Following and , we define peatlands as ecosystems with accumulated surface peat layers. conservatively proposes a minimum criteria of 5 % organic carbon content to 10 cm depth. Our training database, Peat-DBase , observed the 5 % organic carbon threshold for non-peat measurements but imposed no depth restrictions. This inclusive approach captures extensive peat depth data across diverse peatland ecosystems, which benefits ML algorithms that can only make reliable predictions in regions similar to their training data .

For model outputs, PeatDepth-ML enforces no specific thresholds around peat depth, enabling predictions across the full range from 0 cm onwards. However, when presenting results, we often highlight depths exceeding 30 cm – a common delineation in peat datasets . See and for further discussion on peat definition variations and implications.

PeatDepth-ML uses the same five arcminute grid and netCDF file format as Peat-ML , requiring conversion of training data and new predictors to this standardised format. New predictors, typically acquired as GeoTiff raster files, were processed using similar geospatial tools: Geospatial Data Abstraction Software Library GDAL; , Climate Data Operators CDO; , and NetCDF Operators NCO;.

PeatDepth-ML employs the same ML algorithm, parameter optimization, cross-validation method, and predictor selection process as the Peat-ML framework , as illustrated in Fig. . Similar to Peat-ML, the PeatDepth-ML workflow incorporates several key steps to enhance model robustness: (1) removal of predictors that promote overfitting through elimination of highly multicollinear features via variance inflation factor (VIF) calculation and recursive feature elimination, (2) Bayesian hyperparameter optimization, and (3) blocked leave-one-out (BLOO) cross-validation for predictor error estimation. An important distinction of the VIF over simple correlation, which only looks at pairwise relationships, is that it quantifies complex interdependencies and the combined effect of multiple predictors simultaneously. Our framework also uses a BLOO cross-validation approach as simple random splits result in overly optimistic error assessments due to spatial autocorrelation impacts . BLOO, conversely, may result in overly pessimistic error estimates , but we prefer the more cautious estimate.

PeatDepth-ML introduces a novel bootstrap resampling method and includes several adaptations to the Peat-ML framework necessary for peat depth modelling. One modification was to allow the model to generate values from zero upward, rather than constraining results to a maximum of 100 as in the original Peat-ML framework.

Like , we established minimum cross-validation block sizes by finding where spatial autocorrelation diminished, determined by calculating mutual information (MI) of model residuals at different lag distances. MI describes similarity between a variable at one location and neighbouring locations. Positive MI indicates similar values cluster together; negative MI indicates dissimilar values cluster. Values closer to zero indicate less clustering and autocorrelation . We calculated z-scores to establish statistical significance, with values near zero indicating low significance . Figure  shows MI approaches zero and z-scores decrease sharply by 6°. Though both metrics decrease further beyond 6°, we chose 6° as minimum block size to balance training data non-uniformity, ensuring the maximum number of blocks access diverse peatland types (high-latitude and tropical). While model residual z scores do not match random datasets' lows, they correspond to low MI values and are deemed acceptable. Geary's C and G statistics corroborated these results . Using 6° minimum block size and aiming for even data distribution produces 17 blocks (Fig. ) that delineate training versus testing cells in cross-validation.

We developed a custom scoring in LightGBM , to counter the training data's zero inflation, where peat-absent test cells received less weight (30 %) than peat-present cells (70 %) during cross-validation – a strategy from binary classification with skewed datasets . While peat-present cells comprise only 8.7 % of observed data, their model error then received greater weight. The custom loss function improved predictions of deeper depths but averaged less than 20 cm difference within known peatland regions (not shown). Grid-cell-level performance declined as expected since most test cells have zero peat depth, which we were now deliberately deprioritizing.

We also tested data preprocessing to reduce training data skew: log transformation , Box–Cox, and Log–sinh . However, we rejected all transformations as they provided minimal improvement while adding considerable computational cost.

The hyperparameters that were optimized for Peat-ML were also optimized for PeatDepth-ML . We performed our Bayesian hyperparameter optimization over 1500 model iterations and chose the values that produced the best results. These hyperparameters and their values can be found in Table  and are the same for every bootstrap run.

For the bootstrapping approach, we ran PeatDepth-ML for each bootstrapped dataset (401 in total) such that for each bootstrap we had the corresponding model output, cross-validation output, and the list of predictors used by that model run to produce said output. We then took the mean of the depth predicted over all the bootstrap model run outputs in each grid cell to produce a final PeatDepth-ML product (discussed in Sect. ).

