Inland Waters

ISSN: 2044-2041 (Print) 2044-205X (Online) Journal homepage: http://www.tandfonline.com/loi/tinw20

Using dynamic factor analysis to show how

sampling resolution and data gaps affect the
recognition of patterns in limnological time series

Rosana Aguilera, David M. Livingstone, Rafael Marcé, Eleanor Jennings,

Jaume Piera & Rita Adrian

To cite this article: Rosana Aguilera, David M. Livingstone, Rafael Marcé, Eleanor Jennings,
Jaume Piera & Rita Adrian (2016) Using dynamic factor analysis to show how sampling resolution
and data gaps affect the recognition of patterns in limnological time series, Inland Waters, 6:3,
284-294, DOI: 10.1080/IW-6.3.948

To link to this article: https://doi.org/10.1080/IW-6.3.948

Published online: 02 Jan 2018.

Submit your article to this journal

View related articles

View Crossmark data

Full Terms & Conditions of access and use can be found at

http://www.tandfonline.com/action/journalInformation?journalCode=tinw20
 284

Article

Using dynamic factor analysis to show how sampling resolution

and data gaps affect the recognition of patterns in limnological
time series
Rosana Aguilera,1*† David M. Livingstone,2 Rafael Marcé,1 Eleanor Jennings,3 Jaume Piera,4 and Rita Adrian5,6
1
Catalan Institute for Water Research (ICRA), Girona, Spain
2
Eawag, Swiss Federal Institute of Aquatic Science and Technology, Dübendorf, Switzerland
3
Department of Applied Sciences, Dundalk Institute of Technology, Dundalk, Ireland
4
Physical and Technological Oceanography Department, Institute of Marine Sciences (ICM-CSIC), Barcelona, Spain
5
Leibniz-Institute of Freshwater Ecology and Inland Fisheries, Berlin, Germany
6
Free University Berlin, Department of Biology, Chemistry and Pharmacy, Berlin, Germany
†
Current address: Marine Science Institute, University of California Santa Barbara, Santa Barbara, CA, USA
* Corresponding author: raguilera@icra.cat

Received 11 November 2015; accepted 23 February 2016; published 17 May 2016

Abstract

Currently, many water quality variables can be automatically monitored at subdaily frequencies, but most environmen-
tal datasets still rely heavily on comparatively low-frequency (e.g., monthly) sampling campaigns. Taking advantage of
the long-term ecological data available from polymictic Lake Müggelsee (Germany) at weekly temporal scales, we
tested whether dynamic factor analysis (DFA), a dimension-reduction technique especially designed for time series
with data gaps, was able to reproduce the same underlying patterns if the resolution of the observations was artificially
reduced from weekly to once every 2 or 4 weeks (2-weekly or 4-weekly, respectively) and whether the same results
were obtained for different variables (water temperature, water chemistry, and plankton). As expected, the reduction in
data resolution and the addition of gaps increased the amount of variance in our case studies; however, water
temperature patterns, which generally vary only slowly if considered at broader temporal scales, were represented well
in the 4-weekly data model. By contrast, total algal biomass and nutrient dynamics, which usually fluctuate more
rapidly, were not. We further assessed the representativeness of spot samples taken by selection at 4-weekly and
seasonal intervals compared to samples obtained by averaging the weekly data over these 2 intervals. In this case, the
resulting observation variance was more pronounced in DFA models of the samples taken by selection than of those
obtained by averaging. The introduction of additional artificial gaps in one univariate and one multivariate model
decreased the amount of information that could be explained by the pattern(s) extracted by DFA. Overall, we found that
a too low sampling frequency could be a limiting factor when studying interactions and responses to changing
conditions in aquatic systems, particularly when biological variables are involved.

Key words: dynamic factor analysis, Müggelsee, sampling resolution, time series

Introduction frequency, defined as half the sampling frequency (Zayed

1993). In other words, the sampling frequency should be at
The Nyquist-Shannon sampling theorem implies that a least twice the highest frequency contained in the signal.
continuous signal can be properly sampled only if it An inadequate sampling rate results in the phenomenon
contains no power at frequencies above the Nyquist known as aliasing, in which power at frequencies higher

