Rising floodwaters: mapping impacts and perceptions of flooding in Indonesian Borneo

Borneo is the world's third largest island, located in South-East Asia, encompassing regions of three countries: Indonesia, Malaysia and Brunei Darussalam (4.242° S–6.965° N, 108.883° E–118.990° E). Borneo's native ecosystems host some of the highest levels of biodiversity globally and include various (and often threatened) vegetation types such as mangroves, lowland dipterocarp tropical forests, peat swamp forests and montane rainforests [30]. These forested systems not only harbour some of the world's most iconic species such as the Bornean orangutan (Pongo pygmaeus) [31] and proboscis monkey (Nasalis larvatus) [32], but also provide essential and varied ecosystem services to some 25 million people across Borneo [26, 33]. Current rates of deforestation and forest degradation, however, are among the highest in the world [34–36], thereby threatening such services. Mean annual rainfall is generally high yet variable [range: 1520–4820 mm; 37]. Heavy rainfalls can lead to severe flooding, especially in lowland and coastal areas [38], which are often targeted for conversion to oil-palm agriculture [36, 39]. Our study focused on riverine and flash flooding in Kalimantan (i.e. Indonesian Borneo), which covers 76% of the island's total area.

We delineated rivers and watersheds in Borneo using a hydrologically corrected Digital Elevation Model at 3 arc-second resolution, and ArcHydro 2.0 tools in ArcGIS 10.2 [40] (see S1—supplementary methods, S1.1). We defined 'major rivers' by a minimum drainage area of 200 km², and the finer network of 'all rivers' by a minimum drainage area of 20 km². We defined watersheds as areas draining to the ocean, and delineated subwatersheds as the area draining to each link of the major rivers. For any point of interest, the 'relative watershed' consists of the surrounding subwatershed, combined with any upstream subwatersheds (i.e., all areas that contribute to flow through the focal subwatershed).

We surveyed village leaders' perceptions of flooding in 364 villages across Kalimantan (figure 1, and figure S1), through interviews with the village head or other official. These interviews were conducted during April–October 2009 (341 villages) and April–October 2012 (23 villages), and were drawn from a larger survey on villagers' perceptions of forests and wildlife, in villages selected at random within 10 km of forests in the approximate range of the Bornean orangutan (i.e. sampling was random with respect to past or present flooding). The larger survey comprised two sets of interviews: (1) the village-level interviews, which we analyse; and (2) interviews on villagers' individual perceptions with 7–12 respondents per village, which we do not analyse, and which formed the basis for previous studies of wildlife and ecosystem services [26, 33].

Zoom In Zoom Out Reset image size

Our study selected the 364 villages within Kalimantan, for which village-level interviews gave detailed responses to questions on flooding. A 'flood' was defined as either a riverine flood or flash flood, in which floodwaters covered the village's main road or path at the centre of the village. We coded responses regarding (1) the frequency of flooding over the past five years (n = 302 villages; coded as 5 classes representing frequencies of 0, 0.1, 0.5, 1 and 2+ floods per year) and (2) trends in flood frequency over the past 30 years (n = 260; responses to the question 'has the frequency of floods declined, stayed the same, or increased over the past 30 years?' were coded as: decline, no change, or increase in flood frequency). For the flood trends, a 'decline' was reported in only 2 of 260 villages (S2.1), so we excluded this class, and modelled flood trends as presence/absence data, i.e. whether flood frequency increased (presence) or showed no change (absence). Further details of the interview methods, quality assessments, and coding are given in S1.2.

We obtained flooding reports from the online archives of six news agencies in Kalimantan, covering 16 regional newspapers, using the search keyword 'banjir' (flood). We collated data from all articles that reported flooding in Kalimantan over 3 years, 20 April 2010–29 April 2013, and that gave estimates of flood height and/or flood impacts (see S1.3). We georeferenced flooding locations using Google Earth and census administrative boundaries. We estimated flood impacts for each flooded settlement as the number of people affected, and number of houses flooded. If a numeric value was not reported directly for a settlement, we assigned low, median and high estimates (based on the median, 10th and 80th percentiles from the statistical distributions of the number of houses and number of people affected per flooded settlement), enabling calculation of summed impacts across all the separate flood events (table S1).

