Based on the permanence of flow, streams can be classified into ephemeral, intermittent or perennial using hydrological (e.g. Hedman and Osterkamp, 1982; Uys and O'Keeffee, 1997; Svec et al., 2005; Wohl, 2017) or ecological (e.g. Hansen, 2001; Leigh et al., 2015; Stromberg and Merritt, 2015) indicators. Statistical indicators solely based on the duration of streamflow allow flexible categorisation of streams according to the predominant annual average climatic conditions. For the western United States Hedman and Osterkamp (1982) define streams with measurable annual surface flow for more than 80 % of the time as perennial, whereas streams with surface flow occurring between 80 % and 10 % of the year are categorised as intermittent and as ephemeral when the duration of streamflow per year falls below 10 %. Streamflow classification can also be based on hydrological processes involved in maintaining the flow and the type of sources which provide water to the stream over the year (e.g. Sophocleous, 2002; Nadeau and Rains, 2007). Under normal conditions perennial streams flow throughout the year and receive substantial amounts of water from various sources, including snowmelt and effluent groundwater and sometimes snowmelt (Sophocleous, 2002; Nadeau and Rains, 2007). Intermittent streams cease to flow during dry periods when the state changes from gaining groundwater (effluent) to losing groundwater (influent), while the source of streamflow can originate from various sources (Sophocleous, 2002; Nadeau and Rains, 2007). During dry periods of the year intermittent streams can dry out completely or partially, i.e. along certain stream segments which are then disconnected from the longitudinal flow (Wohl, 2017). In ephemeral streams the groundwater level is always lower than the channel (influent) and therefore flow only occurs in direct response to precipitation (Sophocleous, 2002).

While public agencies rarely monitor discharge in intermittent and ephemeral streams (Uys and O'Keeffe, 1997; Svec et al., 2002), the dynamics of the total stream network and therefore reliable information on the temporal dynamics of the spatial extent is nevertheless of great interest to many hydrological studies (e.g. Godsey and Kirchner, 2014; Shaw, 2016; Stoll and Weiler, 2010; Jensen et al., 2017). Several methods can be used to obtain information about the presence of flow in stream networks, ranging from point measurements at conventional gauging systems or flow monitoring sensors over line feature information through visual assessment during field surveys to spatial information from remote sensing products. Stoll and Weiler (2010) propose the use of remote sensing products as ground truth to verify simulated spatial dynamics of stream networks in ungauged basins. Remote sensing products usually suffer from low temporal resolution on the one hand (as a result of the return interval of the satellites as well as data gaps caused by cloud cover) and from total lack of information for tree-shaded stream reaches on the other hand.

Field surveys of wet vs. dry and subsequent mapping are commonly used to monitor the spatial extent of the stream network (e.g. Godsey and Kirchner, 2014; Shaw, 2016; Goodrich et al., 2018). This is labour intensive and hence the mapping is usually carried out either only seasonally (Godsey and Kirchner, 2014), in uneven intervals during the dry season and directly after strong rainfall events (Day, 1978; Davids et al., 2017), or dependent on the discharge of a reference gauge (Jensen et al., 2017). Recently citizen science approaches have been developed for wet/dry mapping in order to minimise labour-intensive fieldwork (Turner and Richter, 2011; Davids et al., 2017).

Various measurement methods of streamflow at the point scale were established for continuous, automated recording of stage height at weirs or flumes at perennial streams (e.g. Sauer and Turnipseed, 2010). While pressure transducers have been used in studies of temporary stream monitoring (Gupta, 2001; Svec et al., 2005), this monitoring method has some drawbacks: the determination of accurate rating curves and thus correct measurements of zero flow in temporal streams can be challenging due to the unpredictability of the active flow channels within the streambed and often significant erosion or deposition (Shanafield and Cook, 2014). Pressure transducers are also considered to be expensive in cases where the aim of monitoring is not discharge but streamflow duration (Bhamjee and Lindsay, 2011). Monitoring of streambed temperature is a robust and inexpensive method to differentiate between presence and absence of streamflow and can help to overcome the financial drawbacks of conventional gauging systems (e.g. Constantz et al., 2001; Selker et al., 2006; Buttle et al., 2013). Constantz et al. (2001) collected streambed temperature data in longitudinal stream transects in order to monitor the spatial and temporal patterns of streamflow in ephemeral channels. However the sensors needed to be constantly calibrated to air temperature in situations where the streambed and thus the temperature sensor can fall dry (Constantz et al., 2001). Buttle et al. (2013) identify the point scale of a single temperature sensor as a limitation of the method. This can be overcome by fibre optic temperature sensing for continuous monitoring of vertical and longitudinal profiles of stream temperatures (Selker et al., 2006).

