Integrated hydrodynamic and machine learning models for compound flooding prediction in a data-scarce estuarine delta

. Flood forecasting based on water level modeling is an essential non-structural measure against compound flooding 10 over the globe. With its vulnerability increased under climate change, every coastal area became urgently needs a water level model for better flood risk management. Unfortunately, for local water management agencies in developing countries building such a model is challenging due to the limited computational resources and the scarcity of observational data. Here, we attempt to solve the issue by proposing an integrated hydrodynamic and machine learning approach to predict compound flooding in those areas. As a case study, this integrated approach is implemented in Pontianak, the densest coastal urban area over the 15 Kapuas River delta, Indonesia. Firstly, we built a hydrodynamic model to simulate several compound flooding scenarios, and the outputs are then used to train the machine learning model. To obtain a robust machine learning model, we consider three machine learning algorithms, i.e., Random Forest, Multi Linear Regression, and Support Vector Machine. The results show that this integrated scheme is successfully working. The Random Forest performs as the most accurate algorithm to predict flooding hazards in the study area, with RMSE = 0.11 m compared to SVM (RMSE = 0.18 m) and MLR (RMSE = 0.19 m). 20 The machine-learning


Introduction
Compound flooding in low-gradient coastal area is a recognized hazard that can be exacerbated by global warming (Hao and 25 Singh, 2020). Compound flooding hazard is derived by the interaction of storm surge penetration, riverine flooding, and intense rainfall over the area (as the impact of extreme meteorological events) that coincide or nearly coincide (Bilskie and Hagen, 2018). This natural hazard can endanger the population and the coastal area's infrastructures, which have been growing fast in the last decade. Without an appropriate mitigation, the consequences of the hazard can be severe for the coastal environment (Costabile et al., 2013) and the coastal communities both economically (Karamouz et al., 2014) and socially (Comer et al.,30 Commented [JS1]: We updated the abstract as follows: Flood forecasting based on hydrodynamic modeling is an essential non-structural measure against compound flooding over the globe. With the risk increasing under climate change, all coastal areas are now in need of flood risk management strategies. Unfortunately, for local water management agencies in developing countries, building such a model is challenging due to the limited computational resources and the scarcity of observational data. We attempt to solve this issue by proposing an integrated hydrodynamic and machine learning approach to predict water level dynamics as a proxy of compound flooding risk in a data-scarce delta. As a case study, this integrated approach is implemented in Pontianak, the densest coastal urban area over the Kapuas River delta, Indonesia. Firstly, we built a hydrodynamic model to simulate several compound flooding scenarios. The outputs are then used to train the machine learning (ML) model. To obtain a robust machine learning model, we consider three machine learning algorithms, i.e., Random Forest, Multi Linear Regression, and Support Vector Machine. Our results show that the integrated scheme works well. The Random Forest (RF) is the most accurate algorithm to model water level dynamics in the study area. Meanwhile, the machine-learning model with the RF algorithm can predict eleven out of seventeen compound flooding events during the implementation phase. It could be concluded that RF is the most appropriate algorithm to build a reliable ML model capable of estimating the river water level dynamics within Pontianak, whose output can be used as a proxy for predicting compound flooding events in the city.

Commented [JS3]:
We added the references here: Compound flooding hazard is derived from the interaction of storm surge penetration, riverine flooding, and intense rainfall over the areas (as the impact of extreme meteorological events) that coincide or nearly coincide (Bilskie and Hagen, 2018;Ikeuchi et al., 2017;Wahl et al., 2015).

Study area 65
The Kapuas River is the longest island's river in Indonesia (Goltenboth et al., 2006). Its basin is located in the western part of Borneo Island (Fig. 1). The delta and the estuary of the river are still mostly natural. There are no dams, dykes, or groins on its stream. The river is vital for the local communities as a source of freshwater and transportation. The population along the Kapuas River banks is sparse, except in some parts of the estuary. The most populated area is Pontianak, located along the main branch of the Kapuas River, namely the Kapuas Kecil River. Therefore, the hydrodynamics of the river significantly 70 influences the occurrence of floods in the city. During the last decades, the expansion of palm oil cultivation and forest fires happened massively in the Kapuas water catchment (Semedi, 2014;Jadmiko et al., 2017). These phenomena led to a change in the Kapuas hydrological regime and triggered more intense flooding in the river's floodplains. Combined with global sea level rise these phenomena could lead to more intense and severe flood events in the future.

