Fjallsjökull is an outlet glacier of Öræfajökull, an ice-covered volcano to the south of the wider Vatnajökull ice cap, in south-east Iceland (Fig. 1). Recession and thinning of Fjallsjökull has been underway since the end of the Little Ice Age, but has substantially accelerated in recent decades owing to climate warming (Howarth and Price, 1969; Chandler et al., 2020). Fjallsjökull terminates in a large (∼4 km2) proglacial lake – Fjallsárlón – which is also increasing in size as Fjallsjökull recedes (Schomacker, 2010). Calving of Fjallsjökull is regular and has increased in frequency in recent decades as the glacier has accelerated, driven by the expansion of Fjallsárlón (Dell et al., 2019). As of September 2021, the calving face of Fjallsjökull was approximately 3 km wide, with ∼2.4 km of this accessible from a boat (the northernmost 600 m had large, stationary icebergs which were dangerous to navigate; see Fig. 1). We selected Fjallsjökull as a study site due to its accessibility, ability to conduct surveys from boat and shoreline, and variation in calving margin heights (ranging from ∼1 to ∼15 m) to test the performance of our camera system under a diverse range of glaciological settings.

We tested the Raspberry Pi high quality camera module with a 16 mm telephoto lens in comparison to images taken from a DJI Mavic 2 Pro UAV. We also tested the Raspberry Pi camera module V2 (of lower resolution, but a cheaper option), due to its science-grade radiometric calibration (Pagnutti et al., 2017), but initial tests indicated the quality of the long-range imagery was too low to proceed with generating 3D data. The Raspberry Pi camera was attached to a Raspberry Pi 4B computer with an LCD display to visualize images, and adjust focus, as they were captured. Technical comparisons of the setups are given in Table 1, and a list of components and our code for acquisition is given as Supplementary Information.

The Raspberry Pi was mounted in a fixed position on a boat which traversed the southernmost ∼2.4 of the ∼3 km Fjallsjökull calving face, around 500 m from the glacier, while the UAV flew above this boat (Fig. 2). The Raspberry Pi was triggered manually approximately once every 10 seconds throughout the transect, capturing 315 images in total. While we operated the system manually herein, it is important to note, however, that the system is also designed to trigger autonomously at any frequency desired by the user. At the same time, the UAV conducted two flights, capturing 729 images, ensuring no calving occurred between collecting data from the two sensors. The UAV flew closer (∼250 m) to the glacier terminus than the boat transect to ensure the highest possible accuracy in data collection and to keep researchers at a safe distance to an active calving margin. This echoes similar approaches in studies of coastal landslides (Esposito et al., 2017) where a UAV flew closer than a boat survey to obtain the best possible quality 3D models for sensor comparison. In the majority of images, the UAV camera was facing the flat calving face of the glacier. While the UAV has onboard software to autofocus images, we manually checked and altered the focus of the Raspberry Pi camera between images during the boat transect to ensure pictures were not blurry as the boat varied in distance from the glacier.

In order to test the limits of the Raspberry Pi, we performed additional analysis on sub-section B (Fig. 1). We collected images of the calving face from a portion of the shoreline of Fjallsárlón, shown as X to Y in Fig. 1, which ranged from 1.2 to 1.5 km from the calving face. Owing to bad weather, we only collected shoreline data for a limited section (covering sub-section B entirely) before the glacier was obscured from view by fog. This experiment allowed us to assess how the Raspberry Pi performed at long-range.

We also conducted an additional experiment on sub-section B to determine the performance of the camera under sub-optimal conditions by removing 21 of the 31 images captured by the boat transect and deriving point clouds from the remaining 10 camera positions. This reflects the reality of the trade-off between data quality and practical considerations. In theory, fewer images should result in a lower point density (Micheletti et al., 2015), but any time-lapse camera array produced using Raspberry Pis could be cheaper with fewer cameras required.

