Discharging groundwater at the study area is replenished in the Judean Mountains either through precipitation or flash floods that infiltrate into the Upper Cretaceous limestone and dolostone of the Judea Group Aquifer (JGA) and flow east towards the Dead Sea Rift (DSR). After passing the transition to the DSR, which is marked by normal faults and block tectonics, fresh groundwater enters the Quaternary fluvio-lacustrine Dead Sea Group (DSG) that is deposited in front of the Cretaceous rocks (Yechieli et al., 2010). The DSG consists of stratified fine-grained lacustrine sediments (clay minerals, aragonite, gypsum, halite), which are intercalated with coarser clastic layers. At wadi outlets, the lacustrine strata are displaced with fluviatile fine to coarse clastic sediments (Yechieli et al., 1995).

Due to the alternations of coarse and fine strata, groundwater flow occurs through separated subaquifers that possess different groundwater levels (Yechieli et al., 2010). In addition, preferential flow paths in the form of dissolution tubes and cavities develop due to dissolution of the evaporitic minerals (Magal et al., 2010; Ionescu et al., 2012). The dissolution process accelerates with the continuous and fast drop of the Dead Sea water level of ∼ 1 m yr-1 (Yechieli et al., 2010), which forces the formation of new groundwater flow paths. The resulting partially karstic flow system in the DSG is highly transient, resulting in immensely variable discharge rates, discharge locations and chemical composition of springs along the lakeshore (Burg et al., 2016). Besides admixing of interstitial brines to the groundwater, the degree of water–rock interaction in the DSG controls the groundwater's composition. Hence, large observable differences in the composition of discharging groundwater, which occur even within the range of a few metres, are an expression of the heterogeneous maturity of the karst system. The less mature the conduits, the larger the ratio between wetted soluble surfaces and volume.

Consequently, diffuse SGD exhibits a high load of dissolved solids and thus is saline (less mature karst system), while focused SGD is less loaded and exhibits a lower salinity (mature karst system).

At the investigation site, focused SGD occurrences were mapped in 2011 and 2012 by scuba divers and hydrochemically investigated. Based on their findings groundwater emerges from mature karst-like cavities down to depths of 30 m below the sea surface (Ionescu et al., 2012; Mallast et al., 2014).

From these cavities groundwater emerges as focused SGD. Subsequently, density differences between SGD water (1.00–1.19 g cm-3) and Dead Sea brine (1.234 g cm-3) trigger a continuous positive buoyancy (buoyant jet) of the emerging groundwater towards the lake's surface. Within the water column and alongside the buoyant jet, considerable turbulences develop, entraining ambient Dead Sea brine. Thus, the ascending water represents a mixture of fresh to brackish groundwater and lake brine. Once the ascending mixture reaches the lake's surface, it develops a radially orientated flow away from its jet centre, causing a circular-like pattern. These patterns are partially visually observable as shown in Mallast et al. (2014) at the Dead Sea and in various other cases (Swarzenski et al., 2001).

Apart from focused SGD, diffuse SGD occurs at the investigation site as well, either in various depths below the water surface (Ionescu et al., 2012) or at the shoreline. At the study area one diffuse SGD site exists directly at the shoreline (Fig. 1e). The discharge seems to occur over a length of approximately 20–25 m alongshore and was only detectable visually through the occurrence of schlieren in the lake brine. However, low discharge rates in combination with the immediate mixing with the lake brine impede any attempt to sample that water.

In order to still be able to compare groundwater characteristics, we sampled five onshore springs (Fig. 1d). All are in close proximity to previously mentioned SGD and to the shoreline, but emerge as focused flow in different elevations of 0.5–10 m above the Dead Sea water level. The water characteristics are shown in Table 2.

