Integrated machine learning and GIS-based bathtub models to assess the future flood risk in the Kapuas River Delta, Indonesia

Climate change has been accelerating at an alarming rate in the last century and is likely to continue in the future (IPCC 2012). Certain regions will experience more intense and frequent rainfalls (Marengo et al. 2020), which will increase the flooding risk. In contrast, other regions will experience decreasing rainfall and increasing evaporation, which will accelerate soil's progressive drying, leading to drought (Mukherjee et al. 2018). In coastal areas, climate change impacts the frequency and intensity of coastal flooding. The coastal flooding hazard arises through changes in mean sea level and the storminess of the atmosphere that creates storm surges in the first place (Lilai et al. 2016; Vousdoukas et al. 2016). Extreme events, such as the 100-year flood under the current climate, with future sea-level projections, will occur much more frequently.

Urban areas in deltas are particularly vulnerable to climate change (Ridha et al. 2022). These urban areas will face multiple ocean and land threats simultaneously. With the areas' rapid population growth (Bhatta 2010), disaster could impact more people and cause more damage. This risk highlights the importance of preparedness in disaster mitigation (Dinh et al. 2012; Chan et al. 2021). Communities in disaster-prone cities should properly assess the hazards and adapt to the changing climate, by for instance mapping their flood hazards and building flood defense systems. Adaptation strategies should be informed by proper risk projections so they can effectively mitigate the impacts (Hallegatte 2009).

The Kapuas River Delta (KRD) is a low-lying marshy delta on the west coast of Borneo Island, Indonesia. Silt deposits cover the delta from the coastline up to about 60 km inland (MacKinnon et al. 1996), creating estuarine floodplains for the Kapuas river downstream. Several urban areas lie in this delta and are prone to floods, which could be exacerbated by a combination of storm surges, intense rainfalls, and high discharges. These hazards are further intensified by sea-level rise caused by climate change (Moftakhari et al. 2017).

The water levels within the low-lying KRD are influenced by tide and wind surges from Karimata Strait, as well as the river discharges from the Kapuas and the Landak rivers (Sampurno et al. 2022). The flood risk here, and in other deltas, is likely to increase in future climate (Kundzewicz et al. 2014). The rising sea levels and excessive rainfall in the coming years will cause a severe impact because the KRD is low-lying and densely populated. Flood will be more intensive and frequent in a delta such as this (Lange 2020). A forecasting system is needed to assist local water management in assessing potential flood hazards, mitigating the risks, and planning adequate measures (Ngo et al. 2018).

Flood forecasting can leverage water-level modeling based on a machine learning (ML) technique (Ruslan et al. 2014; Gallien 2016; Habert et al. 2016; Noymanee & Theeramunkong 2019; Nguyen & Chen 2020). ML is based on evidence of relationships manifested in records of input and output data without analyzing the internal structure of the physical process. Using only historical data, an ML model can represent a complex input and output relationship, such as the relationship between water level and its predictor variables (Bishop 2006). The knowledge can subsequently be used to predict future water levels.

This study aims to assess the flood risk of an urban area within the KRD in future climates. We used an ML algorithm, i.e., multiple linear regression (MLR), to predict the water level in the city of Pontianak, the most populated urban area in the KRD. However, since data was scarce, we selected three months of data containing many extreme events (26 flood events) in 2020 as a reference case. The predictors we set comprised sea surface elevation in the river mouth, river discharge, wind velocity, and rainfall. The predicted water levels were used as a proxy to estimate the flood hazards in the city. Next, using a GIS-based bathtub model, we created flooding hazard maps under a 100-year flood scenario for the city with three climate change plots. Lastly, we conducted a risk analysis of these hazards to the nearby infrastructures (buildings and roads). The assessment output is essential for the local water resource managers in Pontianak to mitigate the impacts and create future adaptation strategies.

Since the study area is located in a tidal river, where water levels are influenced by ocean tides, rivers discharges, and weather conditions, we used seven predictors: the sea surface elevation (SSE); the weather variables measured in Pontianak, i.e., precipitation, average wind speed, maximum instantaneous wind speed, and average wind direction; discharges from Kapuas River retrieved at about 50 km upstream from the observation point; and discharges from Landak River retrieved at about 18 km upstream from the observation point (Table 1). Before training the model, we split the dataset between 80% for training and 20% for testing purposes.

