Climate change-induced peatland drying in Southeast Asia

When organic peat soils are sufficiently dry, they become flammable. In Southeast Asian peatlands, widespread deforestation and associated drainage create dry conditions that, when coupled with El Niño-driven drought, result in catastrophic fire events that release large amounts of carbon and deadly smoke to the atmosphere. While the effects of anthropogenic degradation on peat moisture and fire risk have been extensively demonstrated, climate change impacts to peat flammability are poorly understood. These impacts are likely to be mediated primarily through changes in soil moisture. Here, we used neural networks (trained on data from the NASA Soil Moisture Active Passive satellite) to model soil moisture as a function of climate, degradation, and location. The neural networks were forced with regional climate model projections for 1985–2005 and 2040–2060 climate under RCP8.5 forcing to predict changes in soil moisture. We find that reduced precipitation and increased evaporative demand will lead to median soil moisture decreases about half as strong as those observed during recent El Niño droughts in 2015 and 2019. Based on previous studies, such reductions may be expected to accelerate peat carbon emissions. Our results also suggest that soil moisture in degraded areas with less tree cover may be more sensitive to climate change than in other land use types, motivating urgent peatland restoration. Climate change may play an important role in future soil moisture regimes and by extension, future peat fire in Southeast Asian peatlands.


Introduction
Peatlands in Insular Southeast Asia contain globally significant carbon stores, estimated at 67 GtC (Page et al 2011, Warren et al 2017. This carbon is maintained through high water tables that prevent peat oxidation or ignition (Hirano et al 2009, Dommain et al 2010. However, in the last half a century, degradation has threatened these carbon stores, as only ∼6% of peat forests remain in pristine condition (Miettinen et al 2016) (Miettinen et al 2017, Andrew andPeters 2021).
Climate also affects peatland carbon loss. During drought years, large-scale burning of peatlands (Van Der Werf et al 2008, Field et al 2016, Taufik et al 2017 also leads to globally significant carbon emissions because dry peat is more flammable. For example, fires associated with the 1997 El Niño Southern Oscillation led to an estimated 0.81-2.56 GtC emitted, 13%-40% of global mean annual fossil fuel emissions at the time (Page et al 2002). Although fire has been a phenomenon in Southeast Asian peatlands for at least 30 000 years (Goldammer et al 1989, Anshari et al 2001, the frequency and scale of these fires has increased dramatically in recent decades (Page and Hooijer 2016). In the second half of the 20th century, periodic droughts only led to large increases in fire during periods when degradation rates were high (Field et al 2009). This evidence suggests that the combined effects of degradation and climate on the soil moisture and groundwater levels in peatlands mediate peat fire (Taufik et al 2017, Dadap et al 2019. Specifically, degradation can worsen the sensitivity of tropical peatland emissions to meteorological drought (Siegert et al 2001), further motivating restoration and conservation efforts (Jaenicke et al 2010, Leifeld and Menichetti 2018, Goldstein et al 2020. Given that fire emissions in Southeast Asian peatlands have historically been largest during drought conditions attributable to El Niño Southern Oscillation and the Indian Ocean Dipole (Van Der Werf et al 2008), future emissions may also be influenced by long-term trends associated with climate change (Li et al 2007). Regional climate simulations have shown that average rainfall will likely decrease in Southeast Asia in future decades (Li et al 2007, Tangang et al 2020, especially during the dry season (Kang et al 2019). Additionally, changes in solar radiation, atmospheric humidity, and temperature may also affect the peat water balance. Understanding how future climate will affect peat vulnerability is necessary to inform management, restoration, and conservations efforts. However, the sensitivity of peatland moisture to climate change is likely highly variable across the region. Several factors influence how different hydroclimatological conditions affect peat moisture including the initial distribution of water table depth, water uptake differences between vegetation types (Hirano et al 2015, Manoli et al 2018, canal properties including their depth, width, and spatial pattern (Page et al 2009, Cobb et al 2020, Dadap et al 2021, microtopography, hydraulic properties of the peat and its macropores (Mezbahuddin et al 2015, Baird et al 2017, Cobb et al 2017, and bulk density (Sinclair et al 2020). Because the distribution of these factors across the region is poorly understood and highly uncertain, it is not feasible to parameterize physical hydrologic models (or using land surface simulations from existing regional climate models (RCMs)) to understand how climate change affects peat moisture across this region.
Here, we instead used observations and a statistical modeling approach to estimate how climate change will influence peat hydrological conditions in the coming decades. In particular, we considered surface soil moisture, which has previously been shown to be closely related to peat fire risk (Dadap et al 2019) and for which observations are widely available across Southeast Asian peatlands using data from the Soil Moisture Active Passive (SMAP) satellite (Entekhabi et al 2010, McColl et al 2017. Here we use the term 'soil moisture' to describe peat moisture content in the upper ground surface, but note that due to the high organic matter in peatlands, this moisture is not contained in the soil matrix in the same way it is in mineral soils. In tropical peatlands, surface soil moisture is closely connected to water table depth (Hirano et al 2014, Dadap et al 2019, the most commonly used metric of peat moisture levels for fire risk studies (e.g. Wösten et al 2008, Hooijer et al 2012. However, future soil moisture regimes are unknown. To our knowledge, only one study has attempted to model future soil moisture (Li et al 2007), but the coarse resolution (>2 degree) of the general circulation models used in that study cannot account for the complex topography and land-ocean-atmosphere interactions associated with this region, nor for the effects of variations in land use and peat properties. In this study, we instead used machine learning to build a statistical model that predicts soil moisture variations across the region as a function of several climate factors. The statistical model was then used to analyze the impact of climate change on soil moisture across the region, including its spatial distribution and variation with land use type.

