Fires in tropical peatlands are a well-recognised source of carbon emissions to the atmosphere [1
], and in 2015 levels of landscape burning in Indonesia were greatly exacerbated by a drought brought on by the strong El Niño. Estimates suggest that over 0.8 × 106
ha of Indonesian peatlands burned in September and October 2015, releasing 11.3 Tg of carbon to the atmosphere per day, equating to more CO2
than the entire European Union emitted in the same period [3
After most of the above-ground biomass has burned, peat fires continue smouldering, spreading deep into the peat layer and burning away the carbon-rich soil [4
]. Fire propagation is via pyrolysis (absorption of heat and release of gas, producing char) and oxidation (organic matter/char is consumed producing ash and gases in the presence of oxygen). The majority of emissions occur through the oxidation of char, rather than the pyrolysis of peat [5
]. Char oxidation is evident from the ash that remains post-fire (Figure 1
), and also the presence of char beneath. Complete combustion (and so the greatest Depth of Burn, DoB) is suggested to occur mostly around the roots of trees [6
], potentially because of reduced soil moisture [7
To estimate carbon and greenhouse gas (GHG) emissions from tropical peatland fires requires quantification of the amount of peat consumed. This can be done by assessing the amount of thermal radiant energy emitted by the burning peat—sometimes coupled with atmospheric assessments of the emitted species [3
]—but is most commonly done by measuring the burned area and multiplying this by the estimated depth of burn (DoB) and pre-fire peat bulk density [10
]. Due to the high bulk density and carbon content of peat, and the fact that DoB can extend to tens of cm, landscape fire fuel consumption per unit area in tropical peatlands is amongst the highest of any biome worldwide. However, the range in DoBs expected between different types of peatland biome and between different fires means that the carbon emission estimates derived from the burned area based approach remain rather uncertain, and whilst burned area and peat bulk density are increasingly well measured [12
], few DoB estimates exist currently. The existing spatially explicit DoB measurements available, beyond a few point-based measures derived via the use of metal rods to assess subsidence after burning [1
], have mostly been assessed using light detection and ranging (LiDAR) approaches [9
LiDAR can measure surface topography (to within a few cm) by timing the return of an emitted laser pulse, and airborne LiDAR enables difficult-to-access landscapes to be surveyed quickly and accurately [16
]. However, the costs of LiDAR surveys are prohibitive. Furthermore, DoB is most accurately measured by differencing pre- and post-burn digital terrain models (DTMs), but collection of the former data requires predicting where (and to some extent when) landscape fires will occur. To date, only one study has estimated peatland depth of burn by differencing the pre- and post- burn peat DTMs of the same area of peat swamp forest (in the USA), and returned a mean DoB of 0.47 m (with a standard deviation in this mean of ±0.18 m) over a 25 km2
burned area [16
]. Other airborne LiDAR-based studies have not had pre-burn elevation data available, and so have either measured the mean elevation difference between a burned and adjacent unburned area [9
], or have reconstructed a pre-burn surface model using interpolations from adjacent unburned topography [15
]. While these two post-burn only approaches may not capture DoB as well as if pre-burn topographic data were available, they have provided unique assessments of peatland DoB over very large areas (3750 and 4000 ha respectively).
Most in situ peatland fire research has focused on temperate or boreal regions [6
], but as tropical peat is derived from different vegetation types and contains a much higher fraction of woody material [21
], more work is required to assess DoB in tropical peatlands. Work in [4
] demonstrates the thermal and physical burn characteristics of tropical peat fires using experimental plots in Kalimantan (Indonesia), albeit at very small-scale (9 m2
area) where point-based measurements can suffice. Unmanned aerial vehicles (UAVs) provide a flexible platform to cover much larger areas, and via photogrammetric analysis, the imagery captured by such platforms may also be suitable for mapping DoB. However, to our knowledge, this has yet to be attempted in any biome worldwide.