For images from both the Raspberry Pi and UAV, far cliffs (rock faces flanking Fjallsjökull; Fig. 2a) were masked out prior to generating tie points in Agisoft Metashape. Images from the UAV were georeferenced using its onboard GNSS real-time kinematic positioning (RTK) system, with an accuracy <2 m (Nota et al., 2022). Images from the Raspberry Pi were georeferenced by aligning them to images captured by the UAV and producing a sparse point cloud, before removing UAV images to produce the final dense point clouds. Point clouds from both sensors were therefore referenced to this RTK system only, rather than having a global reference (akin to Luetzenburg et al., 2021). While the Raspberry Pi images could be successfully aligned without UAV images, our workflow was designed to unify the coordinate systems of the point clouds and thereby avoid confounding co-registration errors in the cloud comparison. Eight high quality point clouds were produced from each of the Raspberry Pi and UAV at various stages along the calving face (locations in Fig. 1) with a mild depth filter using Agisoft Metashape. Sub-sections were computed at natural break points in the glacier front geometry, at approximately 250–350 m intervals, owing to limitations in computer processing. We then cropped point clouds to the calving face, cleaned with a noise filter, and finely aligned the Raspberry Pi clouds to the UAV clouds assuming a 95 % overlap in CloudCompare.

Differences between point clouds from the Raspberry Pi and UAV were compared using the multiscale model to model cloud comparison (M3C2) tool in CloudCompare (Lague et al., 2013). M3C2 calculates a series of core points from the Raspberry Pi cloud and quantifies the distance to the UAV cloud about those points using projection cylinders. This requires users to define key parameters, including the width of normal (D), projection radius (d), and maximum depth of the cylinder (h) (all parameters in metres). We followed approaches developed by Lague et al. (2013), and applied to glacierized environments by Westoby et al. (2016) and Watson et al. (2017), of calculating the normal width to take into consideration surface roughness and the scale of the model. We used a standardized value of 0.6 m across all models as this fell within the range of 20–25x surface roughness for the vast majority (>98 %) of points, following equations presented in Lague et al. (2013). Projection diameter was calculated as a function of point density, so to ensure each projection cylinder had a minimum of five points, we used a value of 1.1 m. Finally, we set the maximum projection depth to 10 m to exclude grossly erroneous values (<0.01 % of all values).

The Raspberry Pi-based camera captured high-resolution imagery across the full length of Fjallsjökull, at distances of up to 1.5 km. Glacier textures and structures, such as debris patches and cracks in the ice, were clearly visible within the photos captured by the Raspberry Pi (see example imagery in Fig. 3) to aid 3D reconstruction. The ground sampling distance (GSD) (the on-ground distance represented by one pixel) of the Raspberry Pi at 500 m range was 3.80 cm and at 1.5 km was 11.41 cm (following calculations by O'Connor et al., 2017). By comparison, trail cameras used by Mallalieu et al. (2017), at a mean distance of 785 m to Russell Glacier, achieved GSD of 28.05 cm. We successfully generated point clouds along the front face of Fjallsjökull using the 315 Raspberry Pi photos captured from the boat survey. Eight point clouds were generated at high resolution, with survey lengths of ∼250–350 m each. The full range of calving face heights observed at Fjallsjökull, from ∼1 to ∼15 m were examined in this analysis. Point clouds were largely complete, though many were speckled in appearance.

Point clouds generated by the Raspberry Pi show a close comparison to those derived from the UAV, with a mean absolute error of M3C2 distance of 0.301 m and a standard deviation of 0.738 m across the Fjallsjökull calving face (Table 2, Fig. 4). Point density of all Raspberry Pi point clouds was high (<10 cm average spacing between points), allowing small features on the ice surface to be distinguished from ∼500 m away. Extremely high M3C2 values (a threshold greater than 1 m difference between the UAV and Raspberry Pi) are found at the far edges of the models where fewer frames are used to produce the point clouds, and at the highest parts of the margin (particularly prominent in panel E of Fig. 4). These values account for 5.03 % of points (3.31 % > 1 m; 1.72 % > -1 m), and there is a slight positive skew (the Raspberry Pi is overestimating the range to the glacier) in the error distribution with a mean M3C2 distance of 4.31 cm (Fig. 5). The difference in colouration between point clouds (demonstrated in Fig. 4) is likely due to the Raspberry Pi exposure, saturation, and ISO settings all remaining as “auto” to ensure good quality images across the transect. These settings are all fully adjustable if a camera is placed in a fixed position.

