A recent study by Chen et al (2020) compared multiple global gridded population datasets for their applicability in regional studies and their accuracy w.r.t statistical data across cities. Results indicated that the population values within the gridded population of the world (GPW) version 4.0 dataset and Global Human Settlement Layer (GHSL) dataset are closest to the statistical value. Also, the absolute value for the relative error of total population is the smallest for GPW v4.0 being within 3% compared to WorldPop and History Database of the Global Environment version 3.2 (HYDE) datasets (Chen et al., 2020). In this study. We used the (GPW) version 4.0 population dataset containing a raster grid of 30 arc-second spatial resolution. This dataset contains gridded population count and population density, as well as the corresponding U.N adjusted values, for years 2000, 2005, 2010, 2015, and 2020 (Doxsey-Whitfield et al., 2015). Following recent international guidelines (Brandmüller et al., 2020), we delineated the urban clusters from the gridded population datasets for the year 2015. An urban cluster (or moderate density clusters) consists of contiguous grid cells with a density of at least 300 inhabitants per km² and has a population of at least 5,000 in the cluster.

We obtained 956 urban clusters in the SEA region based on 2015 population data. The clusters are distributed among all the 11 countries in the SEA region (Figure 1A). Each of the 11 SEA countries had between 1 (Singapore) and 93 (Indonesia) clusters, except the Philippines that had a very high number of urban clusters (659, Supplementary Figure S1A). This may be due to the topographic structure of the nation and due to considering many small islands. Around 56% of the total number of clusters range within 10 and 50 km² (Supplementary Figure S1B). There are also around 37 clusters whose size is greater than 1,000 km².

The primary LULC dataset used in this study was created by downscaling land-use change data from the Land Use Harmonization 2 (LUH2) Integrated Assessment Model datasets (Hurtt et al., 2020). The LUH2 data are reported at 30 km resolution, which is too coarse for our study, so we used downscaled versions of the LUH2 data that provided 10 arc-second (approximately 300 m at the equator) resolution LULC maps for each SSP scenario using the Spatial Economic Allocation Landscape Simulator (SEALS) model, documented in Johnson et al. (2020); Suh et al. (2020); Johnson et al. (2021). The LULC dataset we created is one of the few publicly available high-resolution products, allowing a finer understanding of LULC transition of seven major classes (water, forest, grassland, cropland, urban, barren and non-forested vegetation). Using these downscaled maps, we computed LULC transition graphs between 2015 and 2050 for each scenario. Further, a comparison analysis was performed with alternative urban scenario (SSP) global datasets to compare the accuracy of the dataset. The datasets used for comparison are the Global National Total Amounts of Urban Land, SSPs and Base Year, v1 (2000–2100) (Gao and O’Neill, 2020) and Global projections of future urban land expansion under SSPs (Chen et al., 2020). The data values in the Gao and O’Neill, 2020 dataset were only available between 2000 and 2100 for every 10-year interval. So, we first interpolated the dataset by country between 2000 and 2100. We then extracted the 2015 value via the best fit trendline between the datapoints (between 2000 and 2,100–10-year interval) Comparison was performed at the SEA level although all three models are developed at the global scale. Note that the (Gao and O’Neill, 2020) dataset had projections (SSP scenarios) for year 2015. However, for Chen et al. (2020) and Johnson et al. (2020), the base data year was 2015. Also, of note the results for Singapore have not been included in the following sections considering the non-comprehensive LULC data coverage for Singapore urban cluster.

For the analyses, we used InVEST Carbon storage and sequestration model (standard and “forest edge” versions, v3.9.0), which represents a stock estimation (Sharp et al., 2021). GIS analysis, data organisation, and visualisation were conducted in ArcGIS 10.8 (ESRI). To obtain carbon pool values for aboveground and belowground biomass, we used the CDIAC (Carbon Dioxide Information Analysis Center) global database of consistent and spatially explicit estimates of vegetation carbon stocks, circa 2000, following the IPCC methodology (Ruesch, 2008). CDIAC also provides a global ecofloristic zones shapefile which was integrated into the LULC datasets in our study. These zones were integrated via spatial processing methods by which the existing forest land cover category was enhanced by deriving the sub forest types in this region.

