1. Introduction
The continuous development of water quality sensors has led to a transition from studying long-term trends and seasonal patterns using the time series of monthly or weekly grab samples to the investigation of highly dynamic phenomena, such as storm events and diurnal patterns, using high-frequency in situ measurements. In the past, such events were studied through intensive sampling campaigns, but these studies were often of short duration (e.g., 2–12 days) [
1,
2,
3,
4,
5,
6]. With the currently available technology and decreasing costs, in situ sensors are more frequently used for longer periods, ranging from months, e.g., [
7,
8] to several years, e.g., [
9,
10,
11]. Although large amounts of data present challenges regarding storage, processing, and analysis [
12], longer term datasets provide an opportunity for detailed investigations of hydrological and biogeochemical processes in dynamic systems, especially in remote areas [
13,
14,
15]. Compared to terrestrial ecosystems, where processes such as primary productivity and respiration are highly correlated to climate variables, aquatic ecosystems are more complex. This is a result of seasonal variation in light supply for photosynthesis through seasonal changes in shading by riparian vegetation, temporal changes in autotroph biomass due to floods and droughts, and allochthonous inputs of detritus and organic matter [
16]. Long-term studies improve our understanding of the potential effect of land use and climate change on river metabolism and the delivery of water-related ecosystem services [
16].
Many terrestrial and in-stream biogeochemical processes are driven by the incidence of solar energy, resulting in distinct patterns on a diel, i.e., a 24-hour timescale. The diel change in solar energy can result in diel patterns in stream temperature, especially in shallow, wide and less shaded streams. Additionally, photosynthetic activity is driven by solar radiation, influencing the concentration of dissolved oxygen (DO) and carbon dioxide (CO
2) in the stream [
17]. In addition, biological activity can result in diel patterns in, for example, dissolved organic carbon (DOC) [
6,
18,
19], phosphorus [
6,
18,
20], and nitrate [
20,
21,
22] through uptake and respiration. The study of diurnal fluctuations is one of the subjects for which the use of in situ sensors is particularly suitable, since current technology allows measuring at intervals of seconds or minutes.
Spectrophotometers measuring in the UV-visible spectrum, also referred to as UV-Vis sensors, can be used to investigate diel patterns for DOC and nitrate. These sensors use algorithms to calculate solute concentrations based on absorbance at a specific wavelength or multiple wavelengths. Although the use of in situ UV-Vis sensors presents several challenges, such as biofouling [
23], local calibration [
24,
25], and power supply [
24,
26], numerous studies have used such sensors in the field [
20,
21,
27,
28,
29,
30]. In most of these studies, grab samples analyzed in the laboratory were used to validate data recorded by the sensor [
24,
31]. This sampling, however, does not allow to check the validity of high-frequency patterns, such as diurnal fluctuations.
We deployed four UV-Vis sensors (spectro::lyser, s::can Messtechnik GmbH, Vienna, Austria) to understand nitrate dynamics in streams draining different land use types (tropical montane forest, smallholder agriculture and commercial tea plantations) in the South West Mau region in western Kenya [
9]. Diurnal patterns in the nitrate concentration differed between sites and seasons (i.e., rainy season and dry season), suggesting an influence of land use and seasonality on in-stream biogeochemical processes. However, abrupt changes in these diurnal patterns were observed when the position of a sensor was adjusted to facilitate measurements during very low flows. Furthermore, different sensors recorded different patterns at the same site. These observations led to the suspicion that some of these ‘diurnal patterns’ could be artefacts, rather than the manifestation of biological processes. To assess the validity of the observed diurnal patterns, we used a mobile set-up, whereby a second UV-Vis sensor was installed next to an existing monitoring system. The mobile sensor was installed parallel to the fixed UV-Vis sensor for at least two weeks. Afterwards, we shaded or changed the position of the mobile sensor to a different depth and/or orientation for at least another week to investigate whether these changes influenced the measurements.
The article aims to present evidence to challenge the interpretation of diurnal patterns in nitrate and DOC concentrations measured by in situ UV-Vis sensors in tropical headwater streams with low solute concentrations. A combination of rapidly changing environmental conditions, e.g., intense sunlight, fluctuations in water level, stream temperature, and turbidity, could result in unexpected artefacts in the data. These have, to our knowledge, not been documented in previous studies. Yet, this information is essential to guide other users of in situ UV-Vis sensors in the interpretation of their data. Although the results of our experiment and data analysis are not conclusive, we provide explanations on potential causes for the inconsistencies in diurnal patterns, especially when solute concentrations are low. We propose how data collected with in situ UV-Vis sensors can still be used to increase the understanding of hydrological and biogeochemical processes. Furthermore, we make suggestions on methods that can be used to further investigate this issue in field and laboratory experiments.
