectively. Panels marked with A show predictions of WorldView-2 based models while
panels marked with B show predictions for Sentinel-2 based models.

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx 11

the short-wave infrared. This is important when monitoring at References

fine spatial grain is needed. In this regard, the use of new gen-
erations of very-high resolution imagery, such as WorldView-2 Adam, E., Mutanga, O., Rugege, D., 2010. Multispectral and hyperspectral remote
sensing for identification and mapping of wetland vegetation: A review. Wetl.
and 3, Pleiades or the new Venls, is especially interesting. Ecol. Manag. 18, 281–296. https://doi.org/10.1007/s11273-009-9169-z.
2. To predict soil moisture from optical data, SWIR information is Ahumada, M., Aguirre, F., Contreras, M., Figueroa, A., 2011. Guía para la
fundamental, especially in areas with similar plant cover but Conservación y Seguimiento Ambiental de Humedales Andinos. Minist. del
Medio Ambient, Santiago, Chile.
differing soil moisture values. In our study, Sentinel-2 allowed Ahumada, M., Faúndez, L., 2009. Guía descriptiva de los sistemas vegetacionales
the monitoring of Andean wetlands with sufficient accuracy. azonales hídricos terrestres de la ecorregión altiplánica (SVAHT). Minist. Agric.
It could hence be used for monitoring and management pur- Chile, Serv. Agrícola y Ganad. Santiago, Chile.
Alexakis, D.D., Mexis, F.K., Vozinaki, A.k., Daliakopoulos, I.N., Tsanis, I.K., 2017. Soil
poses, particularly in environments where the size of some of
moisture content estimation based on sentinel-1 and auxiliary earth
the wetlands are too small for the use of traditional passive observation products. a hydrological approach. Sensors 17, 1966–1987.
microwaves-thermal and active radar approaches. However, a https://doi.org/10.3390/s17061455.
basic requirement of such a monitoring system would be a lim- Austin, P.C., Steyerberg, W.E., 2015. The number of subjects per variable required in
linear regression analyses. J. Clin. Epidemoil. 68, 627–636.
ited number of soil moisture measurements from repeated Baker, C., Lawrence, R., Montagne, C., Patten, D., 2006. Mapping wetlands and
field-surveys. riparian areas using Landsat ETM+ Imagery and decision-tree-based models.
Soc. Wetl. Sci. Wetl. 26, 465–474.
Betbeder, J., Rapinel, J., Corgne, S., Pottier, E., Hubert-Moy, L., 2015. TerraSAR-X dual-
The presented approach and results may be of high importance pol time-series for mapping of wetland vegetation. ISPRS J. Photogramm.
for the monitoring and management of Andean wetlands in Chile Remote Sens. 107, 90–98.
in a local and regional scale, and particularly in areas where the Cepeda-Pizzaro, J., Pola, M.L., 2013. Relaciones de abundancia de órdenes de
hexápodos terrestres en vegas altoandinas del desierto-transicional de Chile.
anthropogenic pressures over water resources is elevated due to Idesia 31, 31–39. https://doi.org/10.4067/S0718-34292013000200005.
the mining industry. Chavez, P.A.T.S., 1988. An Improved dark-object subtraction technique for
atmospheric scattering correction of multispectral data. Remote Sens.
Environ. 24, 459–479.
Chen, C.F., Valdez, M.C., Chang, N. Bin, Chang, L.Y., Yuan, P.Y., 2014. Monitoring
spatiotemporal surface soil moisture variations during dry seasons in central
Acknowledgements america with multisensor cascade data fusion. IEEE J. Sel. Top. Appl. Earth Obs.
Remote Sens. 7, 4340–4355. https://doi.org/10.1109/JSTARS.2014.2347313.
This investigation was funded by the project ‘Seguimiento de Contreras, R., 2007. Uso de vegas y bofedales de la zona cordillerana y
precordillerana de la región de Atacama. Universidad de Chile.
Vegas Altoandinas con Imágenes Satelitales de Alta Resolución
Davranche, A., Lefebvre, G., Poulin, B., 2010. Wetland monitoring using classification
Espacial” financed by ALTO MAIPO SpA’ (AM-CO292) and the Fac- trees and SPOT-5 seasonal time series. Remote Sens. Environ. 114, 552–562.
ulty of Forestry and Nature Conservation of the University of Chile. https://doi.org/10.1016/j.rse.2009.10.009.
Davranche, A., Poulin, B., Lefebvre, G., 2013. Mapping flooding regimes in Camargue
We further thank Dr. Eduardo Martínez, Silvicuture Departament,
wetlands using seasonal multispectral data. Remote Sens. Environ. 138, 165–
U. de Chile, for providing and pre-processing the field data, Luis 171. https://doi.org/10.1016/j.rse.2013.07.015.
Enrique Olivera-Guerra for his advices on SAR technologies and Feilhauer, H., Asner, G.P., Martin, R.E., Schmidtlein, S., 2010. Brightness-normalized
Florian Hartig for his advice on data transformation and validation Partial Least Squares Regression for hyperspectral data. J. Quant. Spectrosc.
Radiat. Transf. 111, 1947–1957. https://doi.org/10.1016/j.jqsrt.2010.03.007.
procedures. Frohn, R., D’Amico, E., Lane, C., Autrey, B., Rhodus, J., Liu, H., 2012. Multi-temporal
sub-pixel Landsat ETM+ classification of isolated wetlands in Cuyahoga County,
Ohio, USA. Wetlands 32, 289–299.
Gao, Z., Xub, M., Wang, J., Yang, H., Huang, W., Feng, H., 2013. A method of
estimating soil moisture based on the linear decomposition of mixture pixels.
Appendix A. Classification model selection Math. Comput. Model. 58, 606–613.
Gao, Q., Zribi, M., Escorihuela, M.J., Baghdadi, N., 2017. Synergetic use of sentinel-1
See Fig. A1. and sentinel-2 data for soil moisture mapping at 100 m resolution. Sensors 17,
1455–1471. https://doi.org/10.3390/s17091966.
García, E., Otto, M., 2015. Caracterización ecohidrológica de humedales alto andinos
usando imágenes de satélite multitemporales en la cabecera de cuenca del Río
Santa, Ancash, Perú. Ecol. Apl. 14, 115–125.
Appendix B. Best regression model information GORE-RMS, 2013. Estrategia Regional para la Conservación de la Biodiversidad en la
Región Metropolitana de Santiago 2015-2025.
Harris, A., Bryant, R.G., Baird, A.J., 2006. Mapping the effects of water stress on
See Fig. A2. Sphagnum: Preliminary observations using airborne remote sensing. Remote
Sens. Environ. 100, 363–378. https://doi.org/10.1016/j.rse.2005.10.024.
Hills, R.C., Reynolds, S.G., 1969. Illustrations of soil moisture variability in selected
● ●
areas and plots of different sizes. J. Hydrol. 8, 27–49.
Kaleita, a.L., Tian, L.F., Hirschi, M.C., 2005. Relationship between soil moisture
1.0

