We used a suite of predictors that fell into four main types: climate, soils, vegetation, and terrain (geomorphology). Table  lists each predictor grouped by predictor source and type. The climate, vegetation, and soil predictors were extracted from the Google Earth Engine data catalogue . The geomorphological dataset was downloaded directly from its authors' website last access: 16 January 2020. Sampling across the different years provided by each dataset is assumed to be relatively unimportant as peatland extent is not highly dynamic across decadal timescales, especially considering the scale of our grid cells . An additional predictor was the calculated length of the longest day of the year (hours) for each cell on the 5 arcmin grid. The longest day of the year could be used by the model to determine tropical versus extratropical regions.

The climate predictors were derived from the TerraClimate dataset . TerraClimate is available at high spatial resolution (1/24∘) and provides monthly climate and climatic water balance variables spanning the 1958 to 2015 period. TerraClimate uses the WorldClim dataset for high spatial resolution climatic normals, which is combined with the time-varying climate of the Climate Research Unit Ts4.0 CRU Ts4.0;, where the time-varying anomalies of CRU Ts4.0 are interpolated to the high-resolution climatology of WorldClim. The Japanese 55-year reanalysis JRA55; is used to fill in where CRU Ts4.0 has no climate stations contributing to its record (such as parts of South America, Africa, and smaller islands) and was the sole data source for solar radiation and wind speeds. notes that the water balance model, used to generate some of the variables listed in Table , is simple and does not account for vegetation heterogeneity or their physiological response under varying environmental conditions. For the climate predictors, we computed seasonal means across the available years, i.e. December–February (DJF), March–May (MAM), June–August (JJA), and September–November (SON). Given that these seasonal means are likely less important in tropical regions, we did investigate using annual minimum and maximum values in place of seasonal ones but did not see a significant impact on predicted peatland fractional cover.

Soil predictors were obtained from the 250 m resolution OpenLandMap including soil bulk density (kg m-3), clay content (%), sand content (%), organic carbon content (%), and soil water content at field capacity (33 kPa). These soil variables are derived from an ensemble of machine learning algorithms trained on a global compilation of soil profiles . We used the 30 cm depth estimate for all soil variables.

Terrain information is provided by the 250 m resolution version of Geomorpho90m for 17 different geomorphometry variables describing numerous aspects of the land surface (see Table ). This geomorphology dataset has an original resolution of 90 m, the same resolution as the Multi-Error-Removed Improved Terrain (MERIT) DEM from which it was derived. MERIT-DEM is a merged and error-corrected product based on the ALOS World 3D – 30 m (AW3D30) and Shuttle Radar Topography Mission (SRTM3) datasets.

Information about the vegetation state was provided by several datasets. created a seamless global mosaic image from the Phased Array type L-band Synthetic Aperture Radar (PALSAR/PALSAR2). This image was created with 25 m grid cells on an annual timescale. In creating the mosaic, at each location within a year the images chosen were those showing minimum response to surface moisture. The images were then ortho-rectified, slope corrected, and had a destriping procedure to equalize differences between strips that could occur due to conditions at time of acquisition. As the dataset's intended purpose was to provide a global mask of forest cover soil moisture differences were purposefully minimized, and thus this dataset is likely of more limited use to predict peatland extent than would otherwise be expected for an L-band radar product . However, likely due to the significant computational effort required to produce a global L-band product, we are not aware of another product publicly available.

The MODIS Terra net primary productivity product (MOD17A3.055 NPP) is available annually on a 1 km grid . This version of MODIS NPP (v. 5.5) is corrected for issues relating to cloud-contaminated MODIS leaf area index fraction of photosynthetically active radiation (LAI-FPAR) inputs to the MOD17 algorithm. We averaged the data over the available 2000–2015 period.