2. Materials and Methods
2.1. Study Area
This study was carried out in the South West Mau region in western Kenya, part of the largest remaining indigenous tropical montane forest in East Africa. The outlets of three sub-catchments (27–36 km
2) within the 1021 km
2 Chemosit catchment were instrumented (
Table 1), following a nested catchment approach. Each sub-catchment drains an area characterized by one of the dominant land use types in the region: tropical montane rainforest (NF), smallholder agriculture (SHA), and commercial tea and tree plantations (TTP). The outlet of the Chemosit catchment, referred to as the main catchment (OUT), was instrumented as well. Elevation in the study area ranges from 1715 m a.s.l. to 2932 m a.s.l. Soils are classified as humic Nitisols [
32]. The geology is characterized by phonolitic nephelinites in the upper part and phonolites in the lower part of the catchment [
33,
34]. The annual precipitation is 1988 ± 328 mm for the years 1905 to 2014 at 2100 m a.s.l. [
35]. A more detailed description of the study area is provided in Jacobs et al. [
35].
2.2. Instrumentation
Automatic measurement stations were installed at the outlets of the three sub-catchments (NF, SHA and TTP) in October 2014 and at the main catchment (OUT) in April 2015. Each station consists of a radar-based water level sensor (VEGAPULS WL61, VEGA Grieshaber KG, Schiltach, Germany) and a UV-Vis spectrophotometer (spectro::lyser, s::can Messtechnik GmbH, Vienna, Austria), measuring turbidity, total and dissolved organic carbon (TOC, DOC), and nitrate (NO3-N) with a wavelength range of 220 to 720 nm, a resolution of 2.5 nm, and a 5 mm optical path length. In addition, the sensor measures stream temperature. The measurement range and accuracy for nitrate, when measured with this path length, are 0 to 60 mg N L−1 and ±2% + 0.2 mg N L−1, respectively, according to the sensor documentation. The measurement range for DOC is 0 to 84 mg C L−1, but no information on accuracy is provided by the manufacturer. Measurements are taken every 10 min. An automated cleaning system uses bursts of pressurized air to remove any particles from the sensor window before each measurement. In addition, the sensors are cleaned manually on a weekly to bi-weekly basis to reduce the influence of more persistent fouling.
The sensor manufacturer provides no specific requirements for the orientation of the sensor. The UV-Vis sensors at NF, SHA, and OUT were installed at a 45° angle at the downstream face of a concrete block in the riverbank. This position protects the sensor from obstruction and damage by woody debris and stones during high flows. Because a similar construction was not possible at TTP, the sensor was installed vertically against the rocky riverbank in a fast-flowing section of the stream. During very low flow (e.g., dry season in February and March), the sensors were mounted in a horizontal position approximately 5 cm above the riverbed. To avoid the trapping of air bubbles from the automated cleaning system at the sensor window, we decided to install the sensor in a way that the measurement gap faces downstream, rather than facing the riverbed. As a consequence, UV radiation from the sun could potentially reach the measurement window during measurements. Due to sensor failure (e.g., energy loss of Xenon lamp, internal dark noise error, corrosion of measurement window), sensors have been replaced and repaired several times at each site (
Figure A1). Sensors A, B, and F were rotated between the sites NF, TTP, and OUT on 2 May 2017.
Rainfall was recorded with tipping buckets (Theodor Friedrichs, Schenefeld, Germany, and ECRN-100 high-resolution rain gauge, Decagon Devices, Pullman, WA, USA) as cumulative precipitation in 10-minute intervals (0.2 mm resolution) at nine sites in the study area. To calculate the total rainfall within a (sub-)catchment, each tipping bucket was assigned a weight based on Thiessen polygons. Malfunctioning tipping buckets were temporarily excluded, and the weights of the remaining tipping buckets were adjusted.