DOI: 10.5268/IW-6.3.948 Inland Waters (2016) 6, pp.284-294

© International Society of Limnology 2016
Effects of sampling resolution and data gaps of limnological data on pattern recognition
than the Nyquist frequency is interpreted as appearing at
frequencies lower than the Nyquist frequency, which
results not only in a loss of information about the system
but also distorts the information obtained. In limnology,
these signals might represent processes and interactions
between variables within a given ecosystem or region.
Processes in lakes are complex, involving overlapping
temporal and trophic-level scales. Thus, when measuring
ecological variables in lakes, the sampling frequency and
the lengths of the time series obtained are critical for
capturing and understanding patterns of temporal evolution
in these variables. These patterns include (but are not
limited to) long-term trends and seasonal variability, which
can usually be identified using lower frequency data (e.g.,
monthly) as well as high-frequency fluctuations, which
typically require finer data resolution.
In the past few decades, studies relying mainly on
weekly to once every 4 week (4-weekly) spot samples
(Halliday et al. 2012) have revealed the complex nature of
ecosystems response to environmental change and the
time scales involved (review in Adrian et al. 2012). Such
studies have encouraged the measurement of long-term,
high-frequency datasets (Parr et al. 2003). The recent
development and deployment of high-frequency sensor
monitoring equipment and the increasing duration of ob-
servational time series available from lakes have in turn
allowed more detailed study of lake responses and
processes (Jennings et al. 2012, Solomon et al. 2013, Rose
et al. 2014). These technological developments also
guarantee that the sampling frequency will normally
exceed the Nyquist frequency, avoiding any numerical
artifacts associated with aliasing.
Despite such advances, most limnological datasets are
still prone to uneven sampling intervals and missing ob-
servations. These sampling campaigns are, in most cases,
subject to logistical and financial limitations and often do
not take into account the ideal frequency requirements
necessary to obtain adequate sample representativeness of
the system being studied. These commonly encountered
problems can represent a limitation in the study of the
underlying patterns in ecosystem responses to climatic and
anthropogenic forcing.
Irregularities in sampling frequency and the presence
of data gaps are often the first challenges encountered
when selecting an appropriate method for analyzing time
series. Consequently, the choice of analytical methodology
becomes crucial to taking advantage of all available
information. Of the various time-series analysis techniques
that can be used to study responses and patterns, few are
able to cope with these challenges. For instance, Schoell-
hamer (2001) applied a modified singular-spectrum
analysis algorithm to obtain spectral estimates from
records with a large fraction of missing data. The weighted
DOI: 10.5268/IW-6.3.948
285
wavelet Z-transform (WWZ) method can handle unevenly
sampled data, but large gaps in the data would complicate
the identification of temporal patterns (Foster 1996).
Another example is Seasonal Kendall trend analysis,
which can cope with missing values (Hirsch et al. 1982).
Nevertheless, commonly used methods such as spectral
analysis, wavelet analysis, and Box-Jenkins modeling
usually require time series to be stationary and evenly
sampled (Zuur et al. 2003a). In addition, these methods are
not particularly suitable for unraveling patterns and inter-
actions between variables over time.
Dynamic factor analysis (DFA) is a dimension-reduc-
tion method designed for time series (Zuur et al. 2003a)
used mainly to estimate underlying common patterns in a
set of time series. In this method, observations are related
to the hidden patterns or states by the state-space model.
This framework provides a means for modeling nonlinear
effects and structural change in time series and is able to
cope with missing observations (Harvey 1989). DFA has
been used widely in econometric and psychological studies
and has recently been applied to fisheries datasets (Zuur et
al. 2003b, Zuur and Pierce 2004, Vilizzi 2012), groundwa-
ter-quality data (Muñoz-Carpena et al. 2005, Kuo et al.
2013), and river water-quality data (Aguilera et al. 2015).
To address the issue of the temporal scale of sampling
and gaps in limnological time series, we took advantage
of an ongoing long-term research program at Müggelsee,
a shallow, polymictic lake in Berlin, Germany, in which
data on physicochemical variables and plankton are
collected weekly. We chose a shallow, polymictic lake
because this type of lake is known to be more strongly
influenced by short-term ambient weather conditions than
are dimictic lakes (Gerten and Adrian 2000). The ultimate
goal of this study was to assess the potential usefulness of
DFA in analyzing patterns in limnological time series
with differing temporal sampling resolutions and data
gaps. To achieve this, we specified 4 objectives. (1) We
tested whether DFA was able to reproduce the same
underlying patterns in the data when the sampling
resolution, in univariate analyses, was reduced from
weekly to 4-weekly. (2) We compared and contrasted
analyses of physicochemical and biological variables
because these 2 types of variables may differ with regard
to their frequency spectra. (3) We aimed to determine
whether the complex dynamics of a set of time series
with well-known interactions could be described using
DFA and differing sampling frequencies. (4) Finally, we
assessed the ability of DFA to cope with missing observa-
tions in 2 ways: (a) by introducing artificial gaps into the
time series and investigating the effect of this on the
patterns extracted; and (b) by resampling the available
weekly data to obtain lower-frequency (i.e., seasonal and
4-weekly) datasets.
Inland Waters (2016) 6, pp.284-294
286
Table 1. Time period and number of observations for the Müggelsee variables included in the study. (Note that for pH the number of observa-
tions was substantially lower than for the other variables).
Number of observations
Variable Weekly 2-weekly
Water temperature 1748 874
Dissolved oxygen 1722 861
pH 1027 514
Ammonium 1721 861
Total algal biomass 1721 861
Copepod species case study 1614 807
Methods
Study site and time-series data
Müggelsee is a shallow, polymictic lake in Berlin
(Germany) with a surface area of 7.3 km2, a mean depth of
4.9 m, and a water retention time of ~6–8 weeks. Data
collection in the lake has been undertaken weekly (from
spring to autumn) or once every 2 weeks (2-weekly; during
winter) since 1979 (Köhler et al. 2005). A comprehensive
description of the characteristics of Müggelsee and its sur-
roundings can be found in Driescher et al. (1993). Physico-
chemical and biological variables from the Müggelsee
dataset were selected to carry out DFA. Variables included
water temperature, dissolved oxygen (DO), and pH (all
measured at 1 m depth), nutrient concentrations
(ammonium), and the abundances of phytoplankton (both
taken from the upper 4 m of the water column) and
zooplankton (taken from the entire water column). A
detailed description of sampling and sample processing is
given in Driescher et al. (1993) and in Gerten and Adrian
(2000).
The time spans covered by the time series ranged from
~19 yr (pH) to 33 yr (water temperature; Table 1). The finest
resolution was weekly. The reduction in resolution from
weekly to 2-weekly and 4-weekly was accomplished by
simple decimation (i.e., by selecting every second or every
fourth value from the original weekly dataset).
Dynamic factor analysis (DFA)
Dynamic factor analysis is based on structural time-series
models viewed from within the state-space model framework
(Harvey 1989). In DFA, all common patterns in a set of time
series are estimated simultaneously using the maximum
likelihood method. These patterns are modeled as random