Vegetation indices are provided by the Suomi National Polar-Orbiting Partnership (S-NPP) NASA Visible Infrared Imaging Radiometer Suite (VIIRS) product VNP13A1, which is generated by selecting the best pixel at 500 m resolution over a 16 d acquisition period. The VIIRS data are generated for three vegetation indices including the normalized difference vegetation index (NDVI), which uses both red and near-infrared (NIR) bands, and two enhanced vegetation indices (EVI, EVI2), which also include the blue band with EVI2 designed for intercomparison with other EVI products that do not use a blue band (Table ). EVI is more sensitive to canopy cover, while NDVI is more sensitive to chlorophyll . All snow, cloud, or cloud shadow pixels and any pixels that were not excellent, good, or acceptable quality (according to the dataset's quality flags) were excluded. Given the original data do not have composite monthly values, the mean, minimum (min), maximum (max), and standard deviation (SD) were all calculated based upon all values within a year and then the average was taken across all years.

Vegetation phenology information is provided by the MCD12Q2 V6 Land Cover Dynamics product . The MODIS vegetation phenology product provides phenological information such as the dates of green-up, peak, and senescence along with variables related to the range and summation of the EVI (see Table ). Since this is an annual product the mean, min, max, and SD values are calculated across all years.

We also considered the global surface water (GSW) dataset of but did not include it as a predictor dataset. We found this dataset to be unsuitable for peatland prediction due to its reliance on Landsat imagery. Treed peatlands, peatlands smaller than 30 m by 30 m, and peatlands where the water table is below the peat surface, such as bogs, would not be well captured by GSW. A visual inspection of GSW over some of our training regions (see Sect. ) showed poor correlation between GSW water presence and mapped peatland area.

The statistical modelling was conducted in the Python programming language (v. 3.8.3). We use a gradient boosting decision tree (GBDT) algorithm called LightGBM . Decision tree algorithms make iterative splits to partition data according to different criteria. The decision tree will split each node at the feature with the largest information gain, i.e. the most informative. For GBDTs, the information gain is usually measured by the variance after splitting. To avoid issues with overfitting of a decision tree, GBDT algorithms use the boosting technique, which combines multiple decision trees in series to achieve better predictive power as each tree in the series attempts to minimize the errors in the previous tree. The error minimization steps occur through a form of gradient descent in function space where each tree is trained on a residual vector that measures the magnitude and direction of the true target relative to the previous tree (loss function), which successive iterations minimize.

To provide estimates of the error associated with the LightGBM predictions we adopted a blocked-leave-one-out (BLOO) strategy, which is recommended for applications where the predictors could be expected to exhibit spatial autocorrelation . BLOO tends to produce estimates of prediction error that are closer to the “true” error , particularly in cases where the sampling strategy is clustered . We chose to block our cross-validation (CV) regions based on longitudinal limits to allow both boreal and tropical peatlands to potentially be represented in each block. The optimal number of training blocks is an important determination. Choosing blocks that are too small risks incorrectly increasing our CV-determined model accuracy due to spatial autocorrelation issues, while choosing overly large blocks will result in information loss and worsens our assessed model accuracy unduly. We determine the optimal number of blocks by comparing the length scale of autocorrelation of the model residuals with our block sizes. Figure  shows the autocorrelation tends to zero at a length scale (sill) of around 500 km. To accommodate this we set a minimum block size of 10∘ of longitude (which corresponds to roughly 500 km at 65∘ latitude). Based on the constraints of our minimum block size and the need for a roughly even number of training cells in each block, we end up partitioning the globe into 14 blocks as shown in Fig. . The CV was performed by holding out one block, training the LightGBM algorithm over the other blocks, and then using that trained model to predict the peatland extent over the held-out block. This was performed for each block in turn and the results averaged to give an estimate of the prediction error.

From the potential peatland covariates listed in Table , and discussed in Sect. , we processed 163 global peatland features that could be used by the machine learning model. However, it is likely that many of these predictors will have low predictive power and duplicate information provided by other predictors, leading to over-fitting by the ML algorithm . To select only the most relevant features we used both iterative feature removal based on the calculated multicollinearity and recursive feature elimination with cross-validation (RFECV) , which is a form of backward feature elimination.

