Conventional luminescence readers rely on photomultiplier tubes (PMTs) to detect luminescence emissions e.g.. However, equivalent dose (De) distributions deduced from multiple-grain aliquots tend to scatter more than individual analytical uncertainties can explain (for a brief overview on various reasons, see ), and a PMT does not allow distinguishing simultaneously emitted signals from individual grains.

Single-grain systems, such as employed by , use a laser to optically stimulate luminescence OSL, single grains sequentially. Hence, luminescence is collected grain-wise. However, such a system does not suite when (1) the stimulation (heating, irradiation) can only be applied simultaneously to all grains, or (2) spatial mapping of the sample is desired e.g.. For these reasons, luminescence imaging systems have been subject of research since, at least, the 1980s cf.. In the 1990s, charged coupled device (CCD) cameras became affordable and gained attraction for luminescence detection due to their high quantum efficiency in conjunction with a relatively simple technical implementation into existing systems. A variety of experimental and commercial image systems based on CCD cameras were developed . However, the number of publications making use of those systems in the context of actual dating appears to be surprisingly small e.g..

Reasons for this lag of attention might be found in the technical complexity of luminescence imaging systems, combined with significant issues such as image noise or signal cross-talk . Thus, luminescence imaging methods appear challenging to apply, and the efforts necessary to analyse the measurements might be considered disproportional to the scientific gain. To prevent spatially resolved IR-RF from failing for the same reasons, we intend to provide our software and methods as accessible, transparent and automated as possible.

IR-RF dating applies ionising radiation to stimulate a fluorescence signal in K-feldspar at a wavelength of around 865 nm . This IR-RF signal decays with the accumulation of dose and resets through optical bleaching (e.g. a few hours to days of sunlight exposure). Determining ages up to ca. 600 ka is reported in the literature .

The early development of conventional non-spatially resolved IR-RF as dating technique was an effort of the group led by Matthias Krbetschek at the TU Freiberg (Germany) . Their work cumulated in the infrared radiofluorescence single aliquot regenerative-dose (IRSAR) protocol . Although the IRSAR protocol is straightforward and promises extended temporal range, its invention had limited impact on the dating practice in Quaternary science see for a detailed review.

One particular issue is the low bleachability of the IR-RF signal (at least 2 h of natural sunlight, ), which potentially provokes partial bleaching effects. Other issues are potential inhomogeneities in the mineral composition and micro-dosimetry of the sample .

Seeking a technical solution,

These results were never formally published. However, we are happy to share their presentation on the 11th International LED conference (2005, Cologne, Germany) upon request.

All measurements presented in this article were performed on a single Freiberg Instruments lexsyg research reader at the IRAMAT-CRP2A in Bordeaux (reader name L2). The system is equipped with a ring-shape type 90Sr /90Y β source delivering ca. 3.5 Gy min-1 to K-feldspar grains with a size of 125–250 µm cf..

For luminescence detection, we used a Princeton Instruments ProEM: 512B+ eXcelon EMCCD camera with a 512×512 pixel unichromatic back-illuminated CCD sensor. The camera sits on an automated detector changer, which allows also for spatially resolved thermal luminescence (TL) and OSL measurements cf.. The camera has a quantum efficiency (QE) of ⩾80 % between 450 and 750 nm. At the K-feldspar IR-RF emission centred at ca. 865 nm the QE is around 60 %. For the IR-RF measurements, we placed a Chroma D850/40x interference filter between β source and EMCCD camera. The custom-made optic has a numerical aperture (NA) of about 0.2 and a lateral magnification of 0.6, leading to an image resolution of 27 µm per pixel. Prior to our experiments, we adjusted the optical focus by manually calibrating the camera's installation height until we obtained the best image quality in terms of sharpness and minimised distortion. The nearby 90Sr /90Y β source shielding emits secondary X-ray photons bremsstrahlung, cf., which induce localised high signal events at the CCD chip upon impact. The result is images speckled with bright spots (see Fig. b). We will refer to this effect as “speckle noise”. It has to be noted that naturally occurring cosmic rays also cause similar bright spots. However, we approximated that cosmic rays are responsible for less than 1 % of the spots.

The system was equipped with a solar light simulator (SLS) system facilitating LEDs with broad peaks centred at 365, 462, 523, 590, 625 and 850 nm . The system is the same as used for the experiments by and .

However, over the years, the system received a couple of hardware upgrades tackling various problems cf.. In 2018, an improved drive train for the sample arm, modernised control hardware and a new 100 W at 48 V Si3N4 heater controlled by a PT1000 thermocouple were installed (personal communication, Freiberg Instruments GmbH, 2019).