content and soil surface reflectance. Trans. Asae 48, 1979–1986. https://doi.org/
10.13031/2013.19990.
Weighted coefficients

●
0.5

Kohavi, R., 1995. A study of cross-validation and bootstrap for accuracy estimation
and model selection. Proc. Int. Joint Conf. Artificial Intelligence, 1137–1143.
Kerr, Y.H., Waldteufel, P., Richaume, P., 2012. The SMOS soil moisture retrieval
−2.0 −1.5 −1.0 −0.5 0.0

algorithm. IEEE Trans. Geosci. Remote Sens. 50, 1384–1403.

Lee, T., Yeh, H., 2009. Applying remote sensing techniques to monitor shifting
wetland vegetation: A case study of Danshui River estuary mangrove
communities, Taiwan. Ecol. Eng. 35, 487–496.
●
Li, W., Guo, Q., 2014. A new accuracy assessment method for one-class remote
sensing classification. IEEE Trans. Geosci. Remote Sens. 52, 4621–4632. https://
●
doi.org/10.1109/TGRS.2013.2283082.
Li, W., Guo, Q., Elkan, C., 2011. A positive and unlabeled learning algorithm for one-
● ● Significant band (p<0.01) class classification of remote-sensing data. IEEE Trans. Geosci. Remote Sens. 49,
717–725.
B1 B2 B3 B4 B5 B6 B7 B8 B9 B10 Liu, B., Dai, Y., Li, X., Lee, W.L., Yu., P.S., 2003. Building text classifiers using positive
and unlabeled examples. In: Third IEEE International Conference on Data
Sentinel−2 bands Mining, Melbourne, Nov 19–22, pp. 179–186.
Luckman, A.J., 1998. The effect of topography on mechanisms of radar backscatter
Fig. A2. Variable importance of the best model using partial least square regression from coniferous forest and upland pasture. IEEE T. Geosci. Remote Sens. 36,
(PLSR) and Sentinel-2 data. Only statistically significant bands are displayed. 1830–1834. https://doi.org/10.1109/36.718651.

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001
12 R.A. Araya-López et al. / ISPRS Journal of Photogrammetry and Remote Sensing xxx (2018) xxx–xxx

