A low-cost and open-source approach for supraglacial debris thickness mapping using UAV-based infrared thermography
Abstract. Debris-covered glaciers exist in many mountain ranges and play an important role in the regional water cycle. However, modelling the surface mass balance, runoff contribution and future evolution of debris-covered glaciers is fraught with uncertainty as accurate information on small-scale variations in debris thickness and sub-debris ice melt rates is only available for a few locations worldwide. Here we present a customised low-cost UAV for high-resolution thermal imaging of mountain glaciers and a complete open-source pipeline that facilitates the generation of accurate surface temperature and debris thickness maps from radiometric images. First, a thermal orthophoto is computed from individual radiometric UAV images using structure-from-motion and multi-view-stereo techniques. User-specific calibration and correction procedures can then be applied to the raw thermal orthophoto to account for atmospheric and environmental influences that affect the radiometric measurement. The corrected thermal orhthophoto reflects spatial variations in surface temperature across the surveyed debris-covered area. Finally, a high-resolution debris thickness map is derived from the corrected thermal orthophoto using in-situ measurements in conjuction with an empirical or inverse surface energy balance model that relates surface temperature to debris thickness. Our results from a small-scale experiment on the Kanderfirn in the Swiss Alps show that the surface temperature and thickness of a relatively thin debris layer (ca. 0–15 cm) can be mapped with high accuracy. On snow and ice surfaces, the mean deviation of the mapped surface temperature from the melting point (∼0 °C) was 0.4 ±1.0 °C. The root-mean-square error of the modelled debris thickness was 1.2 cm. Through the detailed mapping, typical small-scale debris features and debris thickness patterns become visible, which are not spatially resolved by the thermal infrared sensors of current-generation satellites. The presented approach paves the way for glacier-wide high-resolution debris thickness mapping and opens up new opportunities for more accurate monitoring and modelling of debris-covered glaciers.
Jérôme Messmer and Alexander R. Groos
Status: open (until 16 Jun 2023)
- RC1: 'Comment on tc-2023-41', Sam Herreid, 22 May 2023 reply
- RC2: 'Comment on tc-2023-41', Anonymous Referee #2, 10 Jun 2023 reply
Jérôme Messmer and Alexander R. Groos
A low-cost and open-source approach for supraglacial debris thickness mapping using UAV-based infrared thermography (v1.0.0) [Data set] https://doi.org/10.5281/zenodo.7692542
Jérôme Messmer and Alexander R. Groos
Viewed (geographical distribution)
The article “A low-cost and open-source approach for supraglacial debris thickness mapping using UAV-based infrared thermography” by Messmer and Groos provides an approach to bypass proprietary software to derive accurate surface temperature measurements and solve for debris thickness. The field experiment was thoughtfully set up and the article was a pleasure to read. It goes slightly against the title that the bulk of the analysis was conducted on data processed with proprietary software. This paper is an addition to a growing set of similar papers (mostly well cited within) which support the underlying premise of a coupled signal between surface temperature and debris thickness. This study, however, conveniently avoids the trickier aspects of this line of research by limiting the study site to a location where debris cover is generally thin, <15 cm, where the signal is expected to be strong. The spatial domain of these similar papers is also seemingly stuck to a small swath of a small glacier. Here, the authors put forth some good suggestions on how to break out of this limited domain in a UAV context, but they do not explore them. To further focus the paper, I think the discussion section could be shortened / organized with more subheaders. I also think it would improve the impact of the paper to add a specific discussion section on scaling the method and method repeatability. The in-line comments below are mostly minor.
L4: Perhaps less accurate and lower resolution, but it seems a little remiss to not mention Rounce et al., 2021 re: a global map of debris cover.
L5: Maybe it’s “describe” the customization of a low cost UAV and “present” a complete open source..
L8: Is it still raw though? I assume during the SfM step there was some statistical resampling. I’m fine if you still call it raw, but to me that is strictly the values of the individual radiometric images.
L9: provided the dt spanning the set of images is reasonably low. I’m sure you say this in the text.
L13: using which method, empirical or inverse EB? Or both?
L15: is this still for a 0-15 cm class? Or for the whole DC area?
L17: “paves the way” Does it? I’m not sure when we’ll have regional, not to mention glacier-wide, UAV coverage and moving up to fixed wing or helicopter increases the cost by quite a bit.