Our software toolchain consisted of three different tools: LexStudio 2 for measurement sequence control, ImageJ for image processing and the R function library Luminescence for data analysis. ImageJ and the Luminescence package are open-source (GPL-3 licence) and freely available for all major platforms (Windows, Linux, macOS). However, our software toolchain was tested so far just on Windows 10 and macOS (v10–v11). Detailed installation guides and additional download links to the SR IR-RF-specific software modules can be found at https://luminescence.de/ (last access: 28 March 2021).

We applied the RF70 single aliquot protocol by , an improved version of the IRSAR protocol . The RF70 sequence (Table ) includes two IR-RF measurements: one for the natural signal (RFnat) and one for the regenerated signal (RFreg). In the data analysis process, the RFnat signal curve is slid vertically and horizontally along the signal curve until the best match is achieved with the RFreg curve. The horizontal sliding distance is the accumulated dose needed to match the natural RF signal, thus defining the equivalent dose . Measurement durations are user-defined. However, RFreg should be longer than the sample's expected natural dose. RFnat should not contain fewer than 70 data points (in our case images) to give sufficient statistical confidence when using the sliding method (, their supplement, proposed at least 40 channels for a resolution of 15 s/channel).

We used the comparable solar simulator settings as in

Contrary to what is quoted erroneously in , these settings are identical to the bleaching settings applied by for their measurements.

While the enhanced LexStudio 2 version automates image acquisition, it does not free the user from parameterising the camera. In the following, we will advise on the most relevant camera settings and their impact on image noise and signal sensitivity. We derive parts of our advice from signal-to-noise ratio (SNR) estimations summarised in Appendix . Table  lists major correlations between CCD camera settings and data quality. Table  lists the camera settings we used in our experiments. For more in-depth insights into the scientific CCD camera technology we may refer to and the Andor Learning Centre (https://andor.oxinst.com/learning/, last access: 28 March 2021; search for “Andor Academy”).

We obtain two image stacks (a series of images) per aliquot from the RF70 protocol. (Table ). Each image stack is saved as a *.tif file. Both image stacks are affected by speckle noise. Besides, the RFreg images might be displaced or rotated compared to the RFnat images due to uncertainties in the aliquot positioning .

Both issues and the grain identification are addressed by the SR-RF macro in ImageJ. The image processing has four steps (Fig. ): (1) speckle noise is removed, (2) both image stacks are geometrically aligned, (3) individual grains are identified, and (4) single-grain RF curves are extracted. Table  gives recommendations for the macro settings. For more details on the macro settings and the detailed sequence of ImageJ commands, we refer to our SR-RF macro documentation available at https://luminescence.de/ (last access: 28 March 2021).

We analysed the single-grain IR-RF data the same way that we would analyse conventional PMT IR-RF measurements. A simple R script to analyse the table.rf file of one aliquot reads as follows (R package Luminescence ≥ v0.9.8 needed):

Here, the new function read_RF2R() converts the table.rf file into a list of RLum.Analysis objects. Each RLum.Analysis object contains the RFnat and RFreg curves of one ROI. The equivalent dose of each ROI is calculated by analyse_IRSAR.RF(), which was already introduced and used by . The resulting dose distribution can be displayed and further evaluated by any of the various functions for dose statistics the Luminescence package provides. In the example above, we allowed vertical sliding after in the function analyse_IRSAR.RF() (parameter vslide_range). Vertical sliding can improve the equivalent dose results' accuracy but needs a significant curvature in the IR-RF decay to work properly. Vertical sliding can be deactivated by setting vslide_range = NULL or removing the parameter. The function plot_ROI() displays and returns the ROI locations and returns the Euclidean distance between them. This information is useful to study the impact of signal cross-talk.

An issue in OSL and TL imaging flagged by and further discussed by is signal cross-talk. Two independent effects cause signal cross-talk: (1) signal light misdirected by optical aberrations and (2) signal light backscattered by the silicone fixation layers and the sample carrier's surface. The visual perceptions are blurry luminescence images and signal halos around individual grains. These halos can extend into the ROIs of other grains located nearby. This cross-talk effect blends the IR-RF curves and potentially narrows the De distribution.

investigated the effects of signal-cross talk in spatially resolved OSL measurements. They found a substantial effect on the equivalent dose outcome and supposed optical aberrations as the primary signal cross-talk source. Like us, they used a lexsyg research device equipped with a ProEM512B camera. However, they measured at the OSL/TL sample position equipped with a custom-made multi-purpose optic . The OSL/TL optic was designed with large opening angles and high UV-to-NIR transmittance. This design decision enabled a maximum of signal yield for various applications but counteracted the optical correction of spherical and chromatic aberration and their secondary effects like astigmatism.

The optic of the RF position has a far smaller aperture (NARF≈0.2 vs. NAOSL/TL≈0.5) and therefore less spherical aberration. In addition, we consider chromatic aberration as negligible because we performed the focus calibration and all measurements at the same wavelength (865 nm). As Fig.  shows, the effect of signal cross-talk appears to be weaker in our measurements as observed by and should be insignificant for inter-grain distances above ca. 500 µm. We tried to maintain this distance by preparing our samples with a very low grain density.