Lopatin, J., Fassnacht, F.E., Kattenborn, T., Schmidtlein, S., 2017. Mapping plant Sadzawka, A., Carrasco, M.A., Grez, R., Mora, M.L., Flores, P., Neaman, A., 2006. M{é}
species in mixed grassland communities using close range imaging todos de an{á}lisis de suelos recomendados para los suelos de Chile. Instituto de
spectroscopy. Remote Sens. Environ. 201, 12–23. https://doi.org/10.1016/j. Investigaciones Agropecuarias, Santiago, Chile.
rse.2017.08.031. Schmidtlein, S., Feilhauer, H., Bruelheide, H., 2012. Mapping plant strategy types
Lopatin, J., Dolos, K., Hernández, H.J., Galleguillos, M., Fassnacht, F.E., 2016. using remote sensing. J. Veg. Sci. 23, 395–405. https://doi.org/10.1111/j.1654-
Comparing generalized linear models and random forest to model vascular 1103.2011.01370.x.
plant species richness using LiDAR data in a natural forest in central Chile. Skowronek, S., Asner, G.P., Feilhauer, H., 2017. Performance of one-class classifiers
Remote Sens. Environ. 173, 200–210. https://doi.org/10.1016/j.rse.2015.11.029. for invasive species mapping using airborne imaging spectroscopy. Ecol. Indic.
MacAlister, C., Mahaxay, M., 2009. Mapping wetlands in the Lower Mekong Basin 37, 66–76.
for wetland resource and conservation management using Landsat ETM images Song, B., Li, P., Li, J., Plaza, A., 2016. One-class classification of remote sensing images
and field survey data. J. Environ. Manage. 90, 2130–2137. https://doi.org/ using kernel sparse representation. IEEE J. Sel. Top. Appl. Earth Obs. Remote
10.1016/j.jenvman.2007.06.031. Sens. 9, 1613–1623. https://doi.org/10.1109/JSTARS.2015.2508285.
Mack, B., 2017. One-class classification in R with the oneClass package. https:// Squeo, F., Cepeda, J., Olivares, N.C., Arroyo, M.T.K., 2006a. Interacciones ecológicas
github.com/benmack/oneClass/blob/master/notebooks/oneClassIntro.ipynb en la alta montaña Del Valle Del Elqui. In: GEOECOLOGÍA de Los ANDES
(Last access: 13. Jan. 2018). Desérticos. La Alta Montaña Del Valle Del Elqui. Ediciones Universidad de La
Mack, B., Roscher, R., Stenzel, S., Feilhauer, H., Schmidtlein, S., Waske, B., 2016. Serena, La Serena, pp. 69–103.
Mapping raised bogs with an iterative one-class classification approach. ISPRS J. Squeo, F., Ibacache, E., Warner, B., Espinoza, D., 2006b. Productividad y Diversidad
Photogramm. Remote Sens. 120, 53–64. https://doi.org/10.1016/j. Florística De La Vega Tambo. In: GEOECOLOGÍA de Los ANDES Desérticos. La
isprsjprs.2016.07.008. Alta Montaña Del Valle Del Elqui. Ediciones Universidad La Serena, La Serena,
Mack, B., Roscher, R., Waske, B., 2014. Can i trust my one-class classification? pp. 325–351.
Remote Sens. 6, 8779–8802. https://doi.org/10.3390/rs6098779. Squeo, F., Warner, B., Aravena, R., Espinoza, D., 2006c. Bodedales: high altitude
Mahdianpari, M., Salehi, B., Mohammadimanesh, F., Motagh, M., 2017. Random peatlands of the central Andes. Rev. Chil. Hist. Nat. 79, 245–255. https://doi.org/
forest wetland classification using ALOS-2 L-band, RADARSAT-2 C-band, and 10.4067/S0716-078X2006000200010.