PeatDepth-ML, , and were harmonized to enable statistical analysis. was resampled to match PeatDepth-ML and resolutions, then all data outside the bounds of were removed from both products. Mean peat depth was calculated for each product over approximately the same area, with PeatDepth-ML bootstrap means calculated then averaged together. PeatDepth-ML uncertainty bounds were derived from bootstrap outputs by determining the 5th and 95th percentiles of bootstrap means, then calculating symmetrical uncertainty as half the difference between these percentiles.

PeatDepth-ML predicts mean peat depth only over the peatland area within each grid cell. To allow estimation of global peatland C stock, we require a peatland fractional extent map that will give the peatland fraction of each PeatDepth-ML grid cell. While any fractional peatland map can be used, we select Peat-ML as our “best” estimate as it is on the same grid, provides fractional extent, and is a peer-reviewed product . To provide an alternative estimate we also use the Global Peatland Map (Greifswald Mire Center, unpublished dataset, https://greifswaldmoor.de/global-peatland-database-en.html, last access: 24 April 2026) although it is unpublished and little information is available about its provenance. To obtain a basic estimate of global peatland carbon stocks we follow an approach similar to : 1Cdens=OC×BD2Cstock=Cdens×Dpeat×Acell where organic C content (OC) is in kg C kg⁻¹, bulk density (BD) is in kg m⁻³, C_dens is C density in kg C m⁻³, Dpeat is peat depth in m, Acell is grid cell area in m², and C_stock is C stock in kg. We calculated Acell using Peat-ML's peatland fractional coverage by determining each grid cell's trapezoidal area and scaling by peat coverage fraction.

We calculated C stock estimates using Cdens, OC, and BD values from other studies to enable comparison. Specifically, we tested C_dens values derived from the mean modelled OC and BD values in . Where their predicted peat depths exceeded 200 cm, they used their 100–200 cm OC and BD values for deeper depths, an approach we adopted. We also tested OC and BD values from two approaches. Both methods use tropical OC and BD from their literature review, while employing high-latitude C_dens values based on estimates (assuming 1.5 m average peat depth; termed “Page1.5”) for one method, and recalculating high-latitude C_dens by combining data with 2.3 m average peat depth from for the second (termed “Page2.3”). These two approaches yield high-latitude C_dens values of 73.30 and 63.74 kg C m⁻³ for Page1.5 and Page2.3, respectively.

Across 401 PeatDepth-ML bootstrap runs, 123 different predictors were selected by one or more runs (bold predictors in Table ). However, only 10 predictors were selected by over 75 % of runs, indicating many were not consistently important. Figure  shows the top 15 predictors by averaged importance across all runs. The highest average importance predictor is average VPD over SON (September–November), selected in 87 % of bootstrap runs and always ranking among the top two when chosen. Second is average SWE over DJF (December–February), selected in 62 % of runs with a more even importance distribution than VPD SON. No other predictors approach VPD SON and SWE DJF importance levels, except overall mean annual VPD which frequently exceeds 10 % importance when selected but appears in only 26 % of runs, yielding lower average importance. Beyond these three, most predictors range from 0 % to 10 % importance. No predictor was selected across all 401 runs, though SWIR3 MAM (March–May) standard deviation came closest (selected in 99 %), followed by Runoff SON and geomorphological landforms indicator (Geomorphon) at 96 % selection each.

Of the top 15 predictors, 10 relate to climate, two to vegetation, two to terrain, and one is a paleo-environment predictor (Table ). In their analysis of peat formation drivers, list climate among the most important globally, alongside vegetation and topography. PeatDepth-ML therefore selects and utilizes predictors associated with peat formation despite strong variability in predictors selected arising from variability in the training data stemming from the bootstrapping. This variability in predictor importance might be also at play in other studies. For example, train separate models for six different large regions. They find the predictor importance to vary strongly between each region (their Fig. 5), which could be at least partially due to the strong dependence upon training data we see here. This also raises the possibility that training models for smaller regions and then stitching together could bring in artifacts due to the smaller domains for each model restricting the amount of data to train upon.

The variability in predictor importance and selection makes any mechanistic interpretation of the predictors to peat formation processes challenging, nevertheless, we can make broad observations about overall PeatDepth-ML behaviour.