We analysed sub-section B (∼250 m long) under a number of other scenarios to explore the limits of Raspberry Pi cameras in SfM studies. Capturing images from the shoreline of Fjallsárlón, between 1.2 and 1.5 km away from the calving face (denoted by X and Y in Fig. 1), increased the standard deviation of M3C2 distance (0.742 m compared to 0.461 from the boat transect, a 61 % increase) and mean absolute error (0.341 m compared to 0.272 m from the boat transect, a 25 % increase). The point cloud itself was largely complete, though visibly more speckled than the point cloud generated from the closer survey (Fig. 6). We observed similar patterns of error in the point clouds captured from the shoreline as from the boat transect, with the highest errors corresponding to ridges of jagged ice.

Sub-section B was generated using 31 images from the Raspberry Pi in Fig. 4 and Table 2, but time-lapse camera arrays are generally limited to 10–15 cameras due to cost. We found that using a reduced set of 10 images had little impact on mean absolute error (0.263 m compared to 0.272 m using all images, a 3 % decrease), but increased the standard deviation (0.627 m compared to 0.461 m when using all images, a 36 % increase). This was most notable towards the periphery of the point cloud (Fig. 6G), though the point cloud contains more gaps than the original. Sub-section B is approximately 250 m long and an individual image captures ∼80 m of the glacier front, which means there was a low level of overlap (2–3 images at the right hand side, which is most speckled (Fig. 6). Given the good quality of images acquired at a greater distance, positioning cameras further away to create more overlap between images would likely address this speckle issue.

Raspberry Pi cameras have rarely been tested in a glaciological setting, but our analysis suggests that they could feasibly be deployed for long-term monitoring purposes and, given their comparable quality to a UAV-derived point cloud, have the potential to capture and quantify dynamic events (e.g. calving). Our data show that, from up to 1.5 km away, Raspberry Pi cameras can detect small features within the ice and, when used to generated 3D data, could identify, with confidence, any displacement of ice over ∼1 m in size. This also holds true for a camera setup using a much-reduced array; our experiments using just 10 camera positions yielded results that were largely comparable in quality to those comprising the full-suite of data (31 camera positions).

Improvements to research design, such as positioning cameras at a more optimal range of heights and angles, including above the glacier, are likely to reduce error in the Raspberry Pi point clouds (James and Robson, 2012; Bemis et al., 2014; Medrzycka et al., 2016; Holmes et al., 2021). A key limitation of our research was that images were captured only from a fixed height in the boat. Indeed, it is no coincidence that we observed the lowest errors between the two sensors at approximately the height level of the boat across all point clouds generated. We also speculate that systematic patterns of error, where high positive error neighbours high negative error such as in Fig. 6e, are due to varying angles of the glacier front being captured in the UAV model but not in the Raspberry Pi which only acquired front-facing images. Therefore, using a greater variety of camera angles and positions, for example by positioning cameras above the glacier front using nearby bedrock or moraines, would likely reduce error across the model (Mosbrucker et al., 2017; Medrzycka et al., 2016; Holmes et al., 2021). While our setup and analysis therefore may represent a conservative view of the potential use of Raspberry Pis in photogrammetry, it also reflects the practical considerations of working in field environments, which are frequently sub-optimal for deploying fixed cameras.