TerraSAR-X imagery. ISPRS J. Photogramm. Remote Sens. 130, 13–31. Stenzel, S., Fassnacht, F.E., Mack, B., Schmidtlein, S., 2017. Identification of high
Martens, H., Martens, M., 2000. Modified Jack-knife estimation of parameter nature value grassland with remote sensing and minimal field data. Ecol. Indic.
uncertainty in bilinear modelling by partial least squares regression (PLSR). 74, 28–38. https://doi.org/10.1016/j.ecolind.2016.11.005.
Food Qual. Prefer. 11, 5–16. https://doi.org/10.1016/S0950-3293(99)00039-7. Stenzel, S., Feilhauer, H., Mack, B., Metz, A., Schmidtlein, S., 2014. Remote sensing of
McCarthy, M.J., Merton, E.J., Muller-Karger, F.E., 2015. Improved coastal wetland scattered natura 2000 habitats using a one-class classifier. Int. J. Appl. Earth
mapping using very-high 2-meter spatial resolution imagery. Int. J. Appl. Earth Obs. Geoinf. 33, 211–217. https://doi.org/10.1016/j.jag.2014.05.012.
Obs. Geoinf. 40, 11–18. https://doi.org/10.1016/j.jag.2015.03.011. Tiner, R., 2015. Classification of wetland types for mapping and large-scale
Merlin, O., Rudiger, C., Bitar, A. Al, Richaume, P., Walker, J.P., Kerr, Y.H., 2012b. inventories. In: Remote Sensing of Wetland. Applications and Advances.
Disaggregation of SMOS soil moisture in southeastern Australia. IEEE Trans. Taylor & Francis Group, Boca Raton, Florida, pp. 19–41.
Geosci. Remote Sens. 50, 1556–1571. Todd Updike, Comp, C., 2010. Radiometric Use of WorldView-2 Imagery Technical
MMA-Centro de Ecología Aplicada, 2011. Diseño del inventario nacional de Note 1 WorldView-2 Instrument Description.
humedales y el seguimiento ambiental. Townsend, P.A., Walsh, S.J., Ecology, S.P., Dec, N., Townsend, A., Stephen, J., J.S., 2001.
Molero, B., Merlin, O., Malbéteau, Y., Bitar, A. Al, Cabot, F., Stefan, V., Kerr, Y., Bacon, Remote sensing of forested wetlands : application of multitemporal and
S., Cosh, M.H., Bindlish, R., Jackson, T.J., 2016. SMOS disaggregated soil multispectral satellite imagery to determine plant community composition
moisture product at 1 km resolution: Processor overview and first validation and structure in southeastern stable URL : http://www.jstor.org/stable/
results. Remote Sens. Environ. 180, 361–376. https://doi.org/10.1016/j. 20051167 and sensing of forested wetland. Plant Ecol. 157, 129–149. https://
rse.2016.02.045. doi.org/10.1023/A:1013999513172.
Muñoz-Schick, M., Núñez, H., Yáñez, J., 1997. Libro rojo de los sitios prioritarios para Wang, J., Ling, Z., Wang, Y., Zeng, H., 2016. Improving spatial representation of soil
la conservación de la biodiversidad en Chile. Ambient. Desarro. 13, 90–99. moisture by integration of microwave observations and the temperature–
Ozesmi, S.L., Bauer, M.E., 2002. Satellite remote sensing of wetlands. Wetl. Ecol. vegetation–drought index derived from MODIS products. ISPRS J. Photogramm.
Manage. 10, 381–402. https://doi.org/10.1023/A:1020908432489. Remote Sens. 113, 144–154. https://doi.org/10.1016/j.isprsjprs.2016.01.009.
Paloscia, S., Pettinato, S., Santi, E., Notarnicola, C., Pasolli, L., Reppucci, A., 2013. Soil Wang, B., Chen, Y., Lü, C., 2015. Evaluating flood inundation impact on wetland