According to on-site measurements, SGD water temperatures at the orifices are 21–31.5 ∘C. During the time of investigation the Dead Sea had a skin temperature of ∼ 21 ∘C providing a temperature maximum difference of 10.5 ∘C between groundwater and ambient Dead Sea brine. Wind speeds amounted to 0.87 m s-1 (±0.16) approaching from SE to E (80–128∘). Occurring waves, which may influence SGD size and shape, had a frequency of 3–7 s with estimated wave heights <15 cm. During the flight, cloud-free conditions and thus homogeneous solar radiation existed, equally reflected at the sea surface throughout the entire covered area.

Thermal radiance image recording with 4–5 Hz results in a total of 670 images recorded within a time span of 167 s. Each image displays thermal radiances emitted by the surface. According to the Stefan–Boltzman law these radiances are directly proportional to the existing surface temperatures and thus are the basis for the present study. Yet, due to the UAV position variability while hovering above the pre-defined flight position, the mapped image footprint is not congruent for each image but varies spatially at the same magnitude as the position variability, which is also described in Holman et al. (2011).

To overcome the varying footprint, we define the first image of the set as the reference image and all remaining ones as input images (Fig. 2). Subsequently all input images are automatically co-registered onto the reference image using an intensity-based image registration within a MATLAB 2016 environment

The MATLAB code used for the present study can be distributed upon request.

The core of the intensity-based image registration consists of comparing an input matrix (reference image) with a transformation matrix (each input image). During an iterative regular step gradient descent optimization, the transformation matrix is transformed, incorporating scaling, rotation and translation. The intensities of both the input and transformation matrices are compared using a similarity measure (e.g. a mean square metric as in the presented study) with the aim to maximize the similarity between both matrices (Viana et al., 2015). The iterative optimization process continues until the maximum iteration criterion reaches 1000 iterations, or the optimization criterion reaches a maximum step length of 0.01. This image registration process is repeated for all input images.

To provide a further independent accuracy measure, we use the two previously described aluminium reflectors. Similar to the automatic approach described in Holman et al. (2017) to find GPS targets, we define search windows in the registered input images, looking for the lowest radiance values (= reflector plates). Since the plates represent an area of several connected pixels, we then extract the mass centre of both plates and each image. Comparing the positions of the mass centres of both reflectors in the reference image with each input image yields an independent spatial accuracy measure. The so-obtained spatial accuracy of the image registration results in an RMSE of 0.58 pixels (1 pixel = 13 cm), a mean of 0.5 pixels, and a standard deviation of 0.3 pixels (see Fig. S1 in the Supplement).

As a consequence of the image registration process and the transformation of the input matrices, the footprint (covered area) varies. In order to be able to analyse the same covered area, we reduce the image sizes to a common footprint extent represented by all images. The so-obtained image size amounts to 561 pixels × 376 pixels. Onto all co-registered images, we then apply a manually derived land mask, masking out any radiances from land parts. Subsequently, radiance values of the remaining water area are normalized using a z-score normalization to account for potential global solar radiation differences that may occur over the time of investigation and would affect the result. The so-obtained processed image set consists of a 3-D data cube (x,y,t) of 670 images (hereafter frames) resembling a total time period of 167 s and showing normalized sea surface radiances in x,y dimensions.

Since SGD at the investigated site consists of focused SGD occurring offshore and diffuse SGD occurring at the land–water interface, we use different approaches to extract relevant discharge spots separately to finally pursue the intended temporal investigations. Given the assumption of a thermally stabilized area over time induced by focused SGD (Mallast et al., 2013b; Siebert et al., 2014), we calculate the thermal variance per pixel of the data cube's temporal dimension t using MATLAB 2016. The resulting low-variance areas represent focused SGD sites with constant and intense discharge. To extract focused SGD sites, we apply a subjective variance threshold of <0.019

Note that the subjective variance threshold is optimal for the present case and will certainly change in other environments. Thus, in order to investigate any spatio-temporal SGD analysis, this threshold needs to be adapted otherwise it would create an erroneous focused SGD outline and hence a wrong focus area.