We performed spatial analyses separately for each of the three datasets: (1) village flood frequency from village interviews over the past five years (n = 302 villages); (2) village flooding trends over the past 30 years (n = 260 villages); and (3) news reports of flood events (n = 380 settlements, i.e. villages, towns or urban areas) (figure S1). We analysed each dataset in relation to 35 spatial predictor variables ('landscape variables'—see S1.6) to assess statistical relationships between flooding and LULC in each watershed (20 LULC classes), soil types and water storage capacities, topography (slope, elevation), climate (temperature and rainfall seasonality and monthly maximum), infrastructure (impervious surface area, and road densities), and socio-economic factors (population density in 2011, and proportions of the population by religion and ethnic group). Initial models included all spatial predictors; final models included those contributing >1% of explained variance.

We performed statistical modelling using Boosted Regression Tree (BRT) models with five-fold cross-validation [41] in the R statistical environment [42, 43]. BRT methods combine many regression trees to form an ensemble model, providing a flexible and robust method for both (1) quantifying relationships and contributions (including nonlinear relationships with the response, and interactions among predictors), and (2) generating predictions for new areas or datasets [41]. Further details of model fitting and assumptions are given in S1.4. We used the resulting models to generate predictions for all populated areas across Kalimantan (S1.4). Model performance was assessed via deviance reductions, cross-fold correlations and estimates of classification error (detailed in S1.4, S2.2).

In 58% of villages (176/302), respondents reported flooding frequencies of one or more times every year over the past 5 years (S2—supplementary results, table S2). Only 10% of villages reported no experience of flooding. Approximately 20% of all villages reported that floods had become more frequent over the past 30 years (table S2; as mentioned in methods 2.3, only two villages or 0.8% reported declines in frequency, and were excluded from modelling of flood trends). The association between flood trends over 30 years, and current frequency over 5 years, was positive but not tight (figure S2); i.e. villages that perceived an increase in flood frequency, reported current frequencies that still span a wide range, from occasional to more than once per year. Current flood frequencies were generally consistent for villages on the same stretch of a river, and either similar or lower for villages further upstream. For example, when looking at the villages with annual or more frequent floods, other villages further upstream (>3.2 km) usually reported similar or lower flooding frequencies (mean difference between frequencies −0.24 floods per year, range 1 to −2; figure S3).

Over the period 20 April 2010–29 April 2013 (3 years), 414 news items reported at least 142 distinct flood events affecting 966 settlements in Kalimantan. Quantifying the total impacts over the 142 flood events between 2010 and 2013 (table S7), we estimate that 197 000 to 360 000 houses were flooded (based on the median and 80th percentile of the distribution of houses flooded per event, respectively, table S1), directly displacing 776 000 people (possibly as many as 1.5 million, based on the 80th percentile method).

Flood events on average affected five settlements, and 60 settlements were flooded in two or more of the events. Flooding was reported in 32 major urban areas and 836 other settlements. Observed locations of flooding based on newspaper reports (2010–2013) and village interviews (2008–2011) showed strong spatial agreement. Almost all interviews from villages close to sites of newspaper-reported floods (within 4 km overland or 60 km downstream of villages) reported experiences of moderate to high flood frequencies (every two years or more frequent).

The most frequently quantified impact of flooding reported was the number of houses flooded, and the number of people directly affected. Other impacts were frequently described, but not quantified, for example 'residents could not work', 'schools were closed', or croplands, plantations, schools, businesses or health facilities were flooded. Financial impacts were often mentioned, but rarely enumerated. Government budget allocations, however, are clearly high: for example the Government of East Kalimantan allocated 605 billion Rupiah (c. US$50 million) for flood mitigation in one city alone from 2011–2013, primarily for four systems of structural defences along the Mahakam River [44].

The village interviews and news-reported floods provide two distinct lines of evidence for the existence of frequent and widespread flooding, and for examining the relationships between flooding and landscape variables. BRT models showed very high performance for all datasets (table S3), and key predictors are shown for each analysis in figure 2. The influence of each class of landscape variable (LULC, climate, topography etc) is summarised in table 1. The influence of each landscape variable is computed from substantial cross-validation, giving a robust estimate of the variable's contribution to the model's predictive goodness-of-fit. Interactions among predictors are naturally incorporated in the tree structure of the models.