Multicollinearity was accounted for by using the calculated variance inflation factor (VIF) to identify and remove highly correlated variables . VIF uses ordinary least-squares regression to determine collinearity with the score determined by 1VIF=1(1-Rj2), where Rj2 is the multiple coefficient of determination for the feature j on the other features (covariates) defined as the ratio between the sum of squares due to the regression (SSR) and the total sum of squares (SST), 2Rj2=SSRSST. One approach would be to simply set a threshold VIF value and remove all features with VIF values higher than this threshold in a single step. However, in order to avoid the elimination of potentially important features, we chose instead to conduct the exclusion process iteratively. First, each feature had its VIF score calculated. Then all features with a VIF value higher than 5 (corresponding to a Rj2 of 0.8) were ranked according to their information gain calculated by LightGBM, and the feature with the lowest gain was removed. The model was then retrained and the VIF value recalculated. If features remained that had a VIF above the threshold value, the same ranking and removal would occur until all remaining features had a VIF value below threshold. This step retained 30 features (listed in Table ). The VIF value chosen is quite stringent, well below what suggest as a critical value (10).

We use RFECV with BLOO CV (using the same blocks as described in Sect. ) in an iterative manner to ascertain the optimal number of features. RFECV first trains the LightGBM algorithm on the original number of features (here 30) with the features ranked for their importance, based on information gain, for the model's root-mean-square error (RMSE) as determined by the BLOO CV. The least important feature is removed and the model is retrained using the new subset of features. By retraining the model after each feature is held out, we avoid the issue of extrapolation that can occur in permutation-based approaches as described in. The algorithm can then produce an estimate of model skill as a function of the number of features trained (Fig. ). The RFECV algorithm will choose an optimal number of features based on the greatest model skill. Based on Fig. , 16 features (highlighted in Table ) were selected as the optimal number to retain for the optimization process and the final model.

GBDT algorithms tend to require hyperparameter tuning to ensure the model is performing optimally. We employed Bayesian optimization on 11 LightGBM hyperparameters (Table ) using the hyperopt package over 500 trials. In each trial, the final 16 predictors identified in the steps above were used in the LightGBM model to optimize the model's calculated RMSE based upon the BLOO CV. The optimized parameters were then used to generate the Peat-ML map.

Global peatland extent as predicted by Peat-ML is shown in Fig. . When Peat-ML is compared to PEATMAP , many major peatlands regions appear similar including Canada, the WSL, the Cuvette Centrale of the Congo, and parts of Scandinavia. However, the two maps also differ substantially. The regions with the most notable difference between the two products include Alaska, parts of Africa excluding the Congo, and eastern Siberia. There are more intermediate peatland extents predicted by Peat-ML, whereas PEATMAP tends to show more regions of 100 % peatland extent with less gradation between peatlands. Our estimated global peatland extent at 4.04×106 km2 is similar to the PEATMAP estimate of 4.23×106 km2 (Table ).

Our Northern Hemisphere (>23∘ N) estimates of 3.0 million square kilometres is lower than the other available estimates including PEATMAP (3.2 million square kilometres), the lower bound of (3.2 million square kilometres), and an older estimate of , but it is at the lower bound suggested by . In the tropics, our model estimate is roughly the same as PEATMAP but only a little over half of the extent estimated by . The map was produced through a hybrid approach that uses hydrological modelling, remote-sensing products, hydro-geomorphology from topographic data, and expert assessment. It is only available across the tropics (maximum 40∘ N).

The spatial distribution of the predicted peatlands will now be examined in detail. We focus on regions that have either multiple other peatland mapping products for comparison or contain large areas of predicted peatlands.

Figure  shows the peatland extent in the WSL, western Russia, and parts of Scandinavia for Peat-ML, its training data, PEATMAP, , and the Boreal–Arctic Wetland and Lake Dataset (BAWLD) . The dataset is derived from the mean of two soil datasets and is only available for the Northern Hemisphere (>23∘ N). The BAWLD product is derived from expert assessment that is then extrapolated through the use of random forest models and geospatial datasets across the boreal and Arctic regions. The original spatial resolution is relatively coarse at 1∘ by 1∘. For the WSL region, all four products are similar, with only slight differences in the peatland fractional cover (rather than its spatial distribution). Peat-ML shows strong similarity with its training data as would be expected. PEATMAP stands out compared to the other maps due to its almost binary peatland coverage showing either high values or no peatlands with little gradation in between. Compared to , Peat-ML shows less peatlands in the northern edge of the Northwestern region of Russia but more by the White Sea. Both PEATMAP and Peat-ML do not show peatlands near the mouth of the Kara River to the northwest of the terminus of the Ural Mountains, as evident in and BAWLD, while Peat-ML and BAWLD show few peatlands on the Yamal Peninsula, where both PEATMAP and suggest appreciable extents. Generally, Peat-ML has more similarity to PEATMAP than and BAWLD over the western Russian domain.