walks and thus are allowed to be stochastic, meaning that
their shape can change over time and is not necessarily
restricted to a straight-line trend or a cosine-function cyclic or
seasonal component (Zuur et al. 2007). The state-space
© International Society of Limnology 2016
Aguilera et al.
Date
4-weekly Start End
437 24 Jul 1978 16 Jan 2012
431 22 Jan 1979 16 Jan 2012
257 18 May 1992 16 Jan 2012
431 01 Jan 1979 19 Dec 2011
431 01 Jan 1979 19 Dec 2011
404 21 Jan 1980 20 Dec 2010
formulation, and thus DFA, also allows a univariate time
series to be decomposed into a pattern and an error term.
DFA was performed using the Multivariate Autoregres-
sive State-Space (MARSS) R-package 3.4 (Holmes et al.
2013), run in R 2.15.1 (R Core Team 2012). Each time series
was standardized by subtracting its mean and dividing by its
standard deviation to facilitate the interpretation and
comparison of resulting patterns and factor loadings. The
basis of the DFA formulation in MARSS includes a process
model (x) and an observation model (y):
xt = xt-1 + wt wt ~ MVN(0,Q)
yt = Zxt + vt vt ~ MVN(0,R). (1)
The observations (yt) are modeled as a linear combination of
hidden patterns (xt) and factor loadings (Z; Holmes et al.
2012). The terms wt and vt describe the errors associated with
the hidden patterns and the observations, respectively,
normally distributed with expectation 0 and with covariance
matrices Q and R. The Q matrix is set to an identity matrix,
and the R matrix is usually taken to be diagonal. The diagonal
values in the R matrix contain the variances of the time series
involved, and the off-diagonal values contain the covariances
between the time series. The MARSS algorithm uses the
Kalman filter and smoother to compute the best estimate for
the hidden patterns, given a set of assumed values for the
unknown parameters Z, Q, and R. The output of the Kalman
filter is then used to compute the likelihood of the data given
the better estimates for the set of parameters. MARSS
estimates the maximum likelihood using the expectation–
maximization (EM) method (Holmes 2013). The number of
iterations necessary is determined by the log-likelihood of
reaching convergence (slope tolerance for log-log test = 0.1;
absolute tolerance test = 0.001).
A considerable advantage of the state-space approach is
the ease of addressing missing observations (Durbin and
Koopman 2012). Holmes (2013) performed a series of
modifications in the EM update equations for missing
values in MARSS models, based on the work of Shumway
and Stoffer (1982, 2010) and Zuur et al. (2003a), and
presented a thorough explanation of the methodology. The
DOI: 10.5268/IW-6.3.948
Effects of sampling resolution and data gaps of limnological data on pattern recognition
main disadvantage of DFA is its intensive computations,
translating into long computational times ranging from
hours to days and weeks, depending on the complexity of
the model. In our study, the High-Performance Computing
Cluster Undarius allowed the simultaneous computation of
multiple independent analyses. Nevertheless, the
computation time for the multivariate analyses (based on
the copepod species and cryptophytes case study described
below) ranged from 1 to 4 weeks.
Effect of sampling resolution
Univariate time-series analysis
Univariate time-series analyses were performed so that
a pattern plus an error term would be extracted from
each variable along a trophic hierarchy (water
temperature, nutrients, DO, pH, phytoplankton,
zooplankton) and at different data sampling resolutions
(weekly, 2-weekly, and 4-weekly). Thus, not only the
effect of differences in the temporal resolution of the
time-series data was assessed, but also the behavior of
different system levels across the trophic hierarchy,
which is affected by biotic interactions and thus
typically incorporates processes of both high-frequency
and low-frequency variability.
Multivariate time-series analysis
A well-established and well-studied relationship
between variables measured in Müggelsee was selected
for the application of DFA to a set of relevant time
series. The case study dealt with the abrupt increase in
the abundance of a copepod species, Cyclops kolensis,
when the abundance of a competing, larger species,
Cyclops vicinus, fell below a critical threshold triggered
by, for example, a change in the abundance of crypto-
phytes, a major food source for the juvenile stages of
both Cyclops species (Scharfenberger et al. 2013).
A set of models was evaluated for the copepod
species case study to test different numbers of common
patterns (m) ranging from 1 to n − 1 (where n is the
number of time-series involved in a particular analysis,
in this case 3) as well as different structures for the
covariance matrix R. Because the number of model
parameters increases drastically when using an uncon-
strained R matrix, we assumed the time series involved
in our multivariate analyses had either an equal
(diagonal and equal) or different (diagonal and unequal)
observation variance but no covariance between time
series. The best model was then selected based on the
Akaike information criterion (AIC; Akaike 1974) as
well as on model inspection of plotted patterns versus
observed values and a visualization of the magnitudes
and signs of factor loadings.
DOI: 10.5268/IW-6.3.948
287
Effect of gaps on pattern extraction in time-series analysis
Gaps were introduced in one representative univariate
time series (total algal biomass) and in the multivariate
copepod species case described earlier (Scharfenberger et
al. 2013). Gaps were systematically introduced by
deleting 3 months of available data per year. The deleted
data consisted of a large gap comprising 2 consecutive
spring months (Apr and May) and a shorter gap in winter
(Dec). Three different analyses with different sampling
resolutions (weekly, 2-weekly, and 4-weekly) were
performed to determine changes to the observed patterns
(previous models without artificial gaps) when data gaps
are larger or evenly distributed across time (i.e.,
systematic missing observations at specific times of the
year). In the copepod species case study, gaps were
introduced only in the driving variable (i.e., cryptophyte
availability).
Effect of sampling representativeness
Using the same univariate time series (total algal
biomass) and multivariate copepod species examples
mentioned earlier, we compared the results of DFA
models for 4-weekly and seasonal observations taken at
those specific intervals (i.e., discrete, equidistant data
points measured every 4 weeks and every season, respec-
tively) with the results of models that averaged the
weekly data over 4-weekly and seasonal intervals. This
simple exercise aimed to assess how representative a
sample could be in the case of discrete, low-frequency
sampling compared to the case in which all available
information (i.e., weekly data) was used by simply
averaging the observations over the 4-weekly or seasonal
intervals.
DFA model evaluation
Comparing weekly, 2-weekly, and 4-weekly data-resolu-
tion models might not be straightforward because these
models seem to extract different temporal patterns,
including long-term trends and seasonal and other cyclic
patterns, although the patterns extracted might vary
depending on the type of variable. The resulting models
for differing data resolutions can first be assessed
visually; because DFA is analogous to linear regression,
the same tools can be applied for model validation (Zuur
et al. 2007). The first tool plots the resulting best fits to
the observed values to identify the extent to which the
model can capture the patterns in the original time series.
The MARSS R-package assesses different model fits
by means of AIC. AIC was used in the copepod species
case study, where the best model fit with m patterns had
to be selected from among a set of different models. In
addition, factor loadings and values in the covariance
matrix R (i.e., the observation error matrix containing
Inland Waters (2016) 6, pp.284-294
288 Aguilera et al.