L44: Rounce et al., 2023 seems to temper this urgency, at least in terms of global sea level predictions. “The inclusion of debris thus delays mass loss over the century especially at local scales but has little impact on sea level rise and the number of glaciers lost by 2100.” Regionally, and for addressing problems on shorter timescales, the urgency seems to remain warranted.
L53: Consider saying more about the low accuracy since this is key motivation for finescale work such as this. You can more explicitly constrain “low accuracy” with repeatability arguments and the often stated ~0.5 m detection limit. From Rounce et al. 2021, we now know where we can anticipate this limit to be met.
L55: Higher resolution mapping techniques are required to do what? Brute force solve the problem with e.g. aerial surveys, or inform satellite based methodology either with sensors in orbit today or wait for civilian satellite surface temperature data to become higher resolution?
L59: See also:
Gök, Deniz Tobias, Dirk Scherler, and Leif Stefan Anderson. "High-resolution debris-cover mapping using UAV-derived thermal imagery: limits and opportunities." The Cryosphere 17.3 (2023): 1165-1184.
Aubry-Wake, Caroline, et al. "Using ground-based thermal imagery to estimate debris thickness over glacial ice: fieldwork considerations to improve the effectiveness." Journal of Glaciology 69.274 (2023): 353-369.
L61: The nadir camera angle is without a doubt a benefit, and a UAV can capture a wider swath of a glacier. But still, the narrow swaths of high resolution thermal data presented here and in all other similar studies I have seen, the spatial coverage is not operative for studies beyond proof of concept. Further, there are radiometric resolution trade-offs between a heavy ground based camera and those that meet the payload of a UAV. I’m not making an argument for more ground based studies, I’m just saying proof of concept studies might benefit from the better (heavier) sensors and the spatial gains from today’s UAVs are not orders of magnitude better.
L66 Or hinting at its futility? For a method to be robust, the cannon of similar papers will need to all start converging on good news and repeatability or the general concept should probably be sidelined.
L69: I assume it would work on thermal imagery acquired by any means?
L73: debris [surface] temperature.
L87: The debris-covered area of the Kanderfirn is already very small, on a bigger debris covered glacier, say, the terminus of Bering Glacier in Alaska as an extreme, how would you decide where to survey? Is there any path for scalability with this method?
Data and Methods
L117: Recommended in the literature or from this study?
Table 1: The smaller TIR area surveyed from the same flight must mean that the overlap was less than for the visible imagery? This difference is due to the view angle? Is the 75% from L118 for RBG or TIR?
L139: Literally random, as in picked by an algorithm before going into the field, or just random sampling while you were in the field?
L148: I’m not sure if it was the correct decision to put the temperature loggers under a rock to shield it from direct radiation since the thermal camera will be measuring the skin temperature and thus include that very warm signal. It’s not unusual to measure 30 degree debris surface temperature on a cool but sunny summer day (as you know from your Fig. 10). It helps a little bit that it was slightly overcast (L113), Herreid, 2021 found overcast conditions to be preferable for TIR / debris thickness / sub-debris melt measurements. Studies like this one that look at a debris cover internal temperature profile along with contact thermistor and TIR surface temperature measurements show smooth and predictable diurnal signals just below the surface and lower (below your shale stone) and more chaotic signals for both TIR and contact thermistors at the very surface (chaotic as a function of clouds passing). Both have the scientific uses, I think in your case the skin measurement is correct, but what you collected should still be close.
L150: If you’re comparing a ground point measurement to a single thermal pixel at this location and the thermal pixel has a resolution of centimeters, it might be good to take a mean surface temp around the location or move the ribbon off a set distance since it will have a different emissivity to the surrounding rock.
L151: How long did you leave the sensors running before the survey? At the surface it should be pretty quick, but the sensors need to become isothermal with their environment.
L159: A few km in the mountains can make a lot of difference, e.g. RH in Table 2. Your final calculations might not be so different with a local sensor, but I think it would be following best practices for a small scale proof of concept study to at least collect local air temperature and RH measurements.
FIg 4. Should “Ice-snow mask” also include debris (L170)? Was this a manual or automated step?
Fig 4. It’s a little confusing if the SfM-MVS steps are open-source or not? L200-214 makes it sound like you used proprietary software but found open source alternatives but didn’t find them suitable for your study? I can completely understand how these decisions happen and end up a bit confusing in text form. However, if there is a simple notebook workflow from start to finish, as Fig. 4 suggests, it would be most clear if this was the workflow used, and then you could have a stand alone validation section with comparisons to common proprietary software packages.