In contrast, the observed diffuse SGD site along the shoreline discharges less water with lower discharge rates and thus has a smaller momentum, which inevitably leads to a smaller, thermally stabilized alongshore area (∼ 101-102 cm perpendicular to the coastline). While thermally stabilized, several direct forces such as breaking waves and currents influence the same area and thus the resulting thermal radiation pattern on the sea surface. These factors lead to rather high variances compared to focused SGD flow. Unfortunately, a similar variance can be expected from ambient areas influenced by highly dynamic flow field induced by waves, currents and discharge. As a consequence, we delineate diffuse SGD from a single frame (frame 210 – not shown) in which thermal radiation patterns and maximum spatial extents induced by high discharge rates are unequivocally detectable (a comparable single image is shown in Fig. 3, upper left image). Analogous to the focused SGD sites, we apply a subjective threshold of >2.5 (normalized radiation) to extract discharge-induced thermal patterns and eliminate extraction artefacts using a morphological closing to clean extracted pattern objects, followed by the deletion of objects smaller than 150 pixels to focus on larger patterns only.

We conduct two forms of (spatio-)temporal analyses: (i) a spatio-temporal analysis to identify the spatial variability of both thermal radiance patterns induced by diffuse and focused SGD and (ii) a periodicity analysis to reveal possible reoccurring temporal discharge patterns. To explore the spatio-temporal behaviour, and within a MATLAB 2016 environment, we construct transects across the maximum extent of each extracted SGD spot, as we expect the most pristine patterns there. Along each transect, normalized thermal radiances per frame are extracted, filtered using a 1-D ninth-order median filter to reduce the white noise portion, and finally plotted, highlighting the spatio-temporal behaviour for each spot.

For the periodicity analysis we use an autocorrelation function, which measures the self-similarity of a signal (Tzanetakis and Cook, 2002). If discharge occurs regularly, it causes a periodic signal, which is expressed as a significant peak (above or below 95 % confidence interval) in the autocorrelation function. As we expect the most pristine discharge-induced thermal pattern signals for both SGD types at the midpoint of each transect, we pursue the periodicity analysis at the specific transect midpoint pixel for each of the transects.

To draw conclusions about the karst maturity of the flow net that feeds on- and offshore springs, we investigate the type and intensity of groundwater–rock interaction at each spring based on their on-site and chemical parameters. Physicochemical on-site parameters (temperature, density, pH, electrical conductivity) of all the above mentioned focused SGD as well as onshore springs (see Fig. 1d, e) are measured in the field using a WTW 350i and Mettler Toledo density meter. The sampling procedure for groundwater samples to analyse the major element concentrations follows the procedure described in detail in Ionescu et al. (2012). Generally, samples for anion and cation analyses are filtered (0.22 µm CA filters), separately filled in HDPE bottles and stored under cool conditions. Cation samples are immediately acidified and later analysed by applying ICP-AES. Anions are analysed using ion chromatography and bicarbonate by Gran titration (see Table 2).

The individual water–rock interactions, which lead to the chemical composition of the respective groundwaters in the springs, are inversely modelled applying PHREEQC and Pitzer thermodynamic database. We apply the latter due to the high activities in the modelled environment, consisting of a sedimentary saline aquifer body, soaked with interstitial brine, which admixes to the fresh groundwater. On the basis of the abundant easily soluble minerals (halite, aragonite and gypsum), we select reactive solid phases in the sedimentary succession, enable ion exchange on clay minerals, and evaluated modelling results based on the probability and lowest sum of residuals.

While the previous variance analysis highlights thermally stable and labile areas useful for identifying SGD spots, time- specific information on spatio-temporal discharge behaviour cannot be derived. We obtain this information through the introduction of transects (see left column in Figs. 5 and 6) along which we extract radiation values of each frame. The transects are constructed across the maximum spatial extent of each extracted, focused and diffuse SGD spot, as we expect the most pristine temporal patterns representative for each spot to occur here.