To obtain soil carbon pool values, we used the Global Soil Organic Carbon (GSOC) map, the first global soil organic carbon map ever produced via a consultative and participatory process involving member countries (FAO, 2018). The LULC data with respective classes were overlaid with the GSOCmap to extract the mean SOC values for each class which were used in the analysis. The interface between agro soil-traits and the weather conditions could influence the growth of local species of plants for carbon storage/sequestration potential at plant species scale. However, this study is addressed at regional scale (300 m spatial resolution) and at land cover level so that we incorporated the Ecofloristic zones classification within SEA region to identify and understand the biomass properties of different types of forest cover from the CDIAC carbon pool database.

The biophysical table derived from these datasets, required for the InVEST model, is provided in Supplementary Table S1. Due to the unavailability of consistent data for dead carbon pools which mainly include litter as well as lying and standing dead wood, they have not been considered for this study.

We performed an uncertainty assessment for the carbon sequestration analysis for SEA urban clusters. For this, we first compared the results from the two carbon storage models (basic and forest edge versions of the InVEST model), and then extracted the available carbon pool data from another global database ForC (Anderson-Teixeira et al., 2018). ForC is the largest database of ground-based measurements of forest ecosystem carbon stocks and annual fluxes. However, the carbon pool data were only available for tropical dry forests, tropical moist deciduous forests and tropical shrubland for SEA urban clusters. Using these carbon pool data for the new database, a modified biophysical table was generated and the InVEST carbon model was run to extract the new carbon storage and sequestration values for SEA urban clusters. The new results were used for uncertainty assessment through comparison with the CDIAC carbon pool database.

Carbon sequestration or losses were compared to the CO₂ emissions from the Emissions Database for Global Atmospheric Research inventory database v5.0 (Crippa et al., 2019). The emission values for the year 2018 were extracted for all the urban clusters in the region. Zonal statistics were applied to extract the total emission per cluster using the 0.1⁰ resolution gridded CO₂ emissions dataset. For those few urban clusters smaller than the grid cell resolution, the emission value of the grid cell was normalized with the cluster area within the grid cell to extract the value for the urban cluster.

Urban cluster areas extracted for the SEA region in the adopted model (Johnson et al., 2020; Figure 1A) were close, ranging between “2% in SSP3 to 50% in SSP5”, to the (Gao and O’Neill, (2020)) reported model values (Supplementary Table S2). The rate of increase between 2015 and respective 2050 SSP scenarios were very similar between the two land use change models, within (+/−) 25% difference, except for the SSP1 scenario (71%, Supplementary Table S2). In general, the urban areas reported by the Chen et al. (2020) model were consistently lower than the Johnson et al. (2020) and the (Gao and O’Neill, (2020)) models, with a similar pattern observed for the rate of increase (%). This analysis confirms the discrepancies in the global urban areas projected under the SSP’s among various global datasets considering the underlying modelling frameworks and base data sources used in each of the model. Also, the Johnson et al (2020) dataset one of the few publicly available high-resolution products, allowing a finer understanding of LULC transition of seven major classes while both the Chen et al. (2020) and the (Gao and O’Neill, (2020)) datasets only consider urban land cover class.

The land-use and land-cover change components of the baseline SSP scenarios cover a broad range of possible futures based on the respective scenario narrative (Riahi et al., 2017). Our study shows that SSP1, SSP4 RCP 6.0, and SSP5 scenarios have higher urban expansion in urban clusters compared to SSP2, SSP3, and SSP4 RCP 3.4 (Figure 2). These observed urbanization patterns were similar to the global projected urbanization rates for SSP scenarios as reported by (Riahi et al., 2017). The study by Riahi et al. (2017) also reports that in SSP3 scenario, urbanization is constrained by slow economic growth. Similarly, our study observed lowest urban growth between 2015 and 2050 for SSP3 RCP 7.0. Most scenarios except SSP2 and SSP4 RCP 3.4 showed a reduction in cropland area between 2015 and 2050 in urban clusters. For forest areas, all scenarios except SSP1 show a reduction between 2015 and 2050 with highest in SSP4 RCP 3.4 (Figure 2). The observed forest increase in the SSP1 RCP 2.6 scenario in our study is consistent with the narrative of this scenario depicting a gradual, global-scale, and pervasive expansion of forests due to little pressure on land resources via low population projections and high agricultural productivity (Riahi et al., 2017). These patterns are observed within the Bangkok urban cluster and illustrated in Figure 1B–H. Based on these results and land cover transition patterns in Figure 2, we propose an urban scenario interpretation for SEA urban clusters:

a. SSP1 RCP 2.6—Urban sprawl and reforestation, where urban growth is highest amongst all scenarios and the only scenario with increase in forest, and reduction in cropland.