Point-scale electric resistance (ER) sensors can provide more accurate estimation of the occurrence of streamflow duration requiring less interpretation compared to temperature measurements (Blasch et al., 2002). ER sensors have been used in many studies to deduce the temporal and spatial dynamics of the occurrence of streamflow in different environments including dry channels and wetlands (e.g. Adams et al., 2006; Goulbra et al., 2009; Bhamjee and Lindsay, 2011; Chapin et al., 2014) and the connectivity of rivers (Jaeger and Olden, 2012). With respect to timing of streamflow the achieved accuracy of the ER sensors was comparable to stream gauge measurements (Blasch et al., 2002). Bhamjee et al. (2016) propose a paired sensor approach combining an ER sensor and a flow-detection sensor, which is based on a flap that opens under water pressure, in order to overcome the limitation in distinguishing between pooling water and flowing water as well as the dry state of the channel.

In recent years the use of image-based methods to identify flow velocity (e.g. Bradley et al., 2002; Muste et al., 2008; Tsubaki et al., 2015), stage (e.g. Shin et al., 2007; Royem et al., 2012; Gilmore et al., 2013; Schoener, 2017) or discharge (Lüthi et al., 2014) found its way into hydrological applications. Streamflow measurements using particle image velocimetry (PIV) require video imagery and have been used in combination with artificial and natural tracers (Bradley et al., 2002; Creutin et al., 2003; Muste et al., 2008). Implementation of PIV in applications for mobile devices enables the use as a mobile measurement device for streamflow (Tsubaki et al., 2015). Luethi et al. (2014) used PIV in a smartphone application in combination with derived water level and measured channel geometry to calculate discharge.

Besides information on streamflow velocity, image-based approaches can also be used to measure water level. Non-contact liquid level measurements using computer vision have been proposed by Chakravarthy et al. (2002) for fuel tank application. Shin et al. (2007) developed a stage measurement method based on images of a gauging staff in a stream. Their image processing is based on the assumption that the camera and staff gauge do not move, and therefore the staff gauge has an identical position in all images which can be defined by a region of interest (ROI). They were able to detect water level after image processing with a mean difference of 2 % between automated measured and visually measured pixels that indicate water level height. Royem et al. (2012) showed the value of wildlife cameras with time-lapse function as a potentially reliable tool to enable stream monitoring in citizen science. They combined the camera with a bright yellow steel ruler as the gauging staff. Evaluation of measured heights from the camera system with stage heights from the United States Geological Survey (USGS) showed a good agreement with a relative difference of 16 %. Gilmore et al. (2013) identified image resolution, lighting effects, perspective, lens distortion and the water meniscus as sources of error. They found image resolution and the meniscus contributing most to errors in detected water level, while the influence of lens distortion largely depends on the consideration of the distortion in the software. Image-based stage recording can provide backup information to existing gauging stations (SEBA-Hydrometrie, 2017). It can also be a practical and cheaper alternative to conventional stream gauging for the temporary monitoring of ephemeral and intermittent streams where costs of installing a conventional gauging system are not reasonable (Peters et al., 2012; Schoener, 2017). Despite these advances in measurement methods, designing monitoring networks for the dynamics of stream network extent remains challenging. According to Bhamjee et al. (2016) determining the adequate spacing of sites along a stream can be difficult as channel and flow characteristics can change over short distances. This can lead to misclassification of streamflow intermittency due to point-scale information and thus incomplete monitoring of linear features.

Setting up an affordable and dense monitoring system for a meso-scale catchment like the Attert requires the use of different sensors for streams with different intermittency characteristics. Thus, we present in this paper streamflow data obtained by a variety of measurement sources, including time-lapse imagery, electric conductivity (EC) measurements and conventional water level gauging. Time-lapse imagery was primarily dedicated to measure streamflow in ephemeral to intermittent streams. EC sensors on the other hand were used to investigate streamflow duration within the intermittent to perennial reaches of the stream network. Compared to conventional stream gauging the information on water levels is lost, but costs of installation and maintenance are reduced. The sensor network was completed by conventionally gauged weirs which allow for evaluation of the data obtained by the EC sensors and time-lapse cameras and allow one to relate the observed pattern of streamflow occurrence to the integrated discharge signal. In total streamflow presence or absence was monitored at 182 locations, allowing us to observe the contraction and expansion of the stream network in the Attert basin over time.