variance values) were used to assess the strength of a

model in extracting underlying patterns in our univariate
and multivariate time-series analyses.

Results

Effect of reducing sampling resolution in

univariate time-series

The reduction in resolution (from weekly to 2-weekly to

4-weekly) did not significantly change the DFA results
for near-surface water temperature (Fig. 1, 4-weekly
data), which exhibited relatively little high-frequency
variability (at least at weekly sampling resolution).
For the ammonium concentration data, the analysis at
a 4-weekly data resolution revealed a slight decrease
over time since the beginning of the sampling program
(Fig. 2a), but neither interannual nor seasonal variability
could be captured (Fig. 2b). The amount of unexplained
information was fairly large in the 4-weekly resolution
model, which was also evident in the fit (Fig. 2b). For
DO, the 4-weekly resolution data analysis resulted in a
constant zero pattern across the 3 decades of data (Fig.
3), also the case for pH (results not shown). The amount
of unexplained information, as reflected in the covariance Fig. 2. Patterns (lines) derived from dynamic factor analyses fitted
matrices, was thus significantly larger for the 4-weekly to the standardized (a) weekly and (b) 4-weekly data points (dots)
data models. for ammonium concentration in Müggelsee (measured at 1 m depth).
The analysis of total algal biomass in Müggelsee
revealed a drastic change in the overall pattern of the
time series after the resolution was reduced from weekly
(Fig. 4a) to 2-weekly (Fig. 4b), and 4-weekly (Fig. 4c).
Further, the variance value was higher for the 4-weekly
resolution model (0.84) than for the 2-weekly (0.51) and
weekly (0.25) models, implying that the observation
variance in the 4-weekly resolution model was at least 3
times higher than in the weekly resolution model.

Fig. 1. Pattern (line) derived from dynamic factor analysis fitted to Fig. 3. Patterns (lines) derived from dynamic factor analyses fitted to
the standardized 4-weekly data points (dots) for near-surface water the standardized (a) weekly and (b) 4-weekly data points (dots) for
temperature in Müggelsee. dissolved oxygen concentration in Müggelsee.