The previous spatio-temporal behaviour already pointed at a certain recurrence pattern of the observed thermal radiation but lacked a distinct statement on whether or not it contains a significant periodicity and thus a dominating force inducing it. In order to provide an unequivocally and temporally pristine discharge signal, we analyse its temporal pattern based on a single pixel of each transect (midpoint of the transect) using a temporal autocorrelation analysis (right columns in Figs. 5 and 6 respectively).

The samples focused on SGD and onshore springs discharge with a temperature between 21 to 31.5 ∘C. Though the groundwater of both focused SGD and onshore springs originates from the freshwater JGA, the springs discharge brackish water with salinities (TDS) ranging between 4.87 and 26.0 g L-1 with the tendency to be on average less saline onshore (TDS = 12.8 g L-1) compared to the focused SGD (TDS = 20.1 g L-1). The inverse geochemical modelling results indicate halite, aragonite and gypsum to be the most important minerals in solution, though ion exchange on clay minerals plays a significant role. Although discharge locations are very close, the amounts of dissolved halite (0–0.01 mol kg-1 H2O), aragonite (0–0.004 mol kg-1 H2O) and gypsum (0–0.02 mol kg-1 H2O) vary significantly between the different springs (Table 2). Translated into cavitation rates, emerging groundwater dissolves and relocates about 59.0–1.552.4 cm3 of halite, aragonite and gypsum per cubic metre from the passed branches of the groundwater flow net into the Dead Sea. Following the above-mentioned approach, those springs with the lowest water–rock interactions, which consequently emerge from the most mature karst pipes, are springs 09/857, 10/30, and 11/121, which all have values <79.8 cm3 of halite, aragonite and gypsum per cubic metre of water. In contrast, springs 09/855, 09/856 or 11/102 possess values >714.7 cm3 of halite, aragonite and gypsum per cubic metre water, which proves a higher water–rock interaction and thus intense dissolution activity that can only occur in less mature karst pipes (Table 2). Focused SGD spots reflect values of 696.5 and 749.7 cm3 (first and third spot respectively) and thus a less mature karst system, while the second focused SGD spot has the lowest value of 414.0 cm3 of halite, aragonite and gypsum per cubic metre of water and thus emerges from a more mature karst system.

Deriving unequivocal SGD indications from single frames such as in Fig. 3 is not trivial, especially in a highly dynamic system as the one presented. For the present case, we suggest the following causes to be relevant: i.

lateral flow dynamics induced by diffuse discharge with higher temperatures (see point ii) govern the investigated area and superimpose thermal radiance signals from vertical flow of focused SGD as mentioned in Mallast et al. (2013a);

Apart from enhancing focused SGD occurrences, the shape of the focused SGD variance pattern at the sea surface along with the location of the lowest variance values (area most thermally stable) gives an indication of SGD emergence locations and the deflection of the resulting vertical plume until it reaches the sea surface. None of the three are perfectly circular, which would indicate an uninfluenced positive buoyancy of discharging water and a SGD emergence directly beneath the centre of the variance pattern (Jirka, 2004). Instead, they are all more or less elliptical, with the lowest variance values at the southern ends (first and second focused SGD spot) and at the northern ends (third focused SGD spot). The remarkable elliptical shape of the first focussed SGD spot implies a crossflow from the south causing a northward deflection of the vertical SGD plume and an elliptic shape of the horizontal plume pattern at the sea surface (Akar and Jirka, 1995). Less pronounced elliptical shapes but with the same northward deflection trend exhibit the second and third focused SGD spot. The northward deflection is most likely induced by flow dynamics as a consequence of diffuse SGD. Since the location of the diffuse SGD spots, especially those with distinctively periodic events with higher discharge rates (Fig. 3), is directly SSW and shows the same northward horizontal plume orientation, we suggest this discharge is the driving force for the deflection.

The variance image provides an average representation of all SGD spots, which are especially useful for reliable size–discharge comparison purposes between SGD spots and likewise allow SGD spots to be outlined. However, as the previous section points out, all are subject to external forces such as currents, waves, and the influence of internal discharge dynamics on resulting pattern shape and size characteristics of the thermal radiance pattern over time.