Hydrodynamic model description and model setup 75
We use the multi-scale hydrodynamic model SLIM 2D to simulate compound flooding for different forcing scenarios. Our model can simulate hydro processes along the land-sea continuum, from the river to the ocean (Vallaeys et al., 2018Frys et al., 2020;Le et al., 2020b). The model solves the 2D Shallow Water Equations (SWE): where is the water column height, is the horizontal gradient operator, U = ̅ is the horizontal transport, R is the rainfall, is the time, ̅ = ( , ) is the depth-averaged horizontal velocity, f is the Coriolis parameter, is the vertical unit vector pointing upward, is the horizontal eddy viscosity, is a constant to define a dry elements ( = 0) and wet elements ( = 1) (Le et al., 2020a), h is the bathymetry, = 9.81 m/s2 is the gravitational acceleration, is the bulk drag coefficient, is the wind stress and is the atmospheric pressure gradient. In order to run the hydrodynamic model, we define a 85 computational domain that covers both the river and the ocean part, create an unstructured mesh to cover the domain, with a resolution of 50 m over the riverbanks, 400 m over the coast near the river mouth, 1 km over the rest of the coastline, and 5 km over the offshore (Fig. 2), set bathymetry and bulkdrag coefficient, imposed some forcings, and define the boundary conditions. More detail about the hydrodynamic model setup can be found in our previous work (Sampurno et al., 2021).
The hydrodynamic model simulation are forced by wind and atmospheric pressure from ECMWF (Hersbach et al., 2020), and 90 tides from TPXO (Egbert and Erofeeva, 2002). As upstream boundary conditions, we imposed discharge from the Kapuas River and the Landak River. We also imposed runoff, which is obtained by converting rainfall over the Kapuas Kecil River catchment area as an inlet water flux at some channels entering the domain. The runoff of every channel was calculated from rainfall data using SWAT+ (Bieger et al., 2017), whose processes involved the pressure, the humidity, and other weather Commented [JS15]: We added more information in this paragraph, and split it into 3 new paragraphs: The Kapuas River is the longest inland river in Indonesia (Goltenboth et al., 2006). The basin is located in the western part of Borneo Island (Fig. 1). The water catchment area spreads over about 93,000 km 2 (about 12.5% of the Borneo Island area, Fig. 1), with about 66.7% of it consisting of forests (Wahyu et al., 2010). The upstream topography comprises hills covered mainly by Acrisol soils (Fig. 2), and the downstream consists of plains with more heterogeneous soil types (Fig. 2), such as Humic Gleysols (derived from grass or forest vegetation) and Dystric Fluvisols (young soil in alluvial deposits). The river is vital for the local communities as a source of fresh water and a transportation system. In the last decades, palm oil cultivation and forest fires expanded massively in the Kapuas water catchment (Semedi, 2014;Jadmiko et al., 2017). These circumstances changed the Kapuas hydrological regime and triggered more intense flooding in the river's floodplains. Combined with global sea-level rise, these phenomena could lead to more intense and severe flood events, particularly in the river delta. The delta of the Kapuas River is still mostly natural, with no dams, dykes, or groins on its downstream. Therefore, the hydrodynamics of the river significantly influences the flood occurrences in the delta. The most populated area over the delta is Pontianak, a city located in the Kapuas Kecil-the middle stream of the second-largest branch of the Kapuas River. As a tidal river, the tidal regime within the Kapuas River delta is mixed but mainly diurnal (Kästner, 2019). The dominant tidal constituent is K1, O1, P1, M2, and S2 (Pauta, 2018). The average tidal amplitude within the delta is set in a microtidal regime, with a mean spring range of 1.45 m at its river mouth (Kästner, 2019).

Commented [JS17]:
We modified the sentence and added references here: To simulate hydrodynamics within the Kapuas River delta, we use the multi-scale hydrodynamic model SLIM 2D (Lambrechts et al., 2008;Gourgue et al., 2009;Remacle and Lambrechts, 2016). point in Pontianak, and compared it with observational data to evaluate the model performance.