Approach
Our general approach in this study was to train statistical models (neural networks) to learn relationships between climate, degradation, location, and soil moisture in Southeast Asian peatlands under present climate. Neural networks have been shown to be a viable and in some cases superior alternative to stateof-the-art models when forecasting hydrologic variables in data scarce regions (e.g. Hsu et al 1995, Kratzert et al 2019. The trained neural networks were then used with projections of future climate to predict future soil moisture. This approach is illustrated in figure 1. Such a climate sensitivity approach has been used previously to understand features of hydrologic projections (Short Gianotti et al 2020).
Here, we directly predict simplified soil moisture statistics to avoid the need for explicit simulation of soil moisture timeseries in the future. These variables were: (a) mean dry season soil moisture (sm dry season ) and (b) percent low soil moisture (pct low sm ), defined here as the percent of time in a given year that the soil moisture is below 0.2 cm 3 cm −3 . For mean soil moisture, we focus on the dry season only because that is more closely tied to fire risk. Here, we assumed that dry season timing will remain the same in the future period. Previous work using both laboratory measurements (Frandsen 1997, Huang et al 2015) and SMAP soil moisture (Dadap et al Figure 1. Overview schematic of the soil moisture modeling approach. Squares denote input data while ovals denote neural network predictions. The model is first trained on ERA5 climate and SMAP soil moisture data. Predictions are then calculated for reference  and future (2040-2060) time periods using climate data from a RCM forced by three global circulation models. Input climate data are bias-corrected to ERA5 reanalysis data using quantile mapping. 2019, figure 3) showed that peat ignition probability (at laboratory scale) and burned area (at remote sensing scales) sharply increase when soil moisture is below a threshold value of about 0.2 cm 3 cm −3 . Thus, the pct low sm statistic represents the fraction of a given year when the peat is at high fire risk and captures the non-linear response of fire to soil moisture.

Study area
This study focused on peatlands in Insular Southeast Asia, an area spanning ∼157 000 km 2 on Sumatra, Borneo, and Peninsular Malaysia. All analyses were limited to pixels covered by at least 50% peatlands, as determined from 30 m land cover maps (Miettinen et al 2016), and were performed on the 9 km EASE-Grid resolution of the SMAP data (Brodzik et al 2012).