Using overlapping, high resolution aerial photographs taken from different viewing angles, Structure from Motion techniques (SfM) reconstruct point clouds (similar to those provided by LiDAR methods), which can then be used to produce digital terrain models (DTMs) and orthomosaic aerial images (orthophotos). The accuracy of DTMs from SfM techniques is affected by the amount of image overlap, camera calibration and sensor size, image quality [22
], photogrammetry algorithm [23
], UAV flying altitude [24
], camera viewing angle [22
], ground control point accuracy, and ground cover heterogeneity (i.e., featureless surfaces can cause additional error [22
]). Fortunately, modern photogrammetric software can automatically detect tie points between photos, regardless of changes in camera orientation or survey distance, providing there is sufficient overlap between photos (50%–90%). DTM vertical accuracy and precision of UAV surveys is reportedly comparable to that of airborne LiDAR, with [28
] achieving a mean 0.015 m bias (root-mean-squared deviation (RMSD) = 0.220 m) in a coastal setting by comparing UAV- and Terrestrial Laser Scanner (TLS)-derived DTMs, [30
] achieving between 0.004 and 0.04 m accuracy and 0.02 and 0.07 m precision for a river area and [31
] generating DTMs accurate to 0.025–0.04 m in a dune environment.
Here, we assess the applicability and accuracy of a low-cost UAV system for assessing peat fire DoB for the first time. Our focus is a previously unburned but degraded peat swamp forest, along with a previously burned shrub land, both located on Sumatra, Indonesia. Pre- and post-fire topography from LiDAR- and UAV-derived DTMs are compared to assess peat depth of burn (DoB) across the area, whilst the aerial photography is also used to assess combustion heterogeneity.
We have presented a structure from motion (SfM) approach for spatially mapping peatland depth of burn (DoB), based around post-fire imagery collected via a small, affordable UAV. To our knowledge, this is the first time such an approach to digtial terrain model (DTM) generation has been applied to assess DoB. We confirmed the versatility, accuracy and optimium setup for the approach; with suitable imagery for deriving surface DTMs collected at a variety of flying altitudes (up to 70 m) and camera angles. DTMs showed accuracies better than 5 cm compared to ground survey measures, comparable to that of airborne LiDAR (e.g., 7–15 cm for [6
]). By comparing pre- and post-burn DTMs of the same 5.2 ha area, peatland DoB was mapped across a degraded peatland (not previously burned) and mean DoB assessed as 0.23 m with <1% uncertainty. The standard deviation of the DoB measures is 0.19 m, indicating wide DoB variations, and maximum DoB extended beyond 1 m in some locations. This range is similar to estimates for forests that have only burned once [15
], but less than the prior estimate provided by [9
] (0.33 ± 0.19 m) for essentially the same biome. The deepest burns occur around the roots of vegetation which is confirmed in Figure 11
and Figure S1
where the some of the highest pre-burn peaks occur in the same location at the lowest post-burn pits, and by the combustion heterogeneity analysis based on analysis of the RGB imagery collected from the UAV (Section 4.1
). These DoB estimates translate to a carbon emissions estimate of 134 ± 29 tC·ha−1
, which is slightly more than the estimate from [15
] for “first fires” in these types of biome, and improves upon previously reported uncertainties of up to 32% [1
Although we used airborne LiDAR data to create our pre-burn DTM, we have also demonstrated that spatial interpolation of the post-burn UAV data alone can be used to create a pre-burn surface very similar in charcter to the actial LiDAR-derived pre-burn DTM (e.g., −0.01 m bias at 10 m resolution). When DTMLiDAR
is substituted for DTMIDW
the ‘post-burn only’ data still provides an overall uncertainty of <0.01 m in mean DoB asssessment. With further investigation, this method could eliminate the need for pre-burn LiDAR data, which are often not available in studies estimating depth of burn at new fire sites [9
Our combustion heterogeneity analysis demonstrates around 21% of the large shrubby peatland area was affected by complete combustion, and that maximum combustion occurs around vegetation roots, where the deepest burns are likely to occur, creating deep burn pits.
This is the first study to difference pre- and post- burn DTMs of the same area in tropical peat burns to assess emissions estimates. Previously, [9
] compared post-fire average elevation in burned areas with adjacent unburned area, and [15
] modeled a pre-burn surface using interpolation of adjacent unburned areas. While our study area is much smaller than these previous efforts due to flying limitations of the UAV, it offers several benefits. Firstly, the derived DTMUAV
is of very high resolution (0.1–0.5 m), and highly precise (with random errors much smaller than in [9
]). Secondly, the UAV can be deployed with very little expense, planning, or effort compared to a full airborne LiDAR campaign. Finally, the UAV system is capable of producing a pre-burn DTM (DTMIDW
) which is comparable to the pre-burn DTMLiDAR
(bias = 0.01 m). Here, the bias is more important than the random error because mean depth is used to calculate the volume of peat burned, and random error is cancelled out as long as elevation distribution is normal.