2.3. Experimental Set-Up
The performance of multiple sensors is best compared by installing all sensors at the same site, but this was logistically not feasible. Instead, an extra sensor (sensor G), referred to as the mobile sensor, was used to test the comparability of patterns recorded by the sensors installed at three of the four measurement stations (fixed sensors). The same sensor had been deployed previously as fixed sensor at TTP and OUT (
Figure A1). The mobile sensor was installed at each site for a minimum of three weeks before being moved to another site (
Table 2). Installation failed at OUT because of problems with the power supply. The mobile sensor was connected to a control box provided by the manufacturer. This control box was then connected to the 12V power supply and to the data logger (con::cube, s::scan Messtechnik GmbH, Vienna, Austria) for data transmission. The mobile sensor was connected to the pressurized cleaning system with the same cleaning frequency as the fixed sensor. Measurements were taken concurrently by both sensors. The mobile sensor was initially installed in parallel to the fixed sensor. After at least two weeks, either the position of the sensor was changed (depth and/or orientation) or the measurement window of the mobile sensor was shaded (
Table 2) to test how this would affect the measurements. Each treatment lasted a minimum of three days.
2.4. Data Processing
All data were subjected to a processing protocol, which was previously applied in Jacobs et al. [
9]. In summary, time stamps with NA values or those indicated as errors by the internal data logger software (moni::tool, s::can Messtechnik GmbH, Vienna, Austria) were flagged automatically. In addition, observations during field visits were used to manually flag periods with unreliable data due to, e.g., burial by sediment, too low water level or a problem with the automatic cleaning system. The median absolute deviation (MAD) for a rolling window of 16 measurements was used to identify outliers [
9,
36]. All flagged data were omitted from further analysis. After flagging, data gaps of <6 h were filled using linear interpolation. For easier identification of diurnal patterns, noise in the measured data was removed by applying a rolling mean with a window width of 3 h.
Discharge was estimated using a site-specific rating curve developed using individual discharge measurements over the full range of measured water levels, see Jacobs et al. [
9]. During weekly to bi-weekly maintenance visits, 100 mL grab samples were taken from the streams, filtered immediately using <0.45 µm polypropylene filters (KX syringe filter, Kinesis Ltd., St. Neods, UK) and stored frozen until analysis in the laboratory of Justus Liebig University Giessen, Germany. The grab samples were analyzed for nitrates using ion chromatography (ICS-2000, Dionex, Sunnyvale, CA, USA) and for dissolved organic carbon (TOC cube, Elementar Analysensysteme GmbH, Hanau, Germany). These data were used to check and calibrate the nitrate and DOC concentrations recorded by the sensors. For each site, linear regression was used to develop a relationship between the grab samples analyzed in the laboratory and the values measured by the sensors. The reverse of the linear regression equation was then applied to calibrate the full dataset (
Appendix B). All data (raw, processed, grab samples for calibration) are available in Jacobs et al. [
37].
2.5. Data Analysis
A rolling median with a window width of 48 h was applied to the processed dataset to calculate the background concentration of nitrate and DOC. The diurnal patterns were estimated by subtracting the background concentration from the processed dataset, resulting in data representing deviation from background concentration in mg N L−1 for nitrate and mg C L−1 for DOC.
Because we observed seasonal differences in the occurrence of diurnal patterns, we classified each day into categories representing seasons (dry, transition, rainy). Due to interannual variability in the onset and end of the dry and rainy seasons, we chose to use discharge instead of fixed dates or months as an objective indicator for the different seasons. For each site, sub-daily discharge was aggregated to mean daily discharge (Qd). We then classified each day into one of three discharge classes: low flow (Qd ≤ Q70), medium flow (Q70 < Qd ≤ Q30), and high flow (Qd > Q30), representing the dry, transition, and rainy seasons, respectively. Q70 and Q30 are the discharge values exceeded on 70% and 30% of the days. The days were further grouped by sensor to investigate potential sensor-specific patterns. At every 10 min of the day, the median and the interquartile range of the deviation from the background concentration were calculated per site, sensor, and discharge class. These data were plotted for a visual investigation of diurnal patterns in nitrate and DOC concentrations in the stream.
The data obtained during the sensor comparison experiment were investigated by comparing the time series of the deviation from the background concentration. Pearson’s correlation coefficients (r) were calculated for the deviations measured by the fixed and the mobile sensor for each treatment, as this goodness of fit criteria particularly addresses the correct timing of events.