L198: I don’t actually see this step in Fig 4. I guess you are referencing only the open-source pipeline but it reads like the validation step is there.
L198: Already stated in L172
L240: It’s not exactly clear to me where the coefficients came from, from the FLIR algorithm or from a different study or from a post processing step, in which case there should be many for each raw image or only one set from the orthoimage?
Equation 2, 3: can you give the citation again for these? Is it all Tattersall (2021a)?
Equation 4: I don’t understand why you take the square of the correction factor?
L255: I’m not so sure if the GCPs meet the criteria, the motivation for the crinkle is to capture surfaces normal to all different angles, in this case I think you’re getting almost exclusively a signal from the upper atmosphere or overhead clouds which will be quite a bit colder than reflective temperature from, say, a valley wall or nearby moraine. Especially if only extracting a central point. I think this is also somewhat of a localized measurement problem, e.g. if someone is having a bonfire just off frame, that heat will “bleed” into neighboring pixels. Do you know what, at your scale, the source of this correction is accounting for?
Table 3: Why do you differentiate between snow and ice if you apply the same emissivity?
L282: Was there snow in your study area? Last year’s or fresh?
L302-308: I think it’s fine to have the context here but could move to results or discussion.
L310: “can be”, but within some well known limitations. Time of day, e.g. if all surfaces go isothermal at night, or signal decoupling from thick debris.
L316: Add citation to statement.
Figure 5: Add n= to each class; maybe stylistic, but I would continue the curve through the origin since 0 h_d will theoretically pass 0C.
L327: Probably best stated with results, putting them into context.
L355: Thermal diffusivity is at least straightforward to solve for from field data and common in debris cover research. I thought the idea behind your page worth of equations is that there isn’t a tuning parameter? Maybe say here that you limited k to realistic values. Unless I’m misreading the purpose of Table 5, add the range/set of values of k rather than just one. Is 1 the best performing that you tried? What did Evatt et al., 2015 use? What metric did you optimize on?
L360: Debris thickness measurements are always a little subjective depending on where you measure. Somewhere, maybe study area section, could you say some qualitative observations of your measurements e.g. in debris less than 1 cm there were fines that still covered the surface or single clasts at the measured debris thickness and bare ice visible between clasts. To those familiar with these field measurements we’ll know if this has the potential to have a higher melt rate that bare ice (in the case of the fines) or about the same as bare ice (if it’s just stones doing a small scale version of the ice pedestal process).
L360: Maybe “in diameter”, of course it’s clear as it is, but sounds a little funny.
L362: [were] not measured or mapped.
Figure 6: I think “modeled” not “mapped” on x-axis
Figure 6: The problem with Fig. 6 is that we look to measured for what the distribution should be, but actually the other two are probably more “true” because they are more inclusive. I think you need to add two more: emp and phy with only intersecting points to measured.
Figure 7: Same comment as above but compounded by being fundamentally different quantities. Still nice that the shape is similar, but are any of your readers skeptical about the physics behind a thermal image? I’m not 100% sure I see the point of this figure. I think there is no good reason to leave the GCP points in the fig.
Figure 8: This figure seems mostly unnecessary and could be merged with Fig. 1. Maybe others feel differently, but I don’t get any new information from seeing a raw DSM, the glacier outline is a pretty unrecognizable shape, and I think most specialists reading the paper will take you at your word about high resolution data resolving boulders, debris and ice cliffs/crevasses.
L376-385: Move to Study Area section
L384: give fig ref with labeled “M”
L384: I have never heard of supraglacial moraines having a lower tail. I think what you mean is a medial moraine structure is crosscut (maybe add: and dispersed) by surface water flowing orthogonal to ice flow. Semantically, all of the rock in your study area is supraglacial moraine.
L389-390: I don’t recall this being a research question raised in the introduction, and I’m not sure if it needs much attention. I would say a network of field based Ts measurements to evaluate the signal of a medium resolution satellite thermal pixel addresses a scientific need, but here I think I trust your drone images to measure Ts more than contact thermistors below the surface. I would flip the axes of Fig 9 to say you can use T below the rock to approximate Ts :)
L399: “More interesting” but in a supplemental figure :) (I’m not saying change it)
L403: It’s important and good that you used both proprietary and open source software, but slightly concerning that you preference the proprietary while open source is a main contribution of this work.