All maps show relatively similar distributions of peatlands surrounding the Baltic Sea (Fig. ). None of the maps indicate peatlands by the Caspian Sea as seen in PEATMAP, except some small extents (1 %–3 % predicted by Peat-ML) to the northwest of those depicted in PEATMAP.

As with Eastern Europe, Western Europe is similar in that PEATMAP shows a more binary representation of the peatland extent compared to the other maps (Fig. ). Peat-ML and have fairly similar peatland distributions and extents. The main differences are expressed in small pockets of peatlands, e.g. eastern Spain has scattered peatlands in that are not found in Peat-ML or PEATMAP, whereas in western Hungary both and PEATMAP show small peatlands not predicted to be as extensive by Peat-ML.

For the northern contiguous USA, for Canada, and for Alaska, peatlands extents are shown in Fig. . Alaskan peatlands predicted by Peat-ML have some similarity to the map and BAWLD, with extensive peatlands in western Alaska (Lower Yukon region). These peatlands are not evident in PEATMAP, which shows less extensive but high-coverage peatlands along the southern and eastern edges of the state. Peat-ML, , and BAWLD predict peatlands along the Alaska North Slope that are not evident with PEATMAP. Other reports suggest extensive wetlands in Alaska e.g., but we are not aware of any mapping product detailing peatland-specific coverage.

For Peat-ML, the Canadian peatlands from and DUC are used as training data, which naturally gives good correspondence between Fig. a and c. For a more informative comparison of the general model skill for boreal peatland regions, Peat-ML predictions from the BLOO CV simulation are also shown, as this would give some indication of predictions without the benefit of training upon a particular region's peatlands (Fig. b). Generally, all datasets shown in Fig.  display some strong similarities, with large peatlands shown for the Hudson's Bay Lowlands (HBL), the Mackenzie Delta, and across the Boreal Plains, yet important differences are also visible. shows little peatland along the southern edge of the Hudson's Bay, perhaps due to their peatland determination model's emphasis on treed peatlands. also show generally higher peatland coverage where peatlands are present than the other datasets. predicts extensive but relatively low coverage across much of the Canadian eastern Arctic that is not found in any of the other peatland maps. Of the five peatland maps, the most closely corresponding peatland extents appear to be between PEATMAP, BAWLD, and Peat-ML.

The northern USA has some peatlands around the Great Lakes evident in PEATMAP and (∼10 %–60 %), which are also predicted but appear less extensive in Peat-ML (usually ∼1 %–15 %). Besides the coverage differences, the products have a similar spatial extent, although PEATMAP's peatlands are more commonly higher coverage per identified peatland.

South American peatlands are shown in Fig. . Peat-ML peatland training data for this region (Fig. a) are currently limited, encompassing only Peru's Pastaza-Marañón foreland basin (PMFB) and the Rio Madre de Dios. In early simulations with Peat-ML and maps from other modelling processes e.g., we noticed predictions for high peatland coverage in areas of South America where peat is not known to occur. This includes seasonally flooded savannas, such as the Llanos de Moxos (Beni Savanna) and Llanos Orientales of Colombia and Venezuela. A recent field expedition searching for peat in the Colombian Llanos failed to discover any peat deposits , which could indicate that these tropical savanna biomes are generally not able to form extensive peat deposits. Additionally, white sand ecosystems are not known to support extensive peatlands, and thus we also excluded the Rio Negro Campinarana ecoregion that corresponds with white sandy soils (Spodosols/Podzols) and not Histosols. Without these negative data, we would likely overpredict peat extent in South America rather severely.

Peat-ML predicts an extensive peatland in the PMFB and central Amazonia. The extent of peatlands in this region is lower than in PEATMAP, mainly due the generally lower extent per grid cell, despite being in broadly similar regions. Both PEATMAP and Peat-ML show peatlands along the northeastern coast of the continent. Peat-ML predicts smaller peatland extents (generally <10 %–15 % coverage) in the Pantanal and along the Paraguay River as it joins the Paraná River down to the Rio de la Plata, which are not evident in PEATMAP.