Frequent to continuous waterlogging from precipitation, groundwater, or both is important for peat accumulation initiation and persistence . Figure  shows that except for downward surface shortwave radiation, wind speed, and year of ice or sea retreat, most of the top 15 climate predictors closely relate to factors that influence water or moisture presence. PeatDepth-ML likely selects VPD SON (and potentially mean VPD) to delineate shallow or non-peatland areas. Given strong zero-inflation in training data, the model likely finds improved scoring during cross-validation when choosing predictors informative for non-peatland regions, even with our custom scoring metric.

SWE DJF may separate high-latitude from tropical peatlands, as suggested for Peat-ML . Peat-ML also ranked SWIR3 among top predictors, suggesting it differentiates wet/dry earth and identifies fens. SWIR3 MAM standard deviation is PeatDepth-ML's most selected bootstrap predictor, and may be used to assess areas experiencing excessive moisture fluctuation that prevents deep peat accumulation, as low moisture increases respiration, limiting peat growth.

PeatDepth-ML frequently selects Runoff SON, potentially indicating adequate levels of precipitation for coastal and tropical peatlands . Geomorphon appears in both Peat-ML and PeatDepth-ML where it provides topographical characteristics conducive to peat development .

Global peat depth averaged across all PeatDepth-ML bootstrap runs, hereafter called the modelled mean, appears in Fig. . The modelled mean indicates peat depths exceeding 30 cm in most major peatland regions including parts of Canada, Congo Basin, Scandinavia and Northern European Plain, West Siberian Lowlands, and Malay Archipelago areas.

Some data gaps exist in PeatDepth-ML outputs due to missing predictor data. Figure  shows the modelled mean lacks data for parts of the Canadian arctic (Victoria Island, parts of Barren Grounds, Baffin Island), all of Greenland, and parts of Russia's Novaya Zemlya archipelago. The model cannot predict for grid cells lacking all provided predictors. Persistent gaps in the modelled mean result from gaps present in identical locations throughout every bootstrap run. Of grid cells with modelled mean data, only 0.5 % had missing data in some bootstrap outputs.

Not all areas predicted from a ML-based model should be considered equally trustworthy. To investigate our spatial applicability, Fig.  shows the results of the AOA analysis across all bootstrap runs, with gradients indicating percentage of runs where grid cells were deemed applicable. For locations with aforementioned missing data in some bootstraps, these count as reduced applicability corresponding to missing data frequency. Of grid cells experiencing any inapplicability, 4.2 % result from occasional bootstrap data gaps, mostly along northern Caspian Sea coasts.

Globally, the modelled mean shows lowest applicability over mountainous regions (Fig. ). Although training data exist in or near mountainous regions, coverage is limited, particularly in Indonesia (Fig. ). This data scarcity likely contributes to higher inapplicability in these areas. Regions in the Siberian Plateau, East Siberian Mountains, and areas extending south from the Chukchi Peninsula to the Kamchatka Peninsula show 80 %–90 % applicability, also reflecting limited training data in those regions. Additionally, since AOA depends on model predictors for each run, it reflects training data uncertainty through bootstrapping. This result is important as it implies that single model instances, as are common in the literature, are potentially unreliable. This also further raises questions about the practice of training different models for different regions of the globe and then stitching them together e.g..

Figure  compares PeatDepth-ML to and ; however, quantified comparisons are challenging due to differing resolutions and spatial extents. provides depth predictions at 1 km spatial resolution over “peat-dominated” areas from the Global Peat Map version 2.0 GPM v.2; , with modifications in Indonesia, Sweden, Denmark, and the Netherlands where peat thickness maps were randomly sampled and treated as peat cores. predicted peat depths north of 23° N at 0.1° resolution. do not include non-peat data in their depth modelling process, instead combining peat depth results with their own coverage map to mask non-peatlands. Thus both and rely upon masking methods to delineate peatland regions and also do not provide estimates of spatial applicability or extrapolation.

Visual comparison of our modeled mean with (Fig. ) shows deepest peat depths in similar regions: Indonesia, West Siberian Lowlands, Hudson Bay Lowlands, Canadian Boreal Plains, Pastaza-Marañón Foreland Basin, and Congo Basin. Key differences include showing deeper peat across more of the Congo Basin, while our model indicates deeper peat in southern Chile and Argentina. In Alaska, predicts shallow depths whereas our model shows relatively deep peat. The Tibetan Plateau, excluded from (via the GPM v.2 mask), appears as a significant peat complex in PeatDepth-ML.