Figure 10: I thought emissivity was segmented for the different surfaces?
Figure 10: Is there a mechanism for sub freezing temperatures or is that a processing error?
L404: Notably the snow patch is lower than 0C and what looks like depressions near the bare ice. Any idea if this is signal or error? Maybe places to suggest putting T ground control for future studies?
L405: I think more technically it is atmospheric temperature not surface temperature. The actual surface temperature of the aluminum is near isothermal with the ground.
L408: What is reporting the average temperature of a random glacier patch for a random time good for? What is a reader learning?
Figure 11: Do you think it's a coincidence that these very cold areas are on the edges of the spatial domain? Maybe a surface resampling residual? Although that seems unlikely with such dense data. Camera angle / vignetting? I think that might be a problem with smaller thermal cameras vs bigger handheld ones. Can you look at the actual raw thermal tiles and see if they also have that cold signal? What’s tripping me up is the surface stream that seems to go cold warm cold, but there is some cross cutting. Could you do an oblique 3D of this so we can see the gradients? The forcings on the ground should be pretty straight forward, 0C water and protection from incoming solar. The change in color of the rock in Fig. 8 at the bottom makes me think it could be a rock wetness / emissivity difference, not so clear about higher/northern cold spots.
L419: That’s fine to mention the camera detection limit, but there’s a difference between random error and a systematic shift. For your empirical relation this is not really relevant but I think the absolute Ts is important for the EB model.
L338: Earlier it was not called a sensitivity experiment, and here it’s unclear half way through the sentence. There’s a difference between a sensitivity analysis and model calibration.
L338: Consider subsections between empirical and EB results, it’s not clear going straight into k that this is a build for EB results.
L440: Point to where in FIg 13 we learn the “very sensitive”
L440: What does “spread” mean? The RMSE?
L451: Nice result. Maybe also speak to repeatability / calibration here if there isn’t a more in depth section in the discussion.
L454: I’m not recalling any explicit mention of the cost?
L459: Trading for an easily acquired time-series.
L469: I wouldn’t say “defacto infeasible”, just a problem needing an innovative solution.
L475: Thermal equilibrium with what? Do you mean a single temperature throughout? Your method itself depends on a thermal gradient within the debris layer.
L530: Why unclear? Within constrained confidence limits, the measurement is not much different from any other that is generally collected without redundancy.
L539-548: This paragraph seems to be repeating established ideas and results, consider cutting.
L548: This study is limited to debris thicknesses that are known to be detectable with thermal imagery. I think the key questions for a promising debris thickness mapping technique are centered around thicker debris, so this statement seems too strong to me given the scope of the study. As Ts climbs the upper limb of the Eq in Fig 5, small variations in Ts produce bigger errors in debris thickness. This study, at least the empirical portion, is conveniently distant from the approaching asymptotic behavior.
L555: Why, because it has a slightly more similar shape than the wider emp? Per my earlier comment, I don’t think these plots are directly comparable.
L591: To be fair, Kanderfirn is a tiny glacier with almost no debris cover. That the global dataset caught it at all seems like a very positive review.
L595-597: Now talking about sub-debris melt seems a little out of place from the bulk discussion of the paper.
L597: I think your study domain would cover 6 ASTER thermal pixels. It’s not the most encouraging advice to say this study would have to be repeated maybe 10 times to cover a more traditionally extensive debris cover and evaluate satellite data.
L603: I don’t really see any “[paving] the way for glacier-wide high-resolution debris thickness mapping” in this study. The following sentence provides some ideas to overcome the 10 minute drone flight limit, but these are not explored here. The discussion is dismissive of using reanalysis data (Gok et al., 2022) but doesn’t put forth an alternative scaling method other than deploying more met stations on glaciers.
L613: I’m not sure I agree this is the most important next step, these thermal data seem quite good and fairly well constrained. I think effort now needs to be focused on thicker debris, repeatability, and wider coverage.
L615: Can you state some statistical measures of performance for the two methods here and in the abstract? It would be useful for the reader to see right away how different the end results are of the two methods.
Rounce, David R., et al. "Distributed global debris thickness estimates reveal debris significantly impacts glacier mass balance." Geophysical Research Letters 48.8 (2021): e2020GL091311.
Rounce, David R., et al. "Global glacier change in the 21st century: Every increase in temperature matters." Science 379.6627 (2023): 78-83.