Flood scenarios
To simulate flood events, we considered 6000 scenarios with oceanic, atmospheric and river forcings (see Table 1). To produce the scenarios, we ran the hydrodynamic model for ten months. Then, we merged the output of all scenarios as a single dataset to train the machine learning model. Using these steps, we tried to encompass all possible flood events resulting from the any 100 combination of the external forcings.

Dependent and predictor variables
To develop machine learning models, we used the river water level at Pontianak as the dependent variable. Then, we considered atmospheric, oceanic, and riverine variables as the predictors (independent variables) of the water level in the city. Atmospheric 105 variables include average and maximum wind speed, wind direction, precipitation, and average atmospheric pressure. Oceanic variables include tide at the river mouth, while the riverine variables include the Kapuas River and the Landak River discharges.
To evaluate the impact of each predictor before the flood event, we also imposed the prior state (one and two hours before) of these parameters (see Table 2). The datasets were recorded hourly and combined the SLIM output (used in the training phase) and observational data (used in the testing phase). 110 atistic tool that can measure the degree of relatedness between variables in a dataset. The greater the value between two variables, the stronger its relatedness, regardless of how nonlinear its dependency (Kinney and Atwal, 2014). between two variables ( and ) is obtained from Choi et al., (2020): where ( , ) is the joint probability distribution.

Machine learning algorithm
Here, we consider three different machine learning algorithms, i.e., random forest (RF), multiple linear regression (MLR), and support vector machine (SVM). RF is a supervised learning algorithm that operates by constructing many decision trees at training time (Breiman, 2001). The algorithm can be implemented for classification or regression. To generate the ultimate 120 output, the model aggregates its multiple decision trees outcomes, which are called sub-sample outcomes (Han et al., 2012).
The technique was enhanced by combining bootstrap in its aggregating processes (Breiman, 2001). Using this strategy, the algorithm became an effective tool for classification and regression. In this study, the RF algorithm was obtained from the R randomForest library (Liaw and Wiener, 2002).

Commented [JS18]:
We modified and move these sentences to the next subsection:

Hydrodynamic model setup and calibration
Commented [JS19]: We changed this subsection:

Hydrodynamic model setup and calibration
Commented [JS20]: We modified this paragraph and added some new paragraphs here: In order to run the hydrodynamic model, we defined a computational domain that covers both the river and the ocean parts. Next, we generated an unstructured mesh to cover the domain, with a resolution of 50 m over the riverbanks, 400 m over the coast near the river mouth, 1 km over the rest of the coastline, and 5 km over the offshore (Fig. 3). The multi-scale mesh was generated using an algorithm developed by Remacle and Lambrechts (2018). Next, we set the bathymetry constructed from two data sets: first, the river and estuary bathymetry maps, obtained from the Indonesian Navy (Kästner, 2019), and second, the Karimata Strait bathymetry, obtained from BATNAS (BATimetri NASional, 2021). Furthermore, we set the bulk bottom drag coefficients, which are 2.5×10 -3 over the ocean (which corresponds to a sandy seabed) and 1.9×10 -2 over the river bed (Kästner et al., 2018). Lastly, we imposed the rainfall, as observed by the Pontianak Maritime Meteorological Station (PMMS). The hydrodynamic model simulation is forced by wind and atmospheric pressure from ECMWF (Hersbach et al., 2020), and tides from TPXO (Egbert and Erofeeva, 2002). As upstream boundary conditions, we imposed discharge from the Kapuas River and the Landak River. The discharge data were retrieved from the Global Flood Monitoring System (GFMS) (Wu et al., 2014). We also imposed runoff, which was obtained by converting rainfall over the Kapuas Kecil River catchment area as an inlet water flux at some channels entering the domain. The runoff of every channel was calculated from rainfall data using SWAT+ (Bieger et al., 2017), which considered the pressure, the humidity, and other weather parameter input. Unfortunately, during the tuning of the ...

Commented [JS22]:
We modified this sentence to: The datasets were recorded hourly and combined with the SLIM output (used in the training and testing phases) and the observational data (used in the implementation phase).

Commented [JS23]:
We modified this sentence: Mutual Information ( ), a statistic tool that can measure the degree of relatedness between variables in a dataset, was implemented to evaluate the relation between each predictor and the dependent variable (Fig. 6).