Nevertheless, for samples with a high grain density or a high grain intensity inhomogeneity, signal cross-talk will be an issue. We propose the following countermeasures to reduce the effect; none of them applied in our experiments though:

Use special sample carriers (punched, black or polished) to minimise backscattered luminescence light.

We selected two potassium-bearing (K-feldspar) samples to apply and test our SR IR-RF tools and their settings. The first sample (TH0, grain size: 125–250 µm) is a modern analogue sample of aeolian origin from Sebkha Tah in Morocco cf.. It is the same sample used by to calibrate the 90Sr / 90Y source of the very same reader used for our measurements here. The sample was exposed to a γ dose of 56.02 Gy (cv∼2 %) in 2015 . The second K-feldspar sample (BDX16651, grain size: 100–200 µm) originates from a coastal dune in the Médoc area (south-western France). For this sample, estimated a palaeodose of 50.7±5.7 Gy for the quartz fraction (green OSL) and 96.2±8.0 Gy (with IR-RF) for the K-feldspar fraction measured here. For details on the sample preparation procedures, we refer to the cited literature.

TH0 allowed us to calibrate the 90Sr / 90Y source with SR IR-RF and compare the results with the PMT's calibration measurements. For this experiment, the detector changed in alternating turns; i.e. after measuring an aliquot with the PMT, another aliquot was measured using the EM-CCD camera, and then we measured an aliquot again with the PMT, and so on. We tested whether both measurements estimate statistically indistinguishable source dose rates.

The measurements of BDX16651 aimed at one main application of single-grain measurements: differentiation between grain fractions with different bleaching history. reported an age of 37.0±4.9 ka (arithmetic average ± standard deviation) for the feldspar fraction (measured with IR-RF) and 26.1±3.5 ka for the quartz fraction (measured with green OSL). While both ages overlap within 2σ, reported consistently older ages for the feldspar fraction compared to the quartz fraction for all samples from the site. Therefore, they argued that the natural bleaching was likely insufficient to reset the IR-RF signal of the feldspar grains. SR IR-RF should confirm the quartz result obtained by and potentially enable us to identify those grains that received a full signal resetting before the last burial.

For both samples, feldspar grains were dispersed randomly on stainless-steel cups aiming at a low grain density. The sample cups were sprayed with a thin layer of silicon oil. However, no mask or other aid was used because this reflects a more realistic aliquot preparation procedure in most laboratories. We aimed at 30 to 50 grains per aliquot. We prepared at least three cups per sample. Irradiation times were equal to values reported in and : 3600 s (RFnat) and 30 000 s (RFreg) for sample TH0; 3600 s and 10 000 s for sample BDX16651.

Figure  shows typical IR-RF curves from one ROI (in our case one grain) for TH0 (Fig. a) and BDX16651 (Fig. b). To obtain the Des, we applied the vertical and horizontal sliding technique to TH0. The vertical sliding ensures that both curves (RFnat and RFreg) match best based on their shape. This approach was first used by to corrected for changed signal intensities due to geometry issues. Still, it can also be used to correct sensitivity changes . For sample BDX16651, only horizontal sliding was used due to the absence of visible curvature in the IR-RF curve.

As rejection criteria, we applied the default test criteria (cf. , their supplement) of the function analyse_IRSAR.RF(). Two of those criteria were of relevance for our contribution: curves_ratio and curves_bounds. The first calculates the ratio of RFnat over RFreg in the range of RFnat. If it exceeds a certain threshold (here 1.001), it usually indicates that the RFnat best matched the RFnat while lying above the RFreg (additionally confirmed by visual inspection), violating the assumption that the highest IR-RF signal is observed for the RFreg after bleaching. If the second, curves_bounds, criterion is flagged, the RFnat cannot match the RFreg within the measured range of RFreg – an observation usually made for very noisy, flat curves.

The raw data of our measurements, along with the applied R scripts and partially pre-processed examples, are available open-access .

While we measured at least three cups with grains per sample, the number of usable cups presentable here finally narrowed down to one cup each. A malfunction in the cooling system of our camera stopped us from conducting more measurements. This cooling system degraded over the last 5 years, continually increasing the lowest reachable temperature from at least -70 ∘C in 2015 to about -45 ∘C in 2019. As we already mentioned above, the CCD chip temperature's spatial and temporal uniformity is necessary to ensure a stable and homogeneous signal background. We provide additional insights in Appendix  and Figs. S1 and S2 in the Supplement. Unfortunately, our camera's cooling system finally lost its ability to maintain stable chip temperatures during our measurements meant for publication. Significant variations in the IR-RF curves' signal background required us to discard most of our measurements (see Fig. S4 for an example).