© International Society of Limnology 2016 DOI: 10.5268/IW-6.3.948

Effects of sampling
Author/sresolution and data gaps of limnological data on pattern recognition
Reducing sampling resolution in multivariate
time-series analysis
In the case study of the coexisting copepod species C.
vicinus and C. kolensis and the cryptophyte biomass
(Scharfenberger et al. 2013), model selection based on AIC
implied that both weekly and 2-weekly data were best
described by 2 common patterns and a diagonal and
unequal covariance matrix. According to the factor
loadings, C. vicinus and the cryptophyte biomass (Fig. 5b)
seemed to be best described by Pattern 1, which decreased
across time. By contrast, C. kolensis (Fig. 5b) was best
characterized by Pattern 2, which tended to be the inverse
Fig. 4. Patterns (lines) derived from dynamic factor analyses fitted to
the standardized (a) weekly, (b) 2-weekly, and (c) 4-weekly data
points (dots) for total algal biomass in Müggelsee.
DOI: 10.5268/IW-6.3.948
289
of Pattern 1. The diagonal values in the R matrix indicated
that the selected model satisfactorily explained the
variability in the time series of the 2 copepod species (0.54
for C. vicinus and 0.12 for C. kolensis) but not the
variability in the cryptophyte biomass time series, which
had a high value of unexplained variance in the covariance
matrix R (0.76).
The 4-weekly data yielded one common pattern; C.
vicinus and the cryptophyte biomass followed a decreasing
trend with time, and C. kolensis followed an increasing
trend related to a negative factor loading (dotted line in Fig.
6a). The factor loadings of the 4-weekly resolution model
(Fig. 6b) were about one order of magnitude lower than
those in the weekly and 2-weekly resolution models.
Overall, the variance reflected in the 4-weekly resolution
model (average diagonal values in R matrix = 0.92) was
about twice that in the weekly model, indicating that a
significant amount of information was lost by using
4-weekly biomass data rather than data of higher resolution.
Unexplained information in time-series analysis
with additional data gaps
In the univariate model of total algal biomass, the amount
of unexplained information (diagonal values in the R
Fig. 5. The best dynamic factor analysis model for the weekly data
of the copepod species case study yielded 2 common patterns. (a)
Temporal variability of the 2 extracted patterns. (b) Pattern 1 (dark
grey) mainly describes the time-series of C. vicinus and crypto-
phytes, whereas Pattern 2 (black) describes the temporal variability
of the abundance of C. kolensis, based on factor loading magnitudes.
Inland Waters (2016) 6, pp.284-294
290
Table 2. Amount of unexplained information in univariate and multivariate DFA models for data of weekly, 2-weekly, and 4-weekly sampling
resolution with and without artificially introduced gaps.
DFA Model m R
Total algal biomass (no artificial gaps) 1 0.25
Total algal biomass (with artificial gaps) 0.35
Copepod abundances (no artificial gaps) 2 0.54
Copepod abundances (with artificial gaps) 0.78
m = number of patterns; R = amount of unexplained information (variance value in covariance
abundance data from Scharfenberger et al. (2013).
covariance matrix) in the weekly resolution model with
artificial gaps increased by 40% (weekly data) compared to
the previous weekly resolution model without such gaps
(Table 2). The increase in unexplained information was
lower in the case of the 2-weekly (27.5%) and 4-weekly
(4.8%) resolution models. These sampling resolutions
initially yielded a large amount of unexplained information
in the previous models (0.51 and 0.84; Table 2), and
therefore the difference in the corresponding artificial gap
models was not as drastic as in the case of weekly data.
Nevertheless, the explanatory power of the 2-weekly and
4-weekly data was much lower than that of the weekly
data.
In the multivariate copepod species model, the
covariance R value increased by 44.4% and 43.9% for the
Fig. 6. The best model after dynamic factor analysis for the 4-weekly
data of the copepod species case study yielded one common pattern.
(a) The dotted line represents the inverse pattern related to the
negative factor loading of C. kolensis shown in (b).
© International Society of Limnology 2016
Aguilera et al.
Weekly
2-weekly 4-weekly
m R m R
1 0.51 1 0.84
0.65 0.88
2 0.57 1 0.92
0.82 0.93
matrix). Copepod
weekly and 2-weekly data resolution models, respectively
(Table 2). The 4-weekly model increased the amount of
unexplained information by only 1.1% because the initial
variance was already high (from R = 0.84 to 0.88). The
total length of gaps introduced in all 3 temporal resolution
cases decreased the number of observations by an average
of 24.6% (range 24.3–25.0%) compared to the correspond-
ing models of the same case study without artificial gaps.
Despite these differences, the extracted patterns and
resulting best fits for models with and without artificial
gaps did not differ substantially from one another; in other
words, the introduction of gaps in the cryptophyte time
series did not substantially alter the shape and behavior of
the patterns extracted by DFA.
Unexplained information and sampling
representativeness
For both univariate and multivariate data, DFA models
based on data obtained by averaging the original
time-series over either 4-weekly or seasonal intervals out-
performed DFA models based on data points simply
selected from the original time series at the same 2
intervals. The values of variance in the R matrices for the
data selection models are considerably larger than those for
the data-averaging models (Table 3). These differences in
variance are notably larger for the seasonal data than for
the 4-weekly data. Overall, our results confirm that, as
expected, averaging the available observations in a time
series to obtain lower-frequency measurements is
preferable to selecting individual values from the time
series at specific target intervals.
Discussion
The responses of lakes to environmental change are
complex, occur on a variety of temporal scales, and can
be associated with changes that occur within critical