The Attert basin is located in Luxembourg and Belgium and covers an area of 247 km2 at the outlet at Useldange (Martínez-Carreras et al., 2012). With slate, marls and sandstone three prevalent geologies can be found in the catchment (Fig. 1). In the north-western part of the catchment (24 % of the area), Devonian slate forms the Luxembourg Ardennes, while the central part of the catchment is dominated by sandy Keuper marls and the southern catchment boundary is defined by the Jurassic Luxembourg Sandstone formation (Martínez-Carreras et al., 2012). Altitudes range from 245 m a.s.l. in Useldange to 549 m a.s.l. at the ridges of the Luxembourg Ardennes. Moderately sloping lowlands dominate the topography in the Keuper marls with steeper slopes at the hilly Luxembourg Sandstone formation (Martínez-Carreras et al., 2012). Surface soil patterns in the Attert basin are strongly related to lithology, land cover and land use (Cammeraat et al., 2018). Stagnosols to Planosols are prevailing soils on the marls; Leptosols, Arenosols and Podzols are predominantly found in the Luxembourg and Buntsandstein sandstone, while soils on Devonian slate are comprised mainly of Cambisols (Martínez-Carreras et al., 2012). Land use in the areas of marls comprises a mixture of forest (29 %), grassland (26 %) and agriculture (41 %) with low proportions of urban areas (4 %), while the sandstone areas are mainly forested (55 %) with lower proportions of grassland and agriculture (39 %). The topography of the slate-rich Ardennes is characterised by plateaus with altitudes between 450 and 550 m a.s.l. and steep valleys incised down to 300 m a.s.l. The plateaus in the Ardennes are mainly used for agriculture (42 %) and urban areas (4 %), whereas slopes and valleys are dominated by forest (48 %) and pasture (6 %). The climate in the area is influenced by the Atlantic Ocean and provides comparable conditions throughout the catchment. There is a slight spatial trend in mean annual precipitation, with annual precipitation averages decreasing from about 1000 mm in the higher altitudes in the north-western part of the catchment to 800 mm in the south-eastern part (Pfister et al., 2017). Mean annual precipitation across the catchment was about 850 mm for the years 1971–2000 (Pfister et al., 2005) and showed on average low variability in monthly precipitation sums ranging from 70 mm to 100 mm per month between the driest and wettest month of the year (Wrede et al., 2014). Mean monthly temperatures show strong seasonal fluctuations and reach a maximum of 17 ∘C in July and a minimum of 0 ∘C in January (Pfister et al., 2005). Consequently, monthly potential evapotranspiration follows the seasonal temperature changes with averages of around 13 mm per month in December up to around 80 mm per month in July (Wrede et al., 2014), adding up to yearly average sums of potential evapotranspiration of around 620 mm yr-1 across the catchment (Pfister et al., 2017). Therefore surface hydrology and the runoff regime is generally characterised by high, mostly rainfall driven flows during the winter season and low flows during the summer season as a result of higher evapotranspiration (Wrede et al., 2014). According to Pfister et al. (2017) bedrock geology has a strong influence on catchment storage, mixing and release of water in the Attert catchment, which results in stronger differences between seasonal flow regimes in areas of impermeable bedrock (slate and marls) compared to regions with permeable sandstone bedrock or diverse geology. Even the topographic map reveals the dependence of hydrological behaviour on bedrock permeability (Fig. 1), with a lower drainage density in the sandstone regions (Le Gouvernment du Grand-Duché de Luxembourg, 2009). Surface flow in the area is impacted by different anthropogenic influences. Tile drains, installation of dams and ditches, and other river regulation measures in the agricultural areas in the central marl-dominated regions of the catchment lower the groundwater table and increase runoff velocity (Schaich et al., 2011). Thus, anthropogenic impacts can alter the behaviour of ephemeral and intermittent streams, resulting in shorter periods of streamflow. Nevertheless, intermittent streams can also become potentially perennial if upstream water treatment plants provide constant flow of treated wastewater even in dry conditions, which can be the case at plants located on the plateaus of the Ardennes (Le Gouvernment du Grand-Duché de Luxembourg, 2018).

All applied methods successfully differentiated between presence and absence of streamflow or water. Sites were equipped for different periods depending on the method and purpose. Most of the time series (162 of 182) end in July 2017, with the longest time series resulting from the conventional gauges starting from either January 2013 (7 gauges) or September 2013 (11 gauges); EC data start between August and November 2015 with a total number of 93 sites. A total of 40 time-lapse cameras were installed in August to November 2015, and an additional 30 sites were equipped in June 2016.

For the overlapping period of August 2015 to July 2017 we measured the presence and absence of streamflow at 142 sites in the Attert basin (Fig. 7). A total of 129 of the sites show less than 20 % of no-data records during that time period (Fig. 5). The best data coverage is given for the period July 2016 to July 2017 with 182 equipped sites. Data series with less than 20 % of no-data records can be found in 155 of the time series for that time period (Fig. 5).