For focused SGD, the observed thermal radiance pattern sizes (distance between boundaries) over time show a spatial variation between 116 % (=2.3 m for the first focused SGD spot) and 155 % (=4.0 m for the third focused SGD spot) as shown in Fig. 5. The variance is a result of occurring lateral flow dynamics constantly influencing the pattern on the sea surface. Yet, the influence is anisotropic in space and time as the lateral flow dynamics are dominated by waves coming from the east and the interaction of horizontal SGD plumes on the sea surface (e.g. second and third focused SGD) as described in Teamah and Khairat (2015), but moreover the strong lateral flow dynamics (crossflow) induced by the discharge impulses of diffuse SGD that in the following is deflected to the NE. The interplay and constant temporal changes lead to an asynchronous boundary movement for most of the observed SGD-induced thermal radiance patterns. Only during the first 200 frames and for the first and third focused SGD spot did the asynchronous behaviour change to synchronous behaviour. During this time only one force seems to dominate the dynamic, causing the synchronous behaviour.

The SGD-induced thermal radiation pattern size variation is different for the observed diffuse SGD spots. While three out of four spots constantly influence a longshore area of 5–8 pixels (0.65–1.04 m), outburst-like events change the cross-shore influence length between 150 % and 600 % and between 0.60 and 4.29 m (Fig. 6). The constant influence reflects a continuous diffuse discharge with lower discharge rates. The latter, however, shows a focused flow with intermittent higher discharge rates. Higher discharge rates induce a higher momentum and consequentially increase the influenced area off the discharge spot. In turn it reveals that karst conduits exist close to the shoreline and next to diffuse SGD. The intermittency with a seemingly recurring temporal pattern, however, points to a steady interplay of different forces that is the subject of the next section.

For focused SGD spots, we could not reveal a significant periodicity, either because of the limited observation length or because no periodicity exists. For diffuse SGD spots, the temporal autocorrelation analysis reveals significant periodicities. The periodicity of discharge rate events varies significantly among given spots between 20 and 78 s (right column in Fig. 6). It primarily provides a further example of the high temporal discharge variability over small spatial scales which, normally, is due to tides or wave set-up that change hydrostatic pressure conditions (Taniguchi et al., 2003b; Burnett et al., 2006). For the present case, tidal influences are irrelevant as the tidal cycles do not exist at the study site. Wave influence on the other hand cannot be excluded, per se. However, most likely it is not the main cause since observed wave frequency of 3 to 7 s would cause high-frequency discharge intermittency of the same magnitude. Precisely this high frequency is observable in the autocorrelation graphs of Fig. 6 close to the shoreline (transect pixel 3). Thus, the frequency proves the minor wave influence on the main discharge events with an observed frequency that is up to 10 magnitudes larger. Along with the focused discharge nature, it rather points to an interplay between wave set-up and a geometry effect within conduits of groundwater flow as the underlying mechanism as described for karst areas in Smart and Ford (1986). Discharge behaviour in this case depends on the maturity and geometric formation of the conduit network. Following Surić et al. (2015) the discharge behaviour is highly anisotropic and heterogeneous, and features a rapid flow. The anisotropic and heterogeneous discharge behaviour is furthermore underlined by the discharge onset and the periodicity that is unequal among the individual spots, even though their spatial location is within 10 m distance (Fig. 3).

According to the modelling results, groundwater passes the DSG through several subaquifers as described in Yechieli et al. (2010), and most probably via conduits, which develop through the fuzzy dissolution of easily soluble minerals that make up large percentages of the sedimentary body. It is further assumed, due to the impregnation of the sediment by Dead Sea brine, that cavitation activity is lower closer and below the freshwater–saltwater interface, although it exists and leads to abundant submarine springs. However, groundwater may also reach the Dead Sea through open faults, which may deeply fracture the sedimentary body as a result of active rift tectonics. Further preferential groundwater pathways may also be created through shallow cracks that develop through the relaxation of the sediment due to the gravitational release of interstitial brine.