We have examined the effects of SfM spatial filtering on both pre- and post-burn datasets. By filtering each DTM using the lowest classified ground point per grid cell, we found it possible to create the required peat surface DTMs. In the pre-burn profile in Figure 11
, the green area shows how higher resolution filtering of DTMLiDAR
(1–10 m) seems to capture vegetation (as seen by the abrupt peaks) rather than ground height values, which are best characterised by 20 m resolution DTMLiDAR
. This suggests that processing pre-burn LiDAR data at high resolution in densely vegetated areas may overestimate pre-burn elevation because true ground points are less likely to be returned in smaller grid cells [44
], as could be the case in [9
]. On the contrary, a larger grid cell flattens and underestimates the post-burn DTM because interpolation may exclude much of the remaining peat. In Figure 11
captures true peat microtopography best at 0.1 and 0.5 m because the filtering effect of lower resolution DTMs excessively smooths the surface topography (as seen by the red dotted/dashed lines in the red area). It is thus important to find a balance between filtering out the impact of residual vegetation (which is improved with larger grid cells in heavily vegetated pre-burn datasets) and maintaining topographic detail in open post-burn datasets. Where possible, filtering algorithms should be applied to point cloud datasets, however the results should be examined carefully to ensure the impact of pre-burn vegetation is adequately filtered out in the resulting surface DTM.
While we have quantified known uncertainties, there remains a very important uncertainty that is not able to be considered here. Despite using recommended algorithms and techniques, there is no guarantee that the derived ground points, output after filtering, actually do represent ground height, rather than the top of a thick surface vegetation layer. The study area was heavily vegetated before the fire, with tall canopy trees (35 m) interspersed with dense ground shrubs. It is not possible to estimate whether LiDAR ground return did in fact come from the ground surface and not from the top of the surface vegetation layer. Previous accuracy assessments of DTM retrieval show their lowest random error of 0.12/0.19 m in unburned/burned tropical peat forest [15
], 0.58 m in unburned tropical forest as a result of incomplete canopy penetration [45
], and up to 0.26 m in densely vegetated wetlands [46
]—values that are in some cases larger than the DoB estimates made in this study. We have no way of quantifying pre-burn DTM error in our study area because there were no control point measurements taken beneath vegetation at the time. Some existing studies estimating DoB have also failed to quantify this uncertainty [9
], and therefore may suffer from a biased pre-burn surface leading to an overestimated DoB. This remains an obvious research question to be answered within the context of DoB measurements from LiDAR.
It took a total of 18 person-hours to set up the ground control points (GCPs) needed for the UAV survey, and a further 2 person-hours were needed to fly the complete survey. The TLS scans required a total of 48 person-hours. Furthermore, the UAV weighs 4 kg with its carry case and batteries, compared to over 30 kg for the TLS. In terms of out-right costs, the TLS costs >USD 110,000 (United States Dollar) compared to ~USD 2000 for the UAV and camera and educational software licenses (however, the UAV system requires either a GNSS for GCP registration, these can be rented from survey companies, or bought second hand for <USD 7500). The cost ($) multiplied by person effort (h) and weight (kg) (user inputs), divided by usable coverage area (m2) (outputs) equates to the UAV survey requiring 0.01 $·kg·h·m−2 compared to the TLS which requires 13.01 $·kg·h·m−2. Furthermore, because the UAV survey is conducted from above, there are fewer gaps in the DTM which require interpolation. Finally, the UAV has the added benefit over all LiDAR systems that it can (albeit more roughly) capture topography through water. Therefore, the UAV system can provide an efficient, cheap and flexible alternative to terrestrial laser scanning systems for point cloud retrieval in rugged, remote and challenging conditions in burned peat swamp forests. While UAV coverage might be much lower than airborne LiDAR, this method might be the only affordable option to REDD+ practitioners without large budgets.