moisture mapping using Sentinel-1 images: Algorithm and preliminary vegetation FPAR of the Macquarie marshes, Australia. Environ. Earth Sci. 74,
validation. Remote Sens. Env. 134, 134–148. 4989–5000.
Pasolli, L., Notarnicola, C., Bruzzone, L., Bertoldi, G., Della Chiesa, S., Niedrist, G., Wang, L., Qu, J.J., Hao, X., Zhu, Q., 2008. Sensitivity studies of the moisture effects on
Tappeiner, U., Zebisch, M., 2012. Polarimetric RADARSAT-2 imagery for soil MODIS SWIR reflectance and vegetation water indices. Int. J. Remote Sens. 29,
moisture retrieval in alpine areas. Can. J. Remote Sens. 37, 535–547. https://doi. 7065–7075. https://doi.org/10.1080/01431160802226034.
org/10.5589/m11-065. Western, A., Zhou, S., Grayson, B.R., McMahon, A.T., Blöschl, G., Wilson, J.D., 2003.
Reschke, J., Hüttich, C., 2014. Continuous field mapping of Mediterranean wetlands Spatial correlation of soil moisture in small catchments and its relationship to
using sub-pixel spectral signatures and multi-temporal Landsat data. Int. J. dominant spatial hydrological processes. J. Hydrol. 286, 113–134.
Appl. Earth Obs. Geoinf. 28, 220–229. https://doi.org/10.1016/ Whiteside, T.G., Bartolo, R.E., 2015. Mapping aquatic vegetation in a tropical
j.jag.2013.12.014. wetland using high spatial resolution multispectral satellite imagery. Remote
Roberts, D.R., Bahn, V., Ciuti, S., Boyce, M.S., Elith, J., Guillera-Arroita, G., Hauenstein, Sens. 7, 11664–11694.
S., Lahoz-Monfort, J.J., Schröder, B., Thuiller, W., Warton, D.I., Wintle, B.A., Wold, S., Sjöström, M., Eriksson, L., 2001. PLS-regression: a basic tool of
Hartig, F., Dormann, C.F., 2017. Cross-validation strategies for data with chemometrics. Chemom. Intell. Lab. Syst. 58, 109–130. https://doi.org/
temporal, spatial, hierarchical, or phylogenetic structure. Ecography 1–17. 10.1016/S0169-7439(01)00155-1.
https://doi.org/10.1111/ecog.02881. Zribi, M., Baghdadi, N., Holah, N., Fafin, O., 2005. New methodology for soil surface
Rundquist, D.C., Narumalani, S., Narayanan, R.M., 2001. A review of wetlands moisture estimation and its application to ENVISAT-ASAR multi-incidence data
remote sensing and defining new considerations. Remote Sens. Rev. 20, 207– inversion. Remote Sens. Environ. 96, 485–496. https://doi.org/10.1016/j.
226. https://doi.org/10.1080/02757250109532435. rse.2005.04.005.
Sadeghi, M., Babaeian, E., Tuller, M., Jones, S.B., 2017. The optical trapezoid model: A Zribi, M., Decharme, B., 2008. A method for soil moisture estimation in Western
novel approach to remote sensing of soil moisture applied to Sentinel-2 and Africa based on the ERS scatterometer. IEEE Trans. Geosci. Remote Sens. 46,
Landsat-8 observations. Remote Sens. Environ. 198, 52–68. 438–448.

Please cite this article in press as: Araya-López, R.A., et al. Monitoring Andean high altitude wetlands in central Chile with seasonal optical data: A com-
parison between Worldview-2 and Sentinel-2 imagery. ISPRS J. Photogram. Remote Sensing (2018), https://doi.org/10.1016/j.isprsjprs.2018.04.001