Although time-lapse cameras were primarily installed at ditches and channels with assumed ephemeral to intermittent streamflow, some of these sites showed continuous streamflow during the observation period. EC measurements show permanent flow for most of the sites with few exceptions at sites which were chosen as reference sites for intermittent stream observations with time-lapse cameras installed in nearby stream sections for validation. At most of the conventionally gauged streams from the dataset provided by the LIST we find perennial streamflow, whereas the sites equipped by the CAOS project show a higher degree of intermittency. Figure 6 shows the linkage between the proportions of bedrock geology found in the catchments of each monitoring site, the catchment area and the degree of intermittency within the period 1 July 2016–1 July 2017. Geology and catchment area are both known drivers of intermittency in the Attert catchment (Pfister et al., 2017), and therefore this additional information is included in the dataset. An overview of the complete dataset featuring data from a time-lapse camera (C), EC sensors (EC) and conventional gauges (CG) is provided in Fig. 7 and reveals fluctuation between wet and dry periods especially at sites monitored with time-lapse cameras.

Data for presence and absence of streamflow for all presented sites are freely available from the online repository of the German Research Center for Geosciences (GFZ) (Kaplan et al., 2019) and have the DOI 10.5880/FIDGEO.2019.010. For each monitoring system (time-lapse cameras, EC sensors, conventional gauges) one text file is available containing a time series starting from 1 January 2013 and ending on 17 July 2017 with a temporal resolution of 30 min. We provide the location of all sites in shapefile format in the projected coordinate system Luxembourg 1930 Gauss (EPSG-WKID: 2169). The shapefile includes the site ID, share of bedrock geology for each watershed, catchment area and mean catchment slope. Additionally shapefiles with catchment boundaries for each of the sites are available. Included is a readme file that contains a detailed description of data file contents, including header information and contact information for additional details. Additional data for the Attert catchment are published in articles of this special issue and are available in the same repository.

Although time-lapse imagery generally provides very accurate information on the presence or absence of streamflow compared to the EC sensor approach or even water level sensors, it has limitations when the quality of images is affected. Various sources can introduce noise, including direct impact of sunlight (Fig. 8a), scattering light reflections on the surface or gauging plate due to moving leaves, and moisture in the lens system (Fig. 8b). The view on the stream can be limited due to plant growth during the growing season (Fig. 8c), fog between the lens and channel, and spiders or insects in front of the lens. Gaps in the time series or additional data losses also happened due to vandalism (Fig. 8d) or theft of equipment. Additionally, data gaps were caused by unintended resetting of the internal camera software after issues with the power supply. This problem increased towards the end of the data acquisition period as cable connections loosened more frequently. Missing data in the time series are marked as NA values. Besides the issue of missing data, classification of streamflow into binary categories of presence or absence can lead to different results when classified by different persons. This issue was tackled by training of persons involved in the visual image analysis with three example sequences of transitions from flow absence to presence and back with images taken under regular light conditions as well as scatter light conditions and images showing a litter-filled streambed. Trained with these data samples and the according predefined classification, all involved image analysts got examples for all major conditions that were observed during the monitoring campaign which increased the homogeneity among the classification of different analysts. Finally, without motion picture one cannot distinguish between the actual presence of flow and pooling water in a wet channel. The additional information of water flow would be beneficial compared to the information of water being present when defining active channels as stated by Shaw (2016). In this study we assume all wet states of a channel to represent active flow conditions.

For the EC measurements to accurately describe streamflow presence or absence, sensor location is critical. Sensors not located at the lowest point of the stream can wrongly indicate a no-flow or no-water situation. With changes in streambed morphology following a high-flow event or anthropological disturbance, sensors might either end up in a dry stream segment with low flows bypassing the sensor or in a puddle within an otherwise dry streambed. Filtering of the EC data to represent flow/no-flow conditions was unproblematic due to the distinct and consistent differences in measured EC values, which can be used to accurately identify whether the sensor is completely dry, enclosed in wet but unsaturated sediment or submerged in water.

Data quality of the conventional gauges was generally high, but at gauges at CAOS sites, which were installed at smaller streams with a higher streamflow variability, issues with eroded weirs and altered channel cross section were reported. No problems were reported with data obtained from the gauges maintained by the LIST, with exceptions for data during low-flow conditions at freezing temperatures in the winter season 2017 which led to uninterpretable data during a short period, leading to data gaps. The gauge at the Schwebich river was under maintenance after the winter of 2017, and thus data from 2017 are missing.