Comparing with both products at northern latitudes reveals it predicts the most widespread and consistently deepest peat, though large areas are filtered during post-processing (continental US, central Eurasia, northern Sahara). Notable differences occur in Alaska, Hudson Bay Lowlands, and West Siberian Lowlands, where predicts shallower peat than our model and . While can predict depths < 30 cm (India, US Coastal Plains), the scarcity of non-peat regions demonstrates this product's reliance on their peatland spatial coverage map.

Regional peat depths for PeatDepth-ML, , and are presented in Table  and Fig. . Due to the skewed nature of the peat depths, the results are presented as medians plus the interquartile range. PeatDepth-ML generally predicts shallower median depths than the other two products, particularly in the tropics where exceeds PeatDepth-ML by over half a metre. In the northern hemisphere, achieves the deepest mean depth, consistent with Fig. .

Figure  shows depth value distributions across products. Since the products treat non-peat regions differently, depths > 30 cm are most comparable (30 cm commonly delineates peatland classification; ). Additionally, the general shapes and peaks of the histogram curves are likely most comparable given the products' different spatial resolutions and peatland area delineations. All three products exhibit similar relative depth distributions with peaks around 250 cm. PeatDepth-ML and show strong agreement, though extends to greater maximum depths. again predicts deeper tropical depths than PeatDepth-ML, with more data points within the harmonised area due to its higher spatial resolution.

Greater training data availability may explain why achieves significantly deeper mean tropical depths than PeatDepth-ML. incorporates more non-zero peat depth training data for tropical regions, particularly Indonesia, partly by including pseudo-observations – grid cells randomly sampled from existing peat depth maps, such as the Indonesian national peat depth category map , using category midpoints as depth values. However, this approach introduces uncertainty by treating spatially extrapolated peat depths as equivalent to discrete soil cores. In fairness, we similarly add pseudo-observations, however the influence is found to be minimal as they are introduced into peat-free regions and since our bootstrapping results demonstrate that depth values assigned to training points can strongly influence model behaviour, we did not incorporate pseudo-observations in regions with potential peat soils.

While direct comparisons are limited by methodological differences (e.g., used regional models; covers only high latitudes), PeatDepth-ML's relatively shallow predictions in Table  likely result from zero-inflated training data and ML algorithms' tendency to predict towards the means of training data . observed this behaviour in their RF model, and reported similar results with both RF and Maximum Entropy models when predicting mean vegetation canopy height. The empirical cumulative distribution in Fig.  confirms PeatDepth-ML follows this pattern. Over 90 % of training and testing data show zero depth, while approximately 75 % of PeatDepth-ML bootstrap predictions are at, or near zero, over the same grid cells. The observation and prediction distributions intersect around 150 cm, with PeatDepth-ML approaching 100 % distribution by 400 cm depth, whereas training data extend deeper without reaching 100 % as quickly. Consequently, PeatDepth-ML poorly predicts extreme values. For combined zero and non-zero depth data, PeatDepth-ML closely matches training data means; however, when considering only non-zero depths, PeatDepth-ML predicts shallower values overall (Fig. ).

Beyond the previous qualitative assessments, we conducted quantitative evaluation of PeatDepth-ML performance using multiple metrics. We began with grid cell comparisons between model-predicted depth values and observations at these locations, calculating root mean square error (RMSE), mean bias error (MBE), and a version of the normalized mean error (NME) developed by (see Table ). RMSE and MBE are commonly used metrics for assessing ML models , with both and employing RMSE, and Peat-ML calculating both RMSE and MBE. Although Peat-ML and use the Coefficient of Determination (R2), we excluded it due to its ambiguous representation of non-linear model performance . We adopted 's NME to enable comparison against two null models: the observed mean null model and the observed random resampling null model (Table ; Figs. d and ).

We calculated these metrics over selected areas for each bootstrap model run using cross-validation results. Figure  shows the regions selected for focussed assessment. We sought equal representation of high-latitude and tropical areas, choosing regions with high peatland fractional coverage according to Peat-ML and substantial observed peat-present grid cells (Fig. ). Each bootstrap model run output is compared to the bootstrapped observed data used in training and testing that model instance. Calculating metrics for each bootstrap run demonstrates variation in model behaviour arising from sampling uncertainty in observed peat depth data. However, this same uncertainty limits the accuracy achievable in our performance assessment. For more realistic representation of model ability, metric calculations use model results produced through BLOOCV (Sect.  and Fig. ). The BLOOCV results represent PeatDepth-ML's predictions without the advantage of learning from observed peat data in the current area.