Except for SSP1 RCP 2.6 (Urban sprawl and reforestation), all scenarios show that urban clusters will have emitted carbon between 2015 and 2050 as shown in Figure 3. SSP1 RCP 2.6 (Urban sprawl and reforestation) shows carbon sequestration mainly attributed to the increase in forest area between 2015 and 2050 indicating around 0.01 Gigatonnes of potential carbon to be sequestered. SSP2 (Compact urbanization with moderate agriculture) and SSP4 RCP 3.4 (Compact urbanization with extensive agriculture) show higher carbon emitted compared to other three scenarios—attributed to having higher decrease in forest area compared to others (Figure 3). Considering the urban sprawl and decline in forests projected, this region could see a carbon loss of up to 0.11 Gigatonnes (SSP4 RCP 3.4).

Our uncertainty assessment reveals that the overall difference between the carbon values derived using the two carbon pools databases (CDIAC and For-C) is very minimal ranging from −0.28% for SSP1 RCP 2.6 to −2.11% for SSP5 RCP 8.5 (Figure 3). We also observed that the overall difference between the carbon values derived using the two InVEST carbon models (Carbon model and Forest carbon edge effect) ranges from −28% for SSP5 RCP 8.5 to −47% for SSP1 RCP 2.6.

Carbon change values normalized by urban cluster area (km²) help compare carbon sequestration and loss values across all urban clusters. The spread of the distribution varies across scenarios (Figure 4). SSP1 RCP 2.6 (Urban sprawl and reforestation) shows the least spread due to its emphasis on increase in forests. This scenario minimizes carbon losses for the majority of urban clusters in SEA. Higher spread of carbon change values is observed in SSP4 RCP 3.4 and SSP2 RCP 4.5 scenarios, which correspond to the highest median carbon losses as well.

Interestingly, there are clusters emitting carbon and clusters sequestering carbon in each scenario. Therefore, we further examined the distribution of urban clusters according to their change in carbon values between 2015 and 2050 across scenarios. We grouped them into four classes representing: 1) Clusters that lost carbon under all scenarios; 2) clusters that sequestered carbon under all scenarios; 3) clusters that have not experienced any change under any scenario; and 4) clusters that either lost or gained carbon depending on the scenario. The analysis aimed to identify clusters which are more likely to sequester carbon or lose carbon by 2050, irrespective of the SSP scenarios. We found that 9% of the urban clusters (85 clusters) are likely to sequester carbon and 46% (438 clusters) are likely to lose carbon in every scenario (Figure 5). 40% (382 clusters) have mixed projections; and 5% (51 clusters) are projected not to change under any scenario. The Philippines and Timor Leste show a majority of urban clusters that are likely to lose carbon (Figure 5). There are also urban clusters sequestering carbon under every scenario for most countries except Brunei, Laos, and Timor Leste.

Comparing carbon change values to urban emissions, we found that the average offset value ranging from −2% (Scenario SSP1 RCP 2.6) to −21% (Scenario SSP4 RCP 3.4, Supplementary Figure S2). We also grouped urban clusters by size: small (0–100 km²), medium (100–500 km²) and large (>500 km²), to examine whether these offset values were driven by the urban extent of the city (Figure 6). Indeed, small urban clusters have higher spread of offset % except for SSP 1 scenario (Urban sprawl and reforestation) compared to medium and large clusters. This is an indication that the LULC changes in the smaller clusters depending on their major land cover influences the magnitude of the offset in the SEA region across the SSP’s. SSP 1, due to reforestation, is the only scenario with a positive average carbon offset for large clusters. Overall, Figure 6 indicates that majority of urban clusters in SEA, irrespective of their sizes, are still contributing to carbon emissions for all scenarios except SSP1 RCP 2.6. Even with the reforestation scenario, the majority of urban clusters will only offset 1% of their carbon emissions by the scenarios we modelled (results not shown).

Finally, we also investigated the historical trend of land cover change between 2010 and 2015. 360 urban clusters out of 956 have sequestered carbon between 2010 and 2015. Most of these urban clusters are projected to lose carbon in 2050 for all the SSP scenarios (Comparison of Offset % of these clusters with the total urban clusters is shown in Supplementary Figure S2).