Zoom In Zoom Out Reset image size

LULC variables accounted for 24.0%–29.8% of the total variance in each dataset (corresponding to 28.4%–50.7% of the variance explained by each model). We note that some of the effect of soil or edaphic changes will be represented here by LULC variables, due to the nature of peatlands and wetlands as both landcovers and hydroecosystems.

Predictions for settlements across Kalimantan are mapped for village flooding frequency in figure 3, village flooding trends in figure 4, and news-reported floods in figure 5.

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

LULC variables consist of distance and area metrics for 20 LULC classes. Hydrography consists of distance to the nearest river systems, and watershed extents. Topography consists of slope, elevation, and distance to coastline. Population and infrastructure consist of population densities (2011), road densities, and urban cover (2010). Climate refers to long-term mean precipitation and temperature variables (1950–2000). Soil variables consist of soil family (oxisol, histosol etc) and changes in soil saturated water content under non-native LULC.

As expected, the frequency of recent flooding (in the past five years) was lower in villages located far from rivers and at higher elevations (figure S4). Both the recent frequency, and the likelihood of a long-term increase in frequency (trend over the past 30 years), were lower for villages near to natural wetlands and when natural wetlands covered a large proportion of the watershed (figure 2, S4, S10). A long-term increase in flood frequency was more likely for areas <200 km from the coastline or on flatter slopes, and in areas with lower long-term average rainfall of the wettest month (less than 400 mm; possibly reflecting lower historical flood frequencies that made any increases in frequency more noticeable), or with higher rainfall seasonality (figure S10).

The frequency of recent floods and the probability of flooding trends over the past 30 years were both larger for villages closer to mines (figure 2), and all 31 villages in watersheds with >0.5% cover of mines were predicted to experience flooding at least every 2 years (figure S6). Flood frequencies were also higher for villages close to other open or disturbed landcovers (e.g. bare or grassed areas and open peatlands).

Recent flood frequencies were slightly lower in watersheds with larger areas of peatland upstream (figure S5), while villages located very close to peatlands were more likely to have experienced an increase in flood frequency over the past 30 years (figure 2). This difference between villages in watersheds with upstream peatlands, versus villages within or near peatlands themselves, is consistent with peatlands having a role in regulating the floodwaters that reach downstream areas, even if flooding of peatlands themselves has increased in some areas over the past 30 years.

For flooding trends, an increase in flood frequency over the past 30 years was less likely in watersheds with greater cover of logged or intact forests, and more likely in watersheds with more extensive oil palm plantations (figure S11). Recent flood frequencies were lower in watersheds with higher cover of logged forests or agroforest/regrowth cover (figure S7). Regarding distance to logged forests, villages very close to the nearest logged forests (<5 km) were predicted to have higher mean recent flood frequencies (figure 2), but were no more likely to have experienced long term increases in frequency. The association with recent frequencies could arise, for example, if logged forests that occurred in areas that naturally flood frequently, were more likely to be left standing (and not converted to other land uses), than forests in less frequently flooded areas.

Of the news-reported flood events, strong positive relationships were identified with the distance from rivers (especially major rivers) and urban areas, higher impervious cover and lower forest cover (figure 2, table S9). Flooding probability increased sharply as impervious surface cover increased from 0%–3% (measured at subwatershed level), without further increases beyond this threshold. Floods were also more likely in watersheds with higher oil palm cover, or at larger distances from intact forest or agroforest/regrowth (table S9). In relation to wetlands, news-reported floods were less likely in subwatersheds with more wetlands (>4% cover), and floods were slightly less likely in watersheds with greater peat soil areas. Despite strong relationships with urban and impervious cover, the news reports did not appear to be biased to only reporting floods in densely populated areas. This is based on three observations, (i) the reports spanned the full range of population densities, and 71% of flooded settlements were rural villages or towns (not cities); (ii) reports also covered areas 220 km from the coastline (figure S14); (iii) predicted flooding was more likely within 50 km of an urban area, but showed no further decline with distance.