Table 1

Independent variables used as water level predictors

Code .	Variable acronym .	Description .	Source .
	SSE	Sea surface elevation retrieved at the river mouth (m)	PUSRIKEL KKP (https://pusriskel.litbang.kkp.go.id/)
	Qkapuas	Hourly discharge of the Kapuas River (m3/s)	Global Flood Monitoring System (Wu et al. 2014)
	Qlandak	Hourly discharge of the Landak River (m3/s)	Global Flood Monitoring System
	RR	Hourly precipitation in Pontianak (mm)	PMMS
	WSavg	Hourly average wind speed in Pontianak (m/s)	PMMS
	WSmax	Hourly maximum instantaneous wind speed in Pontianak (m/s)	PMMS
	WD	Hourly average wind direction in Pontianak (degree, in the range: 0–360)	PMMS

Code .	Variable acronym .	Description .	Source .
	SSE	Sea surface elevation retrieved at the river mouth (m)	PUSRIKEL KKP (https://pusriskel.litbang.kkp.go.id/)
	Qkapuas	Hourly discharge of the Kapuas River (m3/s)	Global Flood Monitoring System (Wu et al. 2014)
	Qlandak	Hourly discharge of the Landak River (m3/s)	Global Flood Monitoring System
	RR	Hourly precipitation in Pontianak (mm)	PMMS
	WSavg	Hourly average wind speed in Pontianak (m/s)	PMMS
	WSmax	Hourly maximum instantaneous wind speed in Pontianak (m/s)	PMMS
	WD	Hourly average wind direction in Pontianak (degree, in the range: 0–360)	PMMS

View Large

Furthermore, we used a GIS-based bathtub model to create a flood extent map associated with the predicted water levels in the study area. Using the bathtub model, we assume that all areas within a model domain will be inundated if their elevation is less than the river's projected water level (Yunus et al. 2016). However, since we only focus on floods connected to the river, we consider the hydrology connectivity between the river streams and the land area (Van de Sande et al. 2012). Any area within the city will be inundated if a channel connects it to the Kapuas River stream, i.e., city canals. Therefore, using an algorithm created by Wang & Liu (2007), we identify and fill surface depressions in our DEM map that are not connected to the river.

Since we employed the MLR algorithm, we also checked the linear regression assumptions between the dependent variables and the predictors to diagnose the model adequacy. Then, we selected the variables that significantly impacted the model and omitted the non-essential ones. To do so, we produced a scatter plot between variables, created histograms of each variable, and calculated the Pearson's correlation coefficients for all pairs of predictor and dependent variables. We also checked the correlation among predictors to detect and avoid multicollinearity. We used the psych library in R to check the linear assumption (Psych: Procedures for Psychological Psychometric and Personality Research 2021). To enhance the prediction skill of the model, we also run a sensitivity analysis before predicting the future flood scenarios impacted by climate change. By doing so, we could evaluate the effect of each parameter on the model's output (Chu 1999). Therefore, we could tune in the model parameters involved in the model-building to improve the prediction results by adjusting only the sensitive parameters (Cacuci 2003). To conduct the sensitivity analysis, we used the konfound R library (Xu et al. 2019).

Two goodness-of-fit coefficients determine the models' performance: the Root Mean Square Errors (RMSE) and the Nash–Sutcliffe Efficiency (NSE). RMSE is commonly used for regression tasks to measure the accuracy of a predicted variable against an observed variable over an entire dataset (Jackson et al. 2019). However, the coefficient was only computed for the maximum values in the dataset to measure how well the model captured the inundation hazards. Meanwhile, NSE was used to assess the ‘skill’ of the ML models compared to the skill of the observed data's mean to predict an unknown dependent variable (Choi et al. 2020).

To evaluate the impacts of climate change, we used the projection of sea-level rises, precipitation changes, and surface wind changes in three climate scenarios for Southeast Asia's regional level (Iturbide et al. 2021). The scenarios are the low (RCP2.6), the medium (RCP4.5), and the high emission scenarios (RCP8.5). Based on these projections, we created projection datasets (SSE, Precipitation, Wind average, and Wind maximum) for each year into the future (from 2021 to 2100). We then re-ran the ML model to predict future water levels and extract each year's annual maximum water level. Then, we computed the flood frequency for every ten years of data using the Gumbel (1958) distribution curve. Here, we took only the flood frequencies between 2020 and 2030 (as the current state) and between 2090 and 2100 (as the future state). Based on the flood frequency curves, we estimated the 100-year flood level as the annual maximum flood hazard levels in 2020 (current) and 2100 (future hazard state).

Next, based on the elevations in the corrected DEM, we defined wet areas as any points on the map with an elevation lower than the 100-year flood level. Wet areas mean either inundated areas (with an elevation greater than zero) or standing water areas (with an elevation less than zero). We repeated this procedure for all future scenarios.