short-term time windows within a season (Adrian et al.
2012). Many past and ongoing long-term research
programs provide high-quality data at short temporal
DOI: 10.5268/IW-6.3.948
Effects of sampling resolution and data gaps of limnological data on pattern recognition
Table 3. Amount of unexplained information in univariate and multivariate DFA models with differing resampling strategies (i.e., data
selection as opposed to data averaging).
DFA Model
Total algal biomass (data selection)
Total algal biomass (data averaging)
Copepod abundances (data selection)
Copepod abundances (data averaging)
m = number of patterns; R = amount of unexplained information (variance value in covariance
matrix). Copepod abundance data from Scharfenberger et al. (2013).
scales. For instance, automated in situ measurements
such as those of the Global Lake Ecological Observatory
Network (GLEON) or Networking Lake Observatories
in Europe (NETLAKE) operate on subhourly temporal
scales (e.g., Hamilton et al. 2014). Most environmental
datasets, however, rely heavily on lower-frequency
sampling campaigns such as that conducted at
Müggelsee. Furthermore, time series usually contain
gaps of different lengths, distributed irregularly with
time. The role of the temporal resolution of long-term
data compared with the distribution of gaps determines
what kind of questions can be addressed to better
understand system dynamics and their response to envi-
ronmental change in the long term. Here, we addressed
the question of the temporal resolution needed to capture
the variability in driving and response variables or the
interaction between them. Within an ecological context,
the ability to deal with irregular sampling data in
univariate or multivariate time series would be
considered an important improvement in time-series
analysis methods (Cazelles et al. 2008). DFA, which
provides the means to analyze time series with varying
sampling resolutions and missing observations,
succeeded in capturing general long-term trends in
physical, chemical, and biological limnological variables
but failed to capture the seasonal variability and
associated interactions between driving and response
variables, particularly when biological data were
involved.
Sampling resolution reduction in
physicochemical and biological variables
Predictably, decreasing the data resolution from weekly
to 4-weekly lowered the capability of the DFA models to
explain the underlying short-term patterns; however, the
decrease in the amount of unexplained information
varied. In the case of an abiotic variable such as near-
surface water temperature, which is usually dominated
by low-frequency variability, 4-weekly data were
sufficient to capture the main long-term pattern and
DOI: 10.5268/IW-6.3.948
291
4-weekly Seasonal
m R m R
0.84 0.81
1 1
0.78 0.64
0.92 0.90
1 1
0.88 0.81
seasonality. This was not the case for the pH and DO
models, which presumably show more high-frequency
variability than does water temperature when considered
at broader temporal scales. The information contained in
the time series of ammonium concentration and total
algal biomass, however, was best explained by the
weekly data model. The aliasing effect caused by under-
sampling is much stronger in processes with a significant
power of high frequency components. Algal growth
(Reynolds 1984) and nutrient dynamics operate on
temporal scales of days or less (Lampert and Sommer
2007), indicating the need, in these cases, for a higher
sampling frequency to avoid or reduce the effect of
aliasing.
Overall, in the univariate time-series analyses of the
chemical and biological variables, we observed that
4-weekly resolution might not suffice to describe the
variability that occurs within short time windows.
Addressing temporal scale is especially relevant in
polymictic lakes, which, unlike dimictic and monomictic
lakes, are subject to frequent changes in thermal regime
during summer. In Müggelsee, for instance, although 79%
of all stratification events last <1 day, the frequency and
period of stratification are predicted to increase (Wilhelm
and Adrian 2008). This change in stratification has
potential consequences for plankton development in the
lake; stratification periods exceeding 3 weeks, for instance,
have been shown to favor the proliferation of cyanobacte-
ria (Wagner and Adrian 2009), and severe climate-related
changes in plankton diversity have occurred on temporal
scales of 2–3 weeks (Wagner and Adrian 2011).
Sampling resolution reduction in multivariate
analysis
In the multivariate copepod species case, the best DFA
models usually yielded fewer common patterns for the
4-weekly data resolution, indicating a reduced capability
of the models to capture interactions apparent in data
sampled at higher resolution (i.e., weekly and 2-weekly).
The 4-weekly data were thus unable to capture the
Inland Waters (2016) 6, pp.284-294
292
underlying common patterns among the 3 interacting
time series of copepod species abundances and
cryptophyte biomass. Note that the overall long-term
pattern in copepod abundances seemed to coincide with
the results of Scharfenberger et al. (2013), who stated
that the decline in C. vicinus followed the driver
threshold scenario of regime shift theory (Andersen et al.
2009), caused by an abrupt change in a driver (in this
case the decline in cryptophyte prey availability),
whereas concomitantly C. kolensis benefited from this
situation and increased in abundance in the decades
following the shift event. The failure to capture the
dynamics in zooplankton development and interactions
critically depends on species-specific life history traits.
Fast-growing species such as cladocerans and rotifers
exhibit life-history durations of only a couple of days
compared to >4 weeks for copepods (Adrian et al. 2006).
Thus, the population dynamics of cladocerans and
rotifers are more easily missed if sampling is at low
temporal resolution. Nevertheless, for copepods with
long and complex life histories, entire generations can be
missed if sampling intervals are in the range of a month