Data sources
2.3.1. Soil moisture data Soil moisture data from SMAP are available every 2-3 d at 9 km resolution from 2015 to present. Each pixel represents a distinct soil moisture observation. An example SMAP soil moisture timeseries from one pixel is shown in Supplementary figure 1. We used soil moisture retrieved from the Multi-Temporal Dual Channel Algorithm (MT-DCA) (Konings et al 2016, Feldman et al 2021, which is a separate dataset from the SMAP baseline science data products. Because the MT-DCA retrievals rely on a dielectric mixing model that was developed for mineral soils (Mironov et al 2004), an empirical correction was applied to account for the high organic matter content of the peat (Bircher et al 2016). Measurements with potentially high error associated with radio frequency interference, urban areas, and precipitation were excluded from the dataset. Microtopography and the presence of organic material on the peat may add error to the soil moisture retrievals, as the presence of litter can affect L-band soil moisture retrievals even in less densely vegetation sites (Kurum et al 2012). Thick vegetation can also block remote sensing measurement of soil moisture where present. Furthermore, no in situ validation of SMAP data has been performed in this region, which remains a limitation of using SMAP data in this context. However, there is evidence that soil moisture retrievals have sufficient accuracy in this region, since triple collocation-based (statistical) error analysis of SMAP soil moisture in the region previously showed that retrieval precision is likely on par with the SMAP mission target error of 0.04 cm 3 cm −3 (Dadap et al 2019).

Input features
Input features were chosen to capture the possible effects of climate, degradation, and location on soil moisture (Supplementary table 1). Climate variables included precipitation and potential evapotranspiration (PET) to represent water supply and evaporative demand; PET was calculated from radiation and temperature using the Priestly-Taylor method. These were represented in the neural networks with mean dry season PET, mean dry season precipitation, mean annual precipitation and precipitation entropy. Precipitation entropy (calculated as the Shannon entropy of monthly precipitation) was included because it is a descriptor of rainfall seasonality (Feng et al 2013), or the degree to which rainfall is distributed between the wet and dry seasons. A smaller entropy value indicates larger seasonal differences in precipitation. Although PET might deviate from actual evapotranspiration (ET), only PET was included here since the RCM and reanalysis data may not capture the differences in water use strategies (and thus, the actual/potential ET ratio) in different land use types.
Because the study area is dominated by coastal areas and topographic complexity, a high resolution simulation is necessary for more accurate prediction of climate variables (Im and Eltahir 2018). Here, we used 25 km regional climate data from the Coordinated Regional Climate Downscaling Experiment-Common Regional Experiment as inputs to the neural networks for the reference   Peatland degradation features used in the neural network model included the percent of different land use types, tree cover fraction, drainage canal density, fire area, and fire count. These factors are likely to change significantly in the future, but it is difficult to predict how they will change due to shifting economic incentives and regulations (Suwarno et al 2018, Schoneveld et al 2019, Humpenöder et al 2020. We therefore only considered changes in climate variables in this study, but incorporated these additional land use and fire inputs to account for their effect on the soil moisture-climate relationship. Location descriptors including latitude, longitude, region, and distance from the edge of the peat dome were also used as predictors to account for possible spatial autocorrelated factors affecting soil moisture, such as land use history, peat physical properties, and land management practices. See Supplementary Text 1 and Supplementary table 1 for more information on the input features and neural network structure.

Neural network prediction of soil moisture
The neural networks were trained using remotely sensed soil moisture from SMAP over the 2015-2020 period. To determine how soil moisture statistics were affected by climate change, the neural networks were then run with a set of regional climate predictions dynamically downscaled from three global climate predictions for a reference  and future time period . To reduce the effect of biases in the global circulation models downscaled by a RCM, all climate inputs were bias-corrected to match the statistics of an observation-driven dataset, here the European Centre for Medium-Range Weather Forecasts ERA5 reanalysis product (Hersbach et al 2019).
We compared predictions of sm dry season and pct low sm between the reference  and future periods . In each case, degradation and location input features were held constant while climate features changed based on bias-corrected RCM predictions. Bias correction of the climate data was necessary because there are biases between the RCM simulations and the pseudo-observational ERA5 data. These differences in distributions would otherwise result in projections of soil moisture incorrectly attributed to changing climate that are instead due to differences between ERA5 and the RCM. We used quantile mapping to correct these biases (Reichle andKoster 2004, Miao et al 2016). Specifically, we matched reference period RCM data to ERA5 data from the same time period, and then applied the same correction to future period RCM data. A separate quantile mapping was applied to each of the three RCM realizations (corresponding to each global circulation model). Both RCM and ERA5 data used for bias-correction were downscaled to 9 km resolution from their original 25 and 30 km grids, respectively, using nearest neighbor resampling.