There are some non-peatland river floodplains that Peat-ML characterizes as peatlands, such as Colombia's Rio Guaviare. This river may be too dynamic to allow extensive peat formation due to relatively rapid meandering that would scour away peat-forming depressions faster than the organic matter can accumulate or else bury potential peat with mineral sediments from the Andes . Given the lack of an appropriate predictor for these hydro-geomorphological processes operating on decadal to centennial timescales, it is not surprising that Peat-ML may overestimate peat extent in these ecosystems. Other areas, like Colombia's Amazon catchment region, might be susceptible to similar processes as these regions are suggested to be floodplain forests in ; however, their map is based on the CORINE Land Cover data for Colombia . Other areas in Colombia where Peat-ML predicts peatlands include parts of the Orinoco catchment region, where shows flooded grassland savannas and riparian wetlands, and the Caribbean catchment region, where peatlands are indicated by CORINE with other wetland types. Given that the CORINE land cover product is based upon remote sensing with little ground truthing, it is possible that several of these wetland regions shown in are actually peat-forming regions, making it difficult to definitively evaluate Peat-ML against this dataset. Besides the occasional small peatland area e.g. in the Páramo of Ecuador;, there are few sources of high-quality peatland mapping products for South America to evaluate Peat-ML against. While Peru has the PMFB mapped by and the Rio Madre de Dios by and is proposed to have extensive peatlands by and Peat-ML, there is presently no national peatland inventory .

Peat-ML predicts more peatland extent than PEATMAP in Central America (Fig. ). Much of the predicted peatlands are close to coastlines, in particular along the Atlantic coasts of Mexico, Nicaragua, Costa Rica, and Cuba. Peat-ML places more extensive peatlands on the Yucatán Peninsula of Mexico, which is not evident in PEATMAP. A desk-based assessment of peatlands based upon cartographic approaches with solicited expert assessment shows similar distributions of peatland extent, but with less peatlands in the Yucatán . The Yucatán peninsula has relatively extensive marsh and mangrove coastal wetlands but is a karstic landscape with a highly permeable carbonate substrate suggesting Peat-ML is overestimating peat extent for the inland portions of the peninsula.

African peatlands (Fig. ) are also poorly mapped, making it difficult to evaluate the Peat-ML results. There are notable differences between Peat-ML and PEATMAP. PEATMAP shows very few peatlands outside the Congo's Cuvette Centrale, whereas Peat-ML has relatively extensive peatlands in South Sudan along the border of the Central African Republic and Chad. This is in general agreement with more qualitative African peatland extent estimates and demonstrates Peat-ML's ability to reasonably determine peatland extents in regions where reliable spatially explicit mapping products are absent. Regardless, Peat-ML may still be underestimating African peatlands due to a lack of appropriate training data. An example is the newly documented peatlands in the Okavango Delta , which have a dominantly herbaceous vegetation cover (sedges, papyrus, grasses), while our only training dataset for Africa is a swamp forest (Cuvette Centrale). Future iterations of Peat-ML may profit from some active mapping campaigns presently underway in East Africa (Alexandra Barthalmes, personal communication, 2021) that could provide much needed training data and thereby improve predictions for the peatland regions of Africa. Improving understanding of African peatland extents will likely remain challenging; however, due to land use pressures that may complicate peatland identification and mapping as suggested by , African peatlands are heavily utilized by rural populations that depend on the peatland's water and organic soils for crop cultivation.

While much of the Indonesian Archipelago contains training data for the ML algorithm, the neighbouring states of Papua New Guinea, Brunei, and Malaysia are entirely model-predicted areas (Fig. ). While Malaysian peatlands appear similar between Peat-ML and PEATMAP, Papua New Guinea is quite different. PEATMAP shows extensive peatlands in the central mountainous region of the country, while Peat-ML has the peatlands placed in the surrounding lowland regions. There is some indication that the mountainous regions should have extensive peatlands . These peatland complexes appear to be sufficiently different from the Peat-ML training data that the ML model is unable to predict them.