Our study used relative georeferencing methods, removing the need for absolute positioning of the clouds using surveyed ground control points. Over glacier calving margins, placing ground control points is especially challenging and alternate methods are required (Mallalieu et al., 2017). For example, there is precedent in using the geospatial data from one point cloud to reference another when comparing sensors (Zhang et al., 2019; Luetzenburg et al., 2021). Alternatively, the positions of the cameras can be used to determine the georeferencing. This “direct georeferencing” can be achieved using GNSS-based aerial triangulation of fixed positions, or an on-board GPS unit that shares the clock of the camera such that a precise time-stamp of location can be associated with each of the acquired images (Chudley et al., 2019). Using this approach would allow comparison between repeat point clouds captured by the Raspberry Pi without any alignment to a UAV-based point cloud. For broader photogrammetry applications of the Raspberry Pi, particularly involving setups with only one camera, control points may be essential in capturing the camera position accurately (Schwalbe and Maas, 2017).

In this study, we cropped our point clouds to show only the front, flat, calving face of Fjallsjökull. This involved significant trimming of point clouds generated by the UAV (up to 40 % of points removed), while the Raspberry Pi only required minor adjustments (∼10 % of points removed). A key limitation of the Raspberry Pi setup in our study design is that it cannot achieve the wide range of viewing angles and heights as a UAV does, and so analysis is limited to the front (i.e. vertical section) of the calving face. While this means the setup can monitor advance/retreat and calving events, the additional ability to generate a 3D model of the top of a glacier surface could potentially provide important information on calving dynamics, such as crevasse formation and propagation, which could be indicative of imminent calving (Benn et al., 2007). In previous work, monitoring the glacier surface in addition to the calving face has enabled the reconstruction of events leading up to major calving events, including the calculation of strain rate and identification of propagation prior to calving (Jouvet et al., 2017). Furthermore, other glacier characteristics, such as surface velocity, can indicate imminent calving but require a more top-down view of the glacier surface (Ryan et al., 2015). Modelling a greater extent of the glacier terminus could be particularly important if such a system was to be integrated into a GLOF early warning system.

For studies making use of a typical DSLR-grade handheld camera, James and Robson (2012) and Smith et al. (2016) suggest a typical relative precision ratio of 1:1000 – an error of 1 m when captured at a distance of 1000 m (though high-quality SfM often far exceeds this; James et al., 2017). At 500 m distance, we achieved a mean absolute precision of 1:1667 and at 1.2–1.5 km distance a mean of 1:978. These values almost match the precision thresholds set for DSLR-grade cameras and exceed the precision achieved by similarly priced trail cameras using in glacierized environments (Mallalieu et al., 2017). While terrestrial laser scanners can achieve greater levels of precision for monitoring glacier fronts (e.g. Pętlicki et al., 2015), their high weight and cost (tens of thousands of dollars) often precludes their use in glaciology research.

Glacier dynamics at a calving margin are complex, but a low-cost time-lapse camera array can offer insight into many key questions. Ice velocities at the terminus of Fjallsjökull range from ∼40 to ∼200 m a-1 for lake terminating ice (Dell et al., 2019). Glacier frontal positions and their diurnal variability can be monitored using this Raspberry Pi approach, as well as calving events that exceed 1 m in depth. Calving dynamics, including characterizing different types of calving and the impact of seasonality and lake drainage, can also be monitored from time-lapse cameras (Mallalieu et al., 2020) to aid in the understanding of how glacier calving contributes to the overall mass balance of a glacier and how this fluctuates over varying timescales (How et al., 2019; Bunce et al., 2021). Using time-lapse photogrammetry, it is theoretically possible to detect precursors (rotation, elevation change, creep) to calving events on the order of magnitude of >1 m, such as observed at Sermeq Kujalleq 65 h prior to calving (Xie et al., 2016).