Independently from this issue, we further discarded grains located at the rim of the stainless-steel cup, where our analysis indicated exceeded IR-RF curve boundaries for unknown reasons (i.e. no match between RFnat and RFreg).

Figure  illustrates the final results for the two remaining cups: one for TH0 (Fig. a, upper part) and one for BDX16651 (Fig. b, lower part). For each sample, we show an image taken with the camera (left-hand side) during the measurements and an Abanico plot of the distribution of the results. ROI pixels (diameter 7 px; see Table ) taken for the De analysis are coloured green and numbered. The numbers are displayed again in the Abanico plots (right-hand side). The results of TH0 display dose rates in Gy s-1 and equivalent doses in Gy for BDX16651. We applied the average dose model to both distributions with an assumed intrinsic overdispersion (σm) of 0.05.

It would be wishful thinking to assume that the work is finished. In comparison to PMT measurements, the number of control parameters exploded similarly to the amount of data. We tried to reduce the complexity by recommending meaningful settings and limit the number of adjustable parameters to a minimum. Still, other systems might have options we did not consider in our contribution.

Besides the technical problem we encountered with the detection chip's cooling system, we also acknowledge that our system is not perfect. A considerable improvement of image quality can be expected from a dedicated RF imaging system. We outline a possible design for such a system in Fig. . A system of one or two concave mirrors would allow the relocation of the camera and the filters further away from the β source and thus minimise bremsstrahlung effects and potential filter degradation cf.. Such a mirror optic would also eliminate chromatic aberration and thus enable the ability to take RF images at different wavelengths without refocusing. A dedicated RF optic would also address further optical aberrations and thus reduce signal cross-talk. As a camera, we propose a modern scientific CMOS camera. CMOS cameras have lower readout noise than traditional CCD cameras, although they do not support EM and hardware pixel binning.

Concerning the software, one subject for future improvements we did not implement is an advanced median filter in the image-processing macro. While the current algorithm proved itself powerful in sufficiently removing speckle noise, it decreases the time resolution and deletes those pixel values which are not identified as median values. More sophisticated algorithms deploy complex running median processes. For example, the (53H, twice) algorithm described in would mostly maintain time resolution while being still as potent in removing spikes. As another example, the (4253H, twice) algorithm of would maintain the shape of the underlying RF curve while smoothing away signal spikes and much of the Gaussian noise.

Another subject of potential improvement is the ROI assignment algorithm. The current algorithm assigns the maximum signal pixel of the grain as the centre of the ROI, no matter if this is the middle of the grain. We suggest a subsequent algorithm which refines the ROI centre towards an estimated grain centre. Consequently, the ROI size could be reduced without losing signal, usually leading to higher inter-aliquot scatter. This would increase grain separation and decrease any influence of signal cross-talk.

Finally, while the presented software toolchain is open-source, hence freely available and open to inspections and improvements, we acknowledge that the combination of three different software tools adds an additional layer of complexity. However, in particular, the image processing through ImageJ has the advantage that the data processing is transparent and available on all platforms and independent of the particular measurement system. Furthermore, users can tap into an extensive repository of available functions and plug-ins to record their own macros and thus adjust the image analysis with ImageJ without a need for programming skills.

This section alludes to the scientific gain and the initially expressed hypothesis that SR IR-RF can unravel the bleaching history of single feldspar grains.

We showed for sample TH0 that obtained source dose-rate results do not differ significantly from conventional IR-RF results using a PMT. This observation is reassuring because it shows that the presented workflow and analysis leads to meaningful results. However, Fig. a also reveals a large scatter between the individual feldspar grains ranging from 0.044 to 0.076 Gy s-1. reported a variation of the radiation field for our source type of only 2 %. Hence, the extreme values might result from microdosimetric effects irradiation, cf., which are related to IR-RF characteristics of single feldspar grains or varying K concentrations e.g..

reported zoning of feldspar grains linked to the geochemical composition. On some of our images (not shown), it appears that the light is not evenly distributed over the grain surface. However, higher optical resolutions would be required to investigate this aspect further.

Sample BDX16651 showed an even higher scatter in the equivalent doses, which is not surprising for a natural sediment sample. While environmental dose rate heterogeneities might add to the observed scatter (Fig. b), the internal K concentration of K-feldspar cf., in our case contributing ca. 23 % to the environmental dose rate cf., weakens the effect. Grain-to-grain variations in the internal K concentrations would undoubtedly broaden the De distribution. However, in our case, the K concentration was sufficiently constrained by energy dispersive X-ray analysis (EDX) their Fig. S16 at 11.4±2.5 %

This value was re-calculated for sample BDX16651 using the original data from but excluding K concentration values ≥14 %.