Figure  shows mean depth and performance metrics for PeatDepth-ML bootstrap runs. Throughout panels (a)–(c), bootstrap variance in selected regions remains broadly consistent with fluctuations of roughly 10–25 cm. The Congo Basin demonstrates the best performance among selected peatland regions (Fig. b). We hypothesize this reflects the Congo Basin's ratio between peat-present and peat-absent training data mirroring the global ratio (see Fig. ); therefore, globally learned patterns may be more applicable in this region. Conversely, the Malay Archipelago shows the worst RMSE performance (Fig. b) and most severe depth under-prediction overall (Fig. c). We attribute this poor performance to insufficient training data in the Malay Archipelago (Fig.  shows the lowest number of training points among all selected regions, despite its size). Additionally, any performance assessment is likely inadequate in the parts of this region where PeatDepth-ML is found to be generally not applicable according to AOA (Fig. ).

The NME results differ markedly from other metrics (Fig. d). Generally, NME ranges between 0.6 and 1.0 for most peatland regions and globally. However, regional variation is less consistent, with the Congo Basin showing particularly wide-ranging NME scores across bootstrap runs. Since the NME denominator is observational variance (Table ), our bootstrapping method can produce more variable values. Additionally, in a grid cell when there are more observed data points with zero peat depth, the NME denominator also approaches zero; leading NME to approach infinity. The Congo Basin contains several grid cells with exceptionally high peat depth measurement densities, allowing bootstrapped mean observed depths to vary more substantially. Moreover, the Congo Basin has the lowest non-zero depth ratio among selected peatland regions (Fig. ). Together, these factors create high observational variance in the Congo Basin, with similarly high variance in NME.

Following , we compared PeatDepth-ML's NME against observed random resampling null models (Table ). These random null models were generated by bootstrapping our 67 208 observed grid cells 1000 times, creating datasets with random selections while maintaining the same total number of grid cells. We used the non-bootstrapped observed data version (Fig. ). NME was calculated for each random null model and plotted in Fig. a. Figure b compares random null model results to global PeatDepth-ML bootstrap results, assessing whether PeatDepth-ML outperforms random resampling. Overall, PeatDepth-ML performs better than both the observed mean null model and random null model. However, Fig. d shows that training data uncertainty can produce extreme cases performing worse than both null models (e.g., outlier bootstraps in the Congo).

PeatDepth-ML's RMSE can be compared to and , acknowledging key methodological differences. used 10-fold cross-validation with balanced random sampling , while randomly sampled 30 % of data for testing. As discussed in Sect. , random cross-validation can produce overly optimistic performance estimates . Our spatial autocorrelation accounting in cross-validation block selection provides more conservative performance estimates , potentially yielding comparatively lower metrics. Additionally, models only northern latitudes, while combines six regional models (see their Fig. 1, but boundaries are not explicitly defined in the text), complicating direct comparisons. Resolution differences further limit precise comparison.

Before bias correction, reported a RMSE of 142 cm (no after-bias correction RMSE was supplied). PeatDepth-ML achieves mean RMSE of 88 ± 1 cm for ≥ 23° N. Table 3 in reports regional RMSEs of 163 cm (North America), 92 cm (Europe and Russia), 55 cm (Latin America), 91 cm (Africa), 215 cm (South and Southeast Asia), and 100 cm (Australia and New Zealand), which we calculate to give an average of 104 cm globally (weighted by the amount of testing data per region). PeatDepth-ML achieves global mean RMSE of 70 ± 1 cm. In the Malay Archipelago – comparable to the South and Southeast Asia domain in – PeatDepth-ML shows mean RMSE of 236 ± 7 cm, likely due to training data scarcity as suggested by AOA analysis. The highest RMSE in also occurs in this region despite greater data availability through the use of pseudo-observations derived peat depth maps, suggesting region-specific modelling challenges.