Population variables (2011 population density in the local area or the watershed) contributed very little to any of the BRT models. Population effects per se can be a concern in studies of flooding trends because of the risk that perceived trends are due to changes in flood detection (exposure and observation), rather than flood occurrence. However, we conclude that village leaders' perceptions of trends in flood frequency over recent decades are largely independent of population per se, and flooding trends relate more to biophysical changes in the landscape that are captured by other variables. This conclusion is based on our definition of a village flood event (flooding of the main road or path, which is less sensitive to population size than other definitions such as a number of houses flooded), the very small effects attributable to current population size, and the lack of detectible effects of population growth at district level (see supplementary results 2.1).

Comparing models based on the two independent data sources (village interviews and news reports), we note strong similarities in the relationships estimated between flooding and landscape variables, and strong spatial concordance in those areas at low or high flooding hazards (figure 6, S16). Regarding relationships with landscape variables, some variables differ in the relative strengths of their influence between the analyses (i.e. % of explained variance), for example, the news dataset spans a larger range of population densities and impervious cover, giving greater power to detect their influences. However, even where the percentage of variance differs, the relationships show similar functional forms in both analyses (figure 2 and supplementary figures S4, S10 and S14).

Zoom In Zoom Out Reset image size

Comparing predictions of village flood frequencies versus news-reported floods (figure 6), we see strong spatial concordance in the areas predicted by both datasets to have high flooding hazards (dark reds), or low flooding hazards (pale greens). News-reported floods were very rarely predicted to occur in areas where the village models indicated low flood frequencies (pale reds). The main difference in predictions occurs in areas where village interviews predicted high flood frequencies, but news-reported floods were predicted to be absent (dark greens). These areas are usually further from major rivers or urban centres, and may reflect lower coverage by news reports of areas that are more remote, or of floods that are very regular or affect very small populations. Further detailed comparisons are given in S2.6.

Comparing predictions of village flood trends, we found that news-reported floods were more likely in areas that were also predicted to have experienced a trend in village flooding (i.e. where village flood frequencies were predicted to have increased over the past 30 yr, table S13). Spatial concordance is marked (figure S16), though not as strong as for village flood frequencies versus news reports (figure 6).

Government knowledge of flood events in Borneo appears to be fragmented and highly incomplete. While national and regional governments on Borneo acknowledge that floods represent major economic and social costs [64, 65], quantitative data and assessments are lacking, especially for regional areas (S2.7). Indonesia's National Action Plan for Disaster Management 2010–12, for example, does not include a detailed flood risk assessment [66]. We compared our estimates from news-reported floods with records from Indonesia's publicly accessible online Disaster Loss Database (DiBi—Data dan informasi Bencana Indonesia) for the same time period. This national database is a model of data transparency and accessibility, however the data it contains is highly incomplete, and dramatically underestimates the number of large flooding events and their impacts (tables S14, S15). For example, the single largest event in our newspaper survey was reported by the regional South Kalimantan Disaster Management Agency to have flooded 24 889 houses in 299 villages along the Martapura River in April 2010 (table S14). The same disaster is recorded in the national database as affecting 929 people, and causing 5 deaths. Overall, the number of people reported in the DiBi database to be impacted by flooding is 10–100 times smaller than that estimated from the newspaper records, for all provinces except Central Kalimantan (table S15).

Indonesia's national methods for flood reporting and risk assessment are geared to prioritisation of emergency response to intensive disasters, i.e., high impact events that affect large, spatially concentrated populations, such as those in the capital Jakarta. However, they have been based on brief and incomplete time series of events, and need to better integrate the monitoring efforts of local and regional agencies (S2.7). Governments do not adequately assess or act on large cumulative risks from events that affect more spatially distributed populations, or occur at high frequency, and add up to large impacts over a region or time period. We also note that the existing method for risk assessment only considers areas with major settlements or agricultural production, and for which ≥1 major event appear in the (incomplete) national database, giving a map with no predicted hazard or risk level for the majority of the land area of Kalimantan (figure S18). The importance of disasters of lower magnitude but high frequency has been highlighted in recent studies in Indonesia, Bolivia, Mexico, Mozambique, Nepal, the Philippines, and Vietnam showing negative impacts on children's education, health, and access to services such as water and sanitation [18]. Note that we do not advocate replacing disaster reporting mechanisms with news surveys. These comparisons indicate the need to support consistent recording at the local level (and maintain existing efforts by agencies such as the South Kalimantan Disaster Management Agency). Secondly, they indicate the scope for better communication of flood events and impacts among local and Provincial authorities, and incorporation into national databases.