However, for whatever reason fresh groundwater is allowed to invade the DSG due to the omnipresent abundance of easily soluble minerals, dissolution activities will immediately start and will enlarge hydraulic apertures of the initial pathways through the sedimentary body. Cavitation rates may be dependent on boundary conditions (e.g. supply of freshwater, hydraulic gradient and microbial activity), leading to different degrees of conduit maturity and random conduit–network geometry.

Consequently, it is thoroughly possible that an initial anisotropic flow system is about to develop close to the observed shoreline. In interaction with the wave set-up, we suggest that such a randomly developed initial flow system is the cause for the different onsets and influence areas for the observed outburst-like events.

The hovering of the UAV over a predefined location and the sensing of thermal radiation at a rate of 4–5 Hz allows a combination of the continuous temporal with the continuous spatial scale for SGD research. In this context it bears an enormous potential as it is possible to provide detailed and high-resolution information on SGD dynamics but also on external forces influencing it. The potential includes the high temporal resolution (sampling intervals), which differs by 1 order of magnitude from classical in situ measurement intervals of 101–102 min (Cable et al., 1997; Mulligan and Charette, 2006; Michael et al., 2011), allowing the illumination of short-term discharge dynamics that could not be reflected with classical methods. The potential furthermore concerns the spatio-temporal continuous characteristic of the presented approach. With the unequivocal indication regarding where diffuse or focused SGD occurs, and where exactly the transition between SGD and ambient fluids is, the indication is that proper sampling sites for each of them could not have been done with a subjective selection of sampling sites. Thus, applying the presented approach before pursuing in situ sampling (which includes the selection of proper sampling sites and sampling intervals) is undoubtedly advantageous.

The third potential concerns SGD monitoring and especially SGD quantification purposes. As mentioned in the Introduction, the basis for SGD quantification is the size of thermal radiance patterns (plumes) in most studies (Kelly et al., 2018; Mallast et al., 2013a; Tamborski et al., 2015; Lee et al., 2016). The presented results show a spatial variation of 150 %–600 %, which indicates the possible uncertainty that underlies a quantification based on single thermal infrared images. With the presented approach these uncertainties could be specified, which, in turn, increases the explanatory power of the quantification.

Apart from these potentials, the approach bears limitations and potential errors that need to be accounted for, if the presented approach is to be applicable to different locations. The first limitation concerns the need for rigid image parts, such as land, to be able to pursue a proper image registration. Equal share of land and water parts, as in the present case, increases the accuracy and thus reduces the potential error due to an erroneous intensity-based image registration, but reduces the investigable area spatially and limits it to areas close to the shoreline. The shoreline bond could be overcome by using rigidly fixed buoys and mounted aluminium plates on top as ground control points anchored offshore. While the investigable area is maximized, the image registration needs to be changed to a procedure based on control points to determine the image transformation and thus the registration as described, for example, in Holman et al. (2017), since the intensity of an image cannot be taken as the basis.

However, independent of the selected approach and the land–water share, flight altitude and camera lens define the size of the footprint and thus the spatial coverage. Due to regulatory framework, flight altitudes are usually restricted, which consequently limits the maximum possible footprint. Thus the restricted flight altitudes represent another spatial limitation of the approach.

The aforementioned spatial limitations are furthermore accompanied by temporal limitations and errors. These temporal limitations are given by the flight times of present-day UAVs that reach up to tens of minutes (Floreano and Wood, 2015). Continuous investigations for several hours, days, or beyond have been, to date, impossible. This sort of long-term and continuous investigation for monitoring purposes, for example, could be possible using a thermal camera system fixed to a mast making flight times irrelevant. Despite other factors coming into play with fixed cameras, such as the viewing angle dependency on emissivity (Norman and Becker, 1995) and the addition of a changing solar reflection component to thermal emission during the day vs. solely thermal emission during the night, the potential lies in the generation of a thermal radiation time series and trend analyses, for example. Similar approaches using fixed video and camera systems, operating in the visible spectrum (red–green–blue, RGB), are operational for near-shore monitoring and management purposes (Holman et al., 2003; Taborda and Silva, 2012). Adapting these operational approaches to fixed thermal camera systems would mean overcoming temporal limitations on the presented UAV approach and generating unforeseen potential in SGD research.