PeatDepth-ML was developed at the global scale to avoid complications from regional approaches. The multi-domain methodology used in requires harmonizing outputs at boundaries, with uncertain effects on AOA. Regional divisions can create disparities in peat observations, affecting model quality and requiring region-specific spatial autocorrelation distances and non-peat data tuning. Preliminary multi-model tests during PeatDepth-ML development showed extreme behaviour (e.g., single-predictor selection) in certain regions, likely from increased zero-inflation in regional datasets. Furthermore, our bootstrapping reveals significant observational uncertainty influence on the model (Fig. ), which may intensify at smaller regional scales and vary by region.

PeatDepth-ML requires a peatland fractional extent map to produce an estimate of peatland carbon stocks as the model estimates the mean peat depth only over the peatland area within each grid cell. We have chosen Peat-ML to provide peatland fractional extent as our “best” estimate (see Sect. 2.5), and will be the focus of further discussion, but any other peatland fractional extent map could be used and we provide an estimate based upon the GPM as an alternative. PeatDepth-ML estimates global peatland carbon stocks between 327 and 373 Pg C, with 7 %–10 % located in tropical peatlands (Fig. ). This tropical proportion is lower than , who estimate 15 %–19 % of global peatland carbon in the tropics. Figure  shows that PeatDepth-ML generally estimates peatland carbon stocks within the lower part of the range provided by other products . reports particularly elevated estimates, with a global value of 942 Pg C, attributed to extensive peat coverage in the Global Peat Map (6.57 million km²) compared to Peat-ML's 4.04 million km². For the tropics, also has a high estimate, over three-fold the next higher estimates (but includes the subtropics as well). However, compared to the more moderate estimates in Fig. , PeatDepth-ML's “best” estimates are commonly on the lower end of the range especially when considering only the tropical peatlands. A potential shallow bias in PeatDepth-ML may contribute to these lower carbon stock estimates, however the larger influence is the selection of Peat-ML to derive our “best” estimate, which has a smaller peatland area estimate than GPM as used by and results in an estimate of peatland C stocks less than half those derived via GPM (Fig. ). The strong influence of peatland fractional cover maps in the calculation of peatland carbon stocks reflects not only differences in total peatland extent (e.g. the large differences between Peat-ML and GPM as mentioned above) but also differing fractional extents in areas of shallow versus deep peat (even if total extents were the same) will have large impacts on total peatland C estimates.

As recognized by and mentioned earlier, we cannot readily draw conclusions about potential peat-forming conditions from the predictors selected by a ML model, as it is challenging to distinguish between cause and effect in these selections. Our examination of PeatDepth-ML's sensitivity to uncertainty in the training data revealed that predictor selection can be highly variable, rendering assumptions about their relationship to peat development potentially even more ambiguous.

PeatDepth-ML, like all data-driven models, is limited by the availability of peat depth observations for training. Both and demonstrated that the accuracy of their products is affected by training data availability, with each dataset biased towards high latitudes. This bias is also present in PeatDepth-ML and is evident in our poorer performance in the Malay Archipelago (Fig.  and possible underprediction of peat depth at low latitudes (Fig. ).

We focused on predicting peat depth using the LightGBM algorithm based on the Peat-ML Framework (Fig. ); however, other algorithms may offer better performance. and demonstrated that RF models can also predict peat depth. Both RF and LightGBM are ensemble models built from decision trees, which effectively handle missing data but are less suited to tasks with small sample sizes . Therefore, other algorithms, such as neural networks, may achieve higher accuracy with highly imbalanced datasets like ours . Alternatively, a first binary classification step of peatland versus non-peatland regions before predicting depth could improve results . Alternative custom scoring methods could be tested, such as weighting based on the imbalance ratio between zero and non-zero depths . Bias correction offers another potential improvement through residual rotation or empirical distribution matching .

Several aspects warrant further investigation. Including more palaeoclimate predictors could improve model performance by better representing the extensive time scales over which many peatlands developed. Distribution-scale assessment using empirical cumulative distributions may yield additional insights into PeatDepth-ML's performance. More thorough testing of regional modelling approaches could examine how model sensitivity and AOA vary by region, though methods to account for extreme zero-inflation in some regional training datasets would likely be required first. Additionally, more detailed C stock estimations could be performed using depth-to-stock relationships derived from peat core data , and further tests on the influence of random sampling of desert data may reveal additional model sensitivities.

Lastly, our peatland carbon stock estimates are heavily influenced by the choice of the peatland fractional extent map as seen in Sect. 3.4. Future work could reduce the grid spatial resolution (e.g. 250 m) to make the use of a peatland fractional map unnecessary.