Neural network models assessment
The neural networks' performances were evaluated in two different ways using cross-validation (CV). First, to assess overall model performance on unseen data, 5-fold 'random' cross validation was performed. This means that a model was trained on a random selection of 80% of the data, then predictions on the unseen 20% of the data were compared to observations. The training data and testing data were cycled through till all data had been tested in this manner. Alternatively, to assess the models' abilities in predicting interannual variability, 6-fold 'temporal' cross validation was performed, meaning that the model was trained on 5 years of data, then tested on the remaining 6th year of data. Prediction accuracy was then assessed using the coefficient of determination (R 2 ) (which varies from 0 to 1 with 1 indicating higher agreement between the model prediction and observation), bias, and root-mean-squared error.

Results and discussion
3.1. Soil moisture models assessment Cross validation for both soil moisture variables, sm dry season and pct low sm , demonstrated that the neural network models could predict out-ofsample data accurately (table 1, Supplementary figure 2). The sm dry season model achieved a CV mean R 2 = 0.83, RMSE = 0.08 cm 3 cm −3 , and a bias of 0.001 cm 3 cm −3 on randomly sampled test data. Similarly, the pct low sm model achieved a CV mean R 2 = 0.73, RMSE = 16%, and a bias of 0.8% on random test data. When the two networks were crossvalidated using a full year's worth of held-out data, R 2 decreased only a slight amount (∆R 2 ≈ 0.1 in both cases), suggesting the networks were able to predict soil moisture behavior on unseen years of data, including simulated future years.

RCM predicts drier future atmospheric conditions
RCM projections show overall drying in the study region, as dry season precipitation is projected to decrease across 89% of the area (figure 2(a)), while PET is projected to increase across 98% ( figure 2(b)). The median change in dry season precipitation is −0.79 mm d −1 and the median PET change is +0.38 mm d −1 between the reference  and future (2040-2060) periods (Supplementary figure 3(a)). Geographically, there are larger decreases in dry season precipitation in southern Sumatra and larger increases in dry season PET in the southern parts of the study region (figure 2). Because ET is the dominant water flux out of peatlands (e.g. Hirano et al 2015, Cobb and Harvey 2019), increased PET is expected to lead to decreases in soil moisture.
Annual precipitation is projected to decrease by ∼0.5-2 mm d −1 in the study region (figure 2(c), Supplementary figure 3(b)). Precipitation seasonality, as captured by precipitation entropy, exhibited a mixed change in signal by latitude in Sumatra: generally decreasing south of the equator and increasing north of it (figure 2(d), Supplementary figure  3(b)). Decreasing entropy suggests higher seasonality, which may cause drier sm dry season , as precipitation may be less evenly distributed between the dry and wet seasons. These results are consistent with those of Kang et al (2019), who found that Aug-Oct precipitation (corresponding to the dry season across most of the study area) generally decreased while Nov-Jan precipitation generally increased. While our model did not account for possible changes in the timing of the dry season, only relatively minor changes are projected in the timing of the monsoon in this region (Ashfaq et al 2020). Overall distributions of climate features shifted under future climate (Supplementary figure 3), but these shifts generally did not extend far beyond the ranges observed under future climate. This builds confidence that the neural networks trained using present climate-soil moisture relationships can accurately assess the impact of future climate scenarios.