(Adrian et al. 2006). Despite the relative dearth of
information obtained by analyzing 4-weekly observa-
tions, DFA could still extract long-term, general patterns
and, in some cases, intraannual and seasonal patterns in
our data (e.g., Fig. 4c and 6a).
Dealing with data gaps and sampling
representativeness
The introduction of additional artificial gaps into the data
used for the univariate total algal biomass models and the
multivariate copepod species case study models decreased
the amount of information that could be explained by the
pattern(s) extracted by DFA (Table 2). The increased
observation variance, however, varied according to the
temporal resolution of the data and was more pronounced
in the weekly data models in both the univariate and mul-
tivariate cases. In real terms, the introduction of gaps for 2
consecutive spring months would be especially critical in
real sampling programs for most biological variables such
as algal or cladoceran biomass because spring is a key
time of year when rapid, short-term changes in external
driving variables such as air temperature and day length
usually take place. Nevertheless, even in winter (a season
much neglected because of the notion that not much
happens), variability in the phytoplankton can be high at
short temporal scales (Özkundakci et al. 2016). All of this
was reflected in the loss of information found in the
weekly models with data gaps.
Intuitively, averaging the highest-frequency observa-
tions to rescale the data to a lower frequency retains a
© International Society of Limnology 2016
Aguilera et al.
larger portion of the available information, whereas the
practice of selecting observations at specific target
intervals excludes much of the available information. We
acknowledge that the weekly sampling frequency in this
study might be below the Nyquist frequency for many
biotic and abiotic processes in limnology. Despite this
limitation, however, we strongly recommend that the
sampling frequency be tailored to include the potentially
distorting effects of aliasing when designing monitoring
schemes. Also noteworthy is that, although high-fre-
quency (e.g., daily, hourly) measurements of biotic
variables would be preferred to guarantee sampling rep-
resentativeness, most measurements of biotic variables
rely on spot-sample campaigns, which are undertaken at
lower frequencies.
Several studies have suggested that long-term routine
monitoring, even at low sampling frequencies, is
important with respect to slow ecological processes, rare
or episodic phenomena, highly variable processes, and
subtle or complex phenomena. For instance, monthly
averaged stream nitrate data in the tropical forest of the
Luquillo Mountains (Puerto Rico) allowed researchers to
quantify the magnitude of the impact of Hurricane Hugo
on water chemistry and to assess the speed at which the
system approached background conditions after the
hurricane hit the island in 1989 (Schaefer et al. 2000,
Dodds et al. 2012). In an analysis of watershed charac-
teristics, Halliday et al. (2012) found that 4-weekly data
allowed the analysis of long-term trends and provided
new insights into the changing amplitude and phase of
the seasonality of water quality variables. Furthermore,
they observed that low-frequency (e.g., monthly) data
provided background information on their study basin
and on the processes occurring necessary to interpret the
complexity of the high-frequency data.
Overall, low-frequency data allow the detection of
long-term trends and seasonality in those cases in which
the power in the low-frequency components of the
processes is much higher than that in the high-frequency
components, specifically in cases in which the aliasing
effect does not significantly affect characterization of the
large-scale processes. High-frequency data are key in the
identification of extremes, short-term trends, and
subdaily variability in limnological variables. High-fre-
quency data are also needed when the power in the high-
frequency components is comparable to that in the low
frequency components; in this case the aliasing effect
caused by undersampling is much stronger. A
compromise between financial and logistical resources
and sample representativeness must be reached to
maximize the information obtained from the observa-
tions taken at the chosen sampling frequency.
DOI: 10.5268/IW-6.3.948
Effects of sampling resolution and data gaps of limnological data on pattern recognition
Acknowledgements
R. Aguilera benefited from a Short-Term Scientific
Mission at the Leibniz-Institute of Freshwater Ecology
and Inland Fisheries, Berlin, funded by the ESSEM COST
Action ES1201 NETLAKE (Networking Lake Observato-
ries in Europe). D.M. Livingstone benefited from time
spent as guest scientist at the Leibniz-Institute of
Freshwater Ecology and Inland Fisheries, Berlin, funded
by the IGB Fellowship Program in Freshwater Science. R.
Adrian was supported through the German Science
Foundation project LakeRisk (AD 91/13-1).
References
Adrian R, Wilhelm S, Gerten D. 2006. Life-history traits of lake
plankton species may govern their phenological responses to climate
warming. Glob Change Biol. 12:652–661.
Adrian R, Gerten D, Huber V, Wagner C, Schmidt S. 2012. Windows of
change: temporal scale of analysis is decisive to detect ecosystem
responses to climate change. Mar Biol. 159:2533–2542.
Aguilera R, Marcé R, Sabater S. 2015. Detection and attribution of
global change effects on river nutrient dynamics in a large Mediterra-
nean basin. Biogeosciences. 12:4085–4098.
Akaike H. 1974. A new look at the statistical model identification. IEEE
T Automat Cont. 19:716–23.
Andersen T, Carstensen J, Hernandez-Garcia E, Duarte CM. 2009.
Ecological thresholds and regime shifts: approaches to identification.
Trend Ecol Evolut. 24:49–57.
Cazelles B, Chavez M, Berteaux D, Ménard F, Vik JO, Jenouvrier S,
Stenseth NC. 2008. Wavelet analysis of ecological time series.
Oecologia. 156:287–304.
Driescher E, Behrendt H, Schellenberger G, Stellmacher R. 1993. Lake
Müggelsee and its environment—natural conditions and anthropo-
genic impacts. Int Rev Gesamt Hydrobiol Hydrogr. 78:327–343.