Next, we conducted a flood risk assessment for further analysis. Here, we identified the infrastructures (buildings and roads) over the study area, retrieved from OpenStreetMap (OpenStreetMap 2020), that may be affected by inundation hazards under the 100-year flood condition in 2100. We analyzed the flood impact on infrastructures using a QGIS plugin: InaSAFE (InaSAFE 2022).

Since the area of interest is low-lying land with a generally low slope (≤ 8%) (Arianti et al. 2020) and in almost a natural state (with no dams, dykes, or levees), a 2.64-m flood level (1% annual exceedance probability) already causes an inundation in a significantly large part of the study area (Table 3). The overflow water runs freely into the city through the drainage canals. Therefore, an increase of 100-year flood level of 28 cm, in a low emission scenario (RCP2.6), already affects infrastructures severely (6.4% increase in impacted buildings and 6.8% increase in impacted road lengths). In the medium emission scenario (RCP4.5), while the 100-year flood level will increase by about 39 cm in 2100, the inundation will cause 8.1% more impacted buildings and 8.4% more affected road lengths. Lastly, in the highest emission scenario (RCP8.5), when the 100-year flood level increases by 70 cm from the current state, there will be 11.9% more impacted buildings and 12.7% more affected roads.

Here, we qualitatively categorized the flood risk for each sub-area (district) as high and low regarding the spatial distribution of flood hazards and exposed infrastructures, i.e., buildings and roads (Environment Agency UK 2013). Figure 7 shows that the high-risk area is located in the eastern and western parts of the city. Most of these areas will be inundated under the 100-year flood level condition. Therefore, we can say that the flood hazard in these areas is high. The exposure is also high because there are many infrastructures in these areas, which means they are highly populated or function as, for example, a business center. The high hazards and dense infrastructures made this area a high-risk flood zone.

Meanwhile, the northern part of the city has a lower infrastructure density. The flood hazard in this area is also low due to its higher altitude. A significant portion of this area will not be inundated under all 100-year flood scenarios. Therefore, the flood risk of this zone is categorized as low (Environment Agency UK 2013). Another non-inundated area is the southern part of the joining point of the Landak River and the Kapuas Kecil River, as well as some parts of the city's southwestern area. Even though these areas have dense infrastructures, we still classified the flood risk of these zones as low.

Nevertheless, this study has some limitations. Firstly, we do not account for potential collapse of the Antarctic ice sheet in future scenarios (van de Wal et al. 2019). Secondly, the flood analysis is limited to the extent of the city of Pontianak. Hazards and exposure areas outside the city are not included in the assessment. Next, the hazards are classified only in two states: inundated and not inundated, without considering the depth of the flood (Islam & Sado 2000). In addition, the impact assessment on the infrastructure (buildings and roads) only depends on the DEM map (FEMA 2003). The analysis result may differ due to local conditions, such as higher local terrain maps and infrastructure types. Lastly, it is also essential to notice that the MLR model is trained by only three months of data in 2020. The water levels could be higher if we considered other flood events in other years.

Despite the limitations, the local water management or government can use this study's results to mitigate future flood events in the study area. The model's evaluation of climate change impacts can guide the adaptation strategies, such as whether or not it is urgent to adjust the height of flood defense structures along the riverbanks within the city. Next, they can track the hazardous area throughout the city and watch what happens as the water comes from all possible causes. Moreover, other deltas with similar characteristics and limitations can adopt the approach to assess their future flood risk.

This study successfully assesses future flood risk in the KRD, particularly in Pontianak, using integrated ML and GIS-based bathtub inundation models. We simulated the water level dynamics and quantified the flood frequency curve of the current and future states as modulated by climate change. We created flood maps with potentially inundated areas in 100-year flood (1% annual exceedance probability) under the current and future scenarios. We found that the 100-year flood level in the study area will increase from the current 2.64 m to 2.92 m, 3.03 m, and 3.34 m in 2100 under each future climate scenario (RCP2.6, RCP4.5, and RCP8.5), respectively. These increases correlate to the increment of flood hazard areas over the region of interest. We found that in 2100 more buildings will be exposed (increased by about 6.4%–11.9%), and more roads will be impacted (increased by approximately 6.8%–12.7%) depending on the climate scenario. This assessment benefits the local water managers in preparing adequate mitigation strategies and the city's disaster management plan.

This work was funded by the Indonesia Endowment Fund for Education (LPDP) under Grant No. 201712220212183.

The relevant data and code for this paper are available at: https://zenodo.org/record/7370291

The authors declare there is no conflict.