Climate changes cause substantially drier soils
Both soil moisture variables exhibited drier conditions under 2040-2060 climate projections compared to 1985-2005 climate, consistent with the changes in climate forcing. Median sm dry season was projected to decrease during the future period by 0.023 cm 3 cm −3 (figures 3(a) and (c)). For context, this decrease is nearly half the magnitude of the 0.056 cm 3 cm −3 decrease in median dry season soil moisture observed by SMAP during the 2015 and 2019 El Niño years relative to non-El Niño years between 2015 and 2020. Recent El Niño years have been associated with a nonlinear increase in fire activity (Yin et al 2016), suggesting that the magnitude of climate-change induced soil moisture drying, absent other changes, could significantly increase fire risk in the region. However, the impacts of climate change relative to recent El Niño years differ geographically. Here, we found that the predicted soil drying due to climate change is generally greater than impacts observed during recent El Niño droughts north of the equator, while the opposite is true south of the equator in the study region (figures 4(a) and (b)).
The pct lowsm variable, a more direct measure of fire risk than sm dry season , increases over almost the entire region. Our neural network projected a median increase in pct low sm of 3% (from 12.5% to 15.5%) (figures 3(b) and (d)), suggesting that extremely dry conditions associated with high fire risk will be more prevalent in the future. To estimate how large the pct low sm associated impact on burned area might be, we consider a single average burned area associated with dry soil moisture (below 0.2 cm 3 cm −3 ) and another average burned area for wet soil moisture conditions (as calculated from the curve in figure 3(a) of Dadap et al 2019). The increase of the 3% in pct low sm would then correspond to a 10% increase in burned area due to future climate change. This calculation, though highly simplified, illustrates the outsized increase in fire risk associated with even small increases in pct low sm driven by climate change.
Drought conditions during recent El Niño years have been attributed primarily to precipitation drought (e.g. Field et al 2016), but our model suggests that future changes in sm dry season are also affected by increased evaporative demand (i.e. increasing PET). This is evident from the higher feature importance of PET compared to precipitation inputs for both neural networks (Supplementary figure 4). Table 1. Cross-validation ('CV') results ± standard deviation across folds. Temporal CV was performed by holding out one year of data at a time for the test set, and training on the other years. For example, the data would be trained on 2015-2019 data and evaluated on unseen 2020 data. This was then repeated for all six years of data. Random CV involved random selection of data from all years (across all pixel-times) when performing five-fold cross validation.

Model
Random  Consistent with this finding, running the model with future (2040-2060) PET but with reference  precipitation resulted in a decrease in median sm dry season that was 0.008 cm 3 cm −3 , or 36% of the change when precipitation drivers were included. Thus, our results suggest that increased evaporative demand will play a significant role in driving soil moisture changes under climate changes. Landatmosphere feedbacks may further exacerbate soil drought and atmospheric aridity under future climate (Zhou et al 2019).