Further limitations concern sensors. A geometric error can be introduced by lens characteristics which distort the thermal image. In particular, a wide-angle lens produces geometric distortions (Meier et al., 2011; Vidas et al., 2012) that can be corrected in order to achieve an image projection that matches the true projection surface. A further sensor limitation is a possible radiometric error. All uncooled microbolometers, including the one applied, have the disadvantage that a thermal drift could occur (Mesas-Carrascosa et al., 2018). Caused by effects of the ambient temperature on the microbolometer detector housing and the consequential energy dissipation from the housing onto the detector array, the thermal drift leads to a non-uniform influence on the thermal image, which manifests in a vignetting effect with radiance reduction towards the borders of a recorded image relative to its projection centre (Meier et al., 2011). Since it additionally changes with time (Wolf et al., 2016), this drift, especially for long term investigations, needs to be accounted for, otherwise Mesas-Carrascosa et al. (2018) estimate the temperature error to increase by 0.7 ∘C min-1.

The aforementioned limitations are all of a technical nature. However, we need to emphasize that natural limitations may also exist that may affect the result. The most prominent factor is the temperature difference between groundwater and ambient water. With a difference approaching 0 ∘C, an unequivocal differentiation is almost impossible, especially if we include entrainment of ambient water and thus the temperature adaption. The higher the difference, the more likely the possibility of identifying SGD-induced thermal anomalies in single images, but also of using a time series of thermal radiance images – as in the presented approach. The advantage of time series data is that the temporal dimension includes dynamics which may enhance subtle temperature differences. These dynamics may be due to waves in which the surface geometries provide the direct indication rather than the surface temperature (recall the directional dependence of the surface emissivity on recorded radiances – see Sect. 5.1). Hence, time series data, whether from an UAV platform or from a mast, are thoroughly recommendable.

Further limitations may exist due to parallel existing strong lateral flow dynamics, as in the present case. On single thermal radiation images, these dynamics may camouflage further focused SGD sites, especially at sites with low groundwater–ambient water temperature differences. Strong lateral dynamics may also, as in the present case, camouflage any bathymetry effect on thermal radiation images as it is described in Xie et al. (2002). If bathymetry has affected the sea surface temperature, we could detect a gradual decrease in temperature from the shoreline towards the sea centre. The reason that we cannot detect any gradual decrease, apart from the camouflaging lateral flow dynamics, may be the bathymetry itself. While the bathymetry decreases gradually during the first 30 m until about 10 m, SGD is found at the bottom of steep walls in depths of up to 30 m (Ionescu et al., 2012) in distances of 50 m to the coastline, which is also visible in Fig. 1a. This sudden morphological step may additionally cause the disappearance of the gradual temperature decrease usually triggered by bathymetry. However, we cannot exclude this effect occurring in other places and different settings where, for example, the bathymetry consists of uniform slopes <5∘. In these occasions, the bathymetry would cause higher sea surface temperatures in summer and lower sea surface temperatures in winter that may be accounted for in the case of SGD detection.

As pointed out, for most technical limitations, solutions and corrections exist to improve and adopt the presented approach independent of the study sites' characteristics. Thus, we propose that the approach is applicable to other areas with diffuse or focused SGD, since the general requirements consist of an UAV with a mounted thermal camera system and rigid areas or fixed points within the covered footprint to allow a proper co-registration of all thermal images. The applicability of the presented approach concerning natural limitations needs to be investigated in the future. However, given a certain discharge rate and sufficient temperature differences between groundwater and ambient water, the suggestion is that time series data of thermal radiance images could prove to be a promising tool for SGD investigations.