Commented [JS24]:
We added new paragraph after this one: All predictors considered in the machine learning model have an MI coefficient greater than zero, which means all predictor variables impact the river water level in Pontianak (Fig. 6). The relationship between these predictors and the water level could be linear or nonlinear (as shown by the MI capturing both relation types). Here, we found that the tidal elevations in the river mouth (X1, X2, and X3) ...

Commented [JS25]:
We added new sentences here: To obtain the optimal parameter for the RF, we first tune the algorithm by searching for the optimal value of the number of variables randomly sampled as candidates at each split (mtry). As a result, the optimal number is 16 (Fig. 7).
least-squared approach (James et al., 2013). In the least-squared approach, the best relationship model will be obtained by minimizing the sum of the squared distance between the calculated values (as model outputs) and the target values (James et al., 2013). This algorithm is the most straightforward approach in machine learning models and generally used as the baseline method. The MLR algorithm which was implemented in this study is obtained from the R RWeka library (Hornik et al., 2008).
SVM is an ML algorithm that works based on statistical learning frameworks (Gholami and Fakhari, 2017). This method is 130 robust for modeling a complex non-linear relationship. It amounts to choosing a kernel function that transforms the input features into a high-dimensional space to tackle the complexity. As a result, the process will transform the non-linear relationship of the input features to become linear. Finally, linear regression is carried out to obtain the ultimate output.
Compared with the other algorithm, SVM needs less computational resources due to its capability to be trained only by a few features (Gholami and Fakhari, 2017). Previously, SVM was only implemented for classification purposes, yet, it has also 135 been implemented for regression purposes after some enhancement. The SVM algorithm which we implement in this study is obtained from the R MARSSVRhybrid library (MARSSVRhybrid: MARS SVR Hybrid, 2021).

Metrics for machine learning model performance evaluation
Here, we use the Nash-Sutcliffe efficiency (NSE) measure to evaluate the models' performance. NSE is used to assess the performance of the machine learning models in producing the predicted water level. A perfect model corresponds to NSE = 1, 140 while a model that has the same predictive skill as the mean of the observed data represents by NSE = 0. If NSE<0 implies that the mean value of observed data is predicting better than the model does. The closer NSE value to 1, the better the predictive skill of the model. The NSE coefficient is calculated as follows: 145 where H represented the water level model at time , 0 represented the observed water level at the same time, and H 0 ̅̅̅̅ is the mean of observed water level.
Root Mean Square Errors (RMSE) of peaks between predicted water level and observation during the flood events are also used as additional performance indicator. RMSE is used to represent the model's ability to predict flood events. The RMSE between the model outputs and the observations is calculated by: 150 where is the water level as the model's output at the i-th peaks, and is the observed water level at the same time. is the number of the total peaks data.

Commented [JS26]:
We modified this sentence into: MLR is a statistical technique that uses several explanatory variables to predict the outcome of a response variable (James et al., 2013). This method fits the linear relationship between input features and the target (observed data) using the least-squared approach.

Commented [JS27]:
We added a new paragraph after this one: To obtain the best performance of the MLR algorithm, we did a statistical analysis to evaluate the multicollinearity among the predictor variables using the Variance Inflation Ratio (VIF). Since multicollinearity negatively affects the performance of the MLR model, VIF can help reduce the number of predictors (Alipour et al., 2020). Here, we found that some variables have VIF more significant than 5, which indicates a potentially severe correlation between these variables in the model (Fig. 8). Therefore, combined with the output of MI analysis, we removed some variables which have low MI and high VIF.

Commented [JS28]:
Modified to: SVM is a supervised machine learning algorithm based on statistical learning frameworks