Degraded areas are more sensitive to climate change
To better understand where soil moisture changes will occur, we separated model predictions by land use (here determined by the majority land use type in each pixel). During the reference period , pristine forest was predicted to have the wettest median sm dry season , while open undeveloped was the driest (figure 4(a)). Nevertheless, reference period distributions of sm dry season were generally found to have little variation across land uses ( figure 4(a)). This was somewhat surprising, as land use is often used as a proxy for hydrologic disturbance (e.g. Miettinen et al 2017, Taufik et al 2020. However, our model predictions were mostly consistent with a meta-analysis of in situ soil moisture measurements, which show similar soil moisture magnitudes across land use types and large variation within land uses (Supplementary figure 5, Supplementary table 2). Such high variability of soil moisture within land use types is likely due to differences in precipitation regimes, peat physical properties, drainage density, and more Kurnianto In particular, open undeveloped areas are predicted to experience the largest changes, while pristine forests are predicted to experience the smallest changes. Open undeveloped areas generally have the lowest starting soil moistures, suggesting that the driest areas will dry further than wetter areas. The differences in soil moisture changes by land use type could be caused by (a) climate changing more in certain land  use types and/or (b) certain land use types are inherently more sensitive to changes in climate. However, the former does not appear to be a major factor, because the soil moisture changes (∆sm dry season and ∆pct low sm ) vary independently of the changes in climate variables (∆precip and ∆PET) when grouped by land use type (figure 6), except for increases in PET with decreases in sm dry season . This suggests that land use could affect the sensitivity of soil moisture response to climate change.
Our results further suggest that tree cover affects soil moisture sensitivity to climate change. We regressed ∆sm dry season and ∆pct low sm with the input metrics that capture peatland degradation (tree cover, canal density, and fire), and found significant relationships for both variables only with tree cover (Supplementary figure 6). These relationships suggest that areas with less tree cover are more sensitive to climate changes (i.e. will experience more drying) than areas with more tree cover. This increased sensitivity with   and future (2040-2060) periods. Box denotes inter-quartile range and median. Change in median (c) sm dry season and (d) pct low sm from reference to future periods. Figure 6. Magnitude of percent change in soil moisture variables (sm dry season and pct low sm ) compared to percent change in climate variables (dry season PET and dry season precipitation). Changes in soil moisture do not appear to vary with changes in climate. Note the signs for sm dry season and for dry season PET denote negative change.
less tree cover can be explained by a number of possible mechanisms. First, tree cover reduces the solar radiation reaching the ground surface. In areas with less or shorter vegetation, this effect is minimized, and atmospheric conditions are more likely to determine changes in soil evaporation (Fan et al 2019, Ohkubo et al 2021). Deforested areas are also more likely to contain degraded soils with increased hydrophobicity (Bechtold et al 2018, Perdana et al 2018. This in turn could decrease rainfall infiltration, increase soil evaporation, and decrease the capillary connection with the water table and the surface soil, making degraded areas more sensitive to climate changes. Furthermore It should also be noted that SMAP soil moisture measurement could be affected by differences in peat microtopography by land use type, complicating comparisons of soil moisture between land use types. For example, the duff and litter layers that form the hummock and hollow topography endemic to pristine peatlands are often replaced by a denser, flatter surface when graded or converted to agricultural use (Lim et al 2012). These differences could in turn affect the profile of soil moisture measurement relative to the groundwater table. For example, Sakabe et al (2018) found high variability in surface soil moisture within pristine forests based on the location of measurement: hummocks averaged 0.06 cm 3 cm −3 while hollows averaged 0.54 cm 3 cm −3 , but the drier value would not necessarily imply higher fire risk. Such small-scale spatial variability would be averaged to a single measurement by SMAP, which integrates measurements over 9 km pixels. However, this variability would not exist in land use types where the ground surface is generally flatter. Thus, in situ validation studies are needed to better understand how to interpret differences in SMAP retrievals between land use types and their implications for fire risk and carbon emissions. Nonetheless, comparisons within land use types would not be affected by this potential issue, and the predicted drying trends observed in all land use types underscores the consistent prediction of drying due to climate change.

Conclusions
Our model projections suggest that future drier climatic conditions across Southeast Asia will lead to lower mean soil moisture and more frequent periods with dangerously dry peat conditions that would lead to increased fire risk. The median predicted decreases in soil moisture are nearly half the magnitude of those experienced during high-fire drought years associated with El Niño under current climate, portending more prevalent fire risk due to climate change. More research is needed to understand the impact of changes in El Niño severity or changes in dry season length, two factors that were not considered in this study. In contrast to recent droughts, future drier soil conditions also appear to be driven by increased evaporative demand in addition to reduced precipitation. Further work is needed to assess the combined and interacting impacts of changing land use-which will mediate how evaporative demand changes will affect future ET and thus ultimate soil conditions-and changing climate, thus requiring the development of detailed land use change scenarios. Our findings suggest that more degraded peatlands with lower tree cover may be especially sensitive to climate change, motivating the importance of restoration in not only reducing current carbon emissions and fire risk, but also towards lessening the impacts from future climate change. Degradation is understood to be a critical determinant of peatland hydrology, but our results suggest that climate change will also play an important role in determining future soil moisture regimes.

Data availability statement
The data that support the findings of this study are openly available at the following URL/DOI: https://github.com/ndadap/future-sm-peatlands, doi: 10.5281/zenodo.6740137.