In addition to calving events, terrestrial-based photogrammetry based on a Raspberry Pi system could monitor other important glacier dynamics at a low cost. There is a long history of using terrestrial photogrammetry for monitoring glacier thinning to quantify mass balance change of mountain glaciers, though this typically involves repeat site visits (Brecher and Thompson, 1993; Piermattei et al., 2015). Where surrounding topography allows, positioning Raspberry Pi cameras to look down on to the glacier surface would allow for SfM-based velocity calculation (Lewińska et al., 2021). Creep rates of rock glaciers have been successfully monitored through terrestrial photogrammetry (Kaufmann, 2012) and UAV surveys (Vivero and Lambiel, 2019), but again requiring repeated site visits. In each of these additional applications, low-cost Raspberry Pi cameras could produce accurate 3D models at a greater temporal frequency, without the logistical challenges, and financial costs, associated with repeating fieldwork.

Our boat-based study provides confidence that terrestrial-based, high cadence setups could produce regular, accurate 3D models. While not reported in these results, this author team have also successfully operated a separate Raspberry Pi camera in the Peruvian Andes, acquiring three images per day for 3 months using a timer switch and solar panel (Taylor, 2022). Given the customizability of Raspberry Pi cameras, their built-in connectivity, accuracy of acquiring 3D models, and robustness in cold environments, we are confident that arrays of fixed Raspberry Pi cameras could produce the first near real-time photogrammetry setup for continuous 3D monitoring of glacier calving fronts. Outside of photogrammetry, the programmability of Raspberry Pis as terrestrial cameras could offer advances in a broad range of settings, including GLOF management (Mulsow et al., 2015), supraglacial lake drainage (Danielson and Sharp, 2013), and iceberg tracking (Kienholz et al., 2019).

We speculate that, given likely sensor innovation and the decreasing cost of technology, the potential of low-cost sensors in glaciology research will only increase (Taylor et al., 2021). We envisage Raspberry Pi computers, or other microprocessors, to play a key role in this expansion. Almost all Raspberry Pi models have built-in WiFi, which allows data sharing between individual devices. With a WiFi radio on-site, providing a range of many hundreds of metres, individual cameras could autonomously send their data towards a central, more powerful Raspberry Pi unit for further analysis. Similar wireless sensor networks in glaciology have been produced to monitor seismicity (Anandakrishnan et al., 2022), ice surface temperatures (Singh et al., 2018), and subglacial hydrology (Prior-Jones et al., 2021). With the development of autonomous photogrammetry pipelines (Eastwood et al., 2019), a Raspberry Pi-based camera array system could, theoretically, run entirely independent of user input. Furthermore, the flexibility of Raspberry Pi computers, particularly their ability to operate multiple sensor types from one unit, opens up the possibility for wide sensor networks across glaciers – creating comprehensive digital monitoring of rapidly changing environments (Hart and Martinez, 2006; Taylor et al., 2021).

There exists considerable potential for low-cost sensors in mountain glacier communities, which are predominantly located in developing countries. Early warning systems situated around glacial lakes in the Himalaya have successfully prevented disaster during a number of GLOF events by allowing time for downstream communities to evacuate (Wang et al., 2022). By reducing the cost of camera-based sensors that are frequently used as part of a monitoring system (for example at Kyagar glacier in the Chinese Karakorum; Haemmig et al., 2014), more cameras can be situated to monitor calving rates, velocity, or stability at higher precision and accuracy in 3D. A low cost also means that more community-driven initiatives based on this Raspberry Pi system are viable. Such systems must be co-designed, and ultimately owned by, the communities they serve. Simple systems (such as Raspberry Pis), with components that are easily replaceable and with open access documentation, lowers the technical knowledge required to maintain an early warning system, and so a greater diversity of stakeholders can engage with its maintenance. Previous work has shown that diversity in engagement, and genuine understanding of the social structures on which communities are built, is essential for the success of early warning systems like these (Huggel et al., 2020).

While we suggest that Raspberry Pi cameras offer an alternative to expensive DSLR cameras for time-lapse camera arrays, based on our experiences we note a series of recommendations to future researchers and communities looking to use this approach in their own systems:

Camera setup must be carefully considered and adopt best practice set by others (e.g. Mallalieu et al., 2017) with regards to angle, overlap, and positioning.