Commented [JS29]:
We added a new paragraph after this one: Since kernel function is critical in SVM, we tuned the SVM algorithm to obtain good results by selecting the most appropriate kernel parameter. We tested four kernels, i.e., linear, polynomial, radial basis, and sigmoid, as the candidates. We found that the radial basis kernel performed the best for the SVM algorithm.
We ran a simulation for January 2019 and compared the simulated water elevation with observations in Pontianak. The model 155 errors correspond to an NSE of 0.87 and an RMSE of 0.12 m, which are deemed sufficiently small to consider model outputs as a good proxy of the real system (Fig. 3).
We then simulated several inundation scenarios to produce datasets used to train the machine learning model. The simulations were run for ten months. Next, we extracted 6000 predicted water levels at Pontianak with their associated input dataset. Based on the Pontianak Maritime Meteorological Station report, floods occurred over the city when the water level exceeds 2.5 m. 160 We therefore set that value as the threshold to define a flood event. It can be seen that several scenarios correspond to a water elevation greater than 2.5 m (Fig. 4).
All predictors considered in the machine learning model have mutual information (MI) coefficient greater than zero, which means that all predictor variables impact the river water level in Pontianak (Fig. 5). The relationship between these predictors and the water level could be linear or non-linear (as MI captured both types of relation). Here we found that the tidal elevations 165 in the estuary (X1, X2, and X3) have the most decisive impact on the river water level in the city (MI > 0.5), while tidal elevation observed one hour before (X2) is the strongest one. Next, the wind speed (max and average), the discharges (from both the Kapuas and the Landak river), and the pressure have a moderate relatedness. Meanwhile, the wind direction and the rainfall have only a weak relatedness (MI < 0.1), which means that both parameters have no a significant impact. Three machine learning models with different algorithms have been built to represent the relationship between dependent 170 (hourly water level modelled with SLIM) and independent parameters (tide, discharges and weather parameters). Models were trained based on 6000 scenarios of the SLIM output data. During the training phase, all NSE coefficients are greater than 0.75, which means that all algorithms perform very well during the training phase (Fig. 6). The most accurate algorithm is the RF followed by the SVM and the MLR. We thus know that all tested machine learning algorithms are promising and need to be evaluated in the next testing phase to select the best one. 175 In order to evaluate the performance of each model in predicting real flooding events, we tested the machine learning models on the observational data. Testing data was selected during the high discharge season for three different months during which inundations occurred (Dec 2018, Jan 2020, Jan 2021). Fig. 7 shows the predicted water levels, produced by each proposed algorithm, compare to the observational data. The accuracy of models to predict flooding events, marked by points in Fig. 7, is evaluated in this phase. 180 Even though all algorithms perform very well during the training phase, the performances are different during the testing phases ( Fig. 8 and Table 3). The RF accurately performs for three different testing phases. Its NSE values range from 0.61 to 0.72, which is a good performance. On the other hand, the MLR and the SVM algorithms succeed in the training phase and in the first and third testing phases but are less successful in the second testing phase. Their NSE on the second testing is only 0.39 and 0.40, respectively. Therefore, we know that the RF is the best machine learning algorithm in modeling water level 185 dynamics for our test case.

Commented [JS30]:
We modified and moved these paragraphs to section 2.

Commented [JS31]:
We modified this paragraph to: During the training and testing phases, all NSE coefficients are greater than 0.8 both in the training and testing phases, which means that all algorithms perform very well. The most accurate algorithm is RF, followed by SVM and MLR (Fig. 9). As such, we know that all the tested machine learning algorithms are promising and need to be evaluated in the implementation phase using observational data.

Commented [JS32]:
We modified this paragraph to: Therefore, we implemented the machine learning models on the selected observational data, which were obtained during the high discharge season for three months in three years when inundations occurred (Dec 2018, Jan 2020, Jan 2021). Fig. 10 shows each proposed algorithm's predicted water levels compared to the observational data. Subsequently, the accuracy of models to predict flooding events, marked by points in Fig. 10, is evaluated.

Commented [JS33]:
We modified and split this paragraph into: Even though all algorithms performed very well during the training and testing phases, the performances differed during the implementation phase (Table 3). However, the RF showed high accuracy in three different implementation phases. From the three different observational datasets, RF's NSE values range from 0.61 to 0.72, which is a good performance. While the MLR algorithm succeeded in the training and testing phases, it only succeeded in the first and third implementation phases, with NSE of 0.72 and 0.65, respectively. The model was less successful in the second implementation phase, with NSE hitting only 0.35 for this implementation dataset. Next, the SVM algorithm's performance is similar to the MLR algorithm. It succeeded in the training and testing phases but only succeeded in the first and third implementation phases, with NSE reaching 0.71 and 0.63, respectively. However, it failed in the second implementation dataset, with an NSE of only 0.41, which is slightly better