Dodds WK, Robinson CT, Gaiser EE, Hansen GJ, Powell H, Smith
JM, Morse NB, Johnson SL, Gregory SV, Bell T, Kratz TK. 2012.
Surprises and insights from long-term aquatic data sets and
experiments. BioScience. 62:709–721.
Durbin J, Koopman SJ. 2012. Time series analysis by state space
methods (No. 38). Oxford University Press.
Foster G. 1996. Wavelets for period analysis of unevenly sampled time
series. Astronom J. 112:1709.
Gerten D, Adrian R. 2000. Climate-driven changes in spring plankton
dynamics and the sensitivity of shallow polymictic lakes to the North
Atlantic Oscillation. Limnol Oceanogr. 45:1058–1066.
Halliday SJ, Wade AJ, Skeffington RA, Neal C, Reynolds B, Rowland
P, Neal M, Norris D. 2012. An analysis of long-term trends,
seasonality and short-term dynamics in water quality data from
Plynlimon, Wales. Sci Total Environ. 434:186–200.
Hamilton DP, Carey C, Arvola L, Arzberger P, Brewer C, Cole JC,
Gaiser E, Hanson PC, Ibelings BW, Jennings E, et al. 2014. A Global
Lake Ecological Observatory Network (GLEON) for synthesising
DOI: 10.5268/IW-6.3.948
293
high-frequency sensor data for validation of deterministic ecological
models. Inland Waters. 5:49–56.
Harvey AC. 1989. Forecasting, structural time series models and the
Kalman filter. Cambridge University Press.
Hirsch RM, Slack JR, Smith RA. 1982. Techniques of trend analysis for
monthly water-quality data. Water Resour Res. 18:107–121.
Holmes EE. 2013. Derivation of the EM algorithm for constrained and
unconstrained multivariate autoregressive state-space (MARSS)
models. Technical Report. arXiv:1302.3919.
Holmes EE, Ward EJ, Wills K. 2012. MARSS: Multivariate Autore-
gressive State-Space models for analyzing time-series data. R
Journal. 4:11–19.
Holmes EE, Ward EJ, Wills K. 2013. MARSS: Multivariate Autore-
gressive State-Space modeling. R package 3.4.
Jennings E, Jones S, Arvola L, Staehr PA, Gaiser E, Jones ID, Weathers
KC, Weyhenmeyer GA, Chiu C, De Eyto E. 2012. Effects of
weather-related episodic events in lakes: an analysis based on high–
frequency data. Freshwater Biol. 57:589–601.
Köhler J, Hilt S, Adrian R, Nicklisch A, Kozerski HP, Walz N. 2005.
Long-term response of a shallow, moderately flushed lake to reduced
external phosphorus and nitrogen loading. Freshwater Biol. 50:1639–1650.
Kuo YM, Jang CS, Yu HL, Chen SC, Chu HJ. 2013. Identifying
nearshore groundwater and river hydrochemical variables
influencing water quality of Kaoping River Estuary using dynamic
factor analysis. J Hydrol. 486:39–47.
Lampert W, Sommer U. 2007. Limnoecology: the ecology of lakes and
streams. 2nd ed. Oxford University Press.
Muñoz-Carpena R, Ritter A, Li YC. 2005. Dynamic factor analysis of
groundwater quality trends in an agricultural area adjacent to
Everglades National Park. J Contam Hydrol. 80:49–70.
Özkundakci D, Gsell AS, Hintze T, Täuscher H, Adrian R. 2016. Winter
severity determines functional trait composition of phytoplankton in
seasonally ice covered lakes. Glob Change Biol. 22:284–298.
Parr TW, Siera ARJ, Battarbee RW, Mackay A, Burgess J. 2003.
Detecting environmental change: science and society—perspectives
on long-term research and monitoring in the 21st century. Sci Total
Environ. 310:1–8.
R Core Team. 2012. R: a language and environment for statistical
computing. R Foundation for Statistical Computing, Vienna
(Austria). ISBN 3-900051-07-0. http://www.R-project.org/
Reynolds CS. 1984. The ecology of freshwater phytoplankton.
Cambridge University Press.
Rose KC, Winslow LA, Read JS, Read EK, Solomon CT, Adrian R,
Hanson PC. 2014. Improving the precision of lake ecosystem
metabolism estimates by identifying predictors of model uncertainty.
Limnol Oceanogr Meth. 12:303–312.
Schaefer DA, McDowell WH, Scatena F, Asbury CE. 2000. Effects of
hurricane disturbance on stream water concentrations and fluxes in
eight tropical forest watersheds of the Luquillo Experimental Forest,
Puerto Rico. J Trop Ecol. 16:189–207.
Scharfenberger U, Mahdy A, Adrian R. 2013. Threshold–driven shifts
in two copepod species: testing ecological theory with observational
data. Limnol Oceanogr. 58:741–752.
Inland Waters (2016) 6, pp.284-294
294
Schoellhamer DH. 2001. Singular spectrum analysis for time series
with missing data. Geophys Res Lett. 28:3187–3190.
Shumway RH, Stoffer DS. 1982. An approach to time series smoothing
and forecasting using the EM algorithm. J Time Ser Anal. 3:253–264.
Shumway RH, Stoffer DS. 2010. Time series analysis and its applica-
tions: with R examples. New York (NY): Springer.
Solomon CT, Bruesewitz DA, Richardson DC, Rose KC, Van de
Bogert MC, Hanson PC, et al. 2013. Ecosystem respiration: drivers
of daily variability and background respiration in lakes around the
globe. Limnol Oceanogr. 58:849–866.
Vilizzi L. 2012. Abundance trends in floodplain fish larvae: the role of
annual flow characteristics in the absence of overbank flooding. Fund
Appl Limnol Arch Hydrobiol. 181:215–227.
Wagner C, Adrian R. 2009. Cyanobacteria dominance: quantifying the
effects of climate change. Limnol Oceanogr. 54:2460–2468.
Wagner C, Adrian R. 2011. Consequences of changes in thermal regime
for plankton diversity and trait composition in a polymictic lake: a
matter of temporal scale. Freshwater Biol. 56:1949 –1961.
© International Society of Limnology 2016
Aguilera et al.
Wilhelm S, Adrian R. 2008. Impact of summer warming on the thermal
characteristics of a polymictic lake: consequences for oxygen,
nutrients and phytoplankton. Freshwater Biol. 53:226 –237.
Zayed A. 1993. Advances in Shannon’s sampling theory. Taylor &
Francis.
Zuur AF, Fryer RJ, Jolliffe IT, Dekker R, Beukema JJ. 2003a.
Estimating common trends in multivariate time series using dynamic
factor analysis. Environmetrics. 14:665–685.
Zuur AF, Tuck ID, Bailey N. 2003b. Dynamic factor analysis to
estimate common trends in fisheries time series. Can J Fish Aquatic
Sci. 60:542–552.
Zuur AF, Pierce GJ. 2004. Common trends in northeast Atlantic squid
time series. J Sea Res. 52:57–72.
Zuur AF, Ieno EN, Smith GM. 2007. Analysing ecological data (Vol.
680). New York: Springer.
DOI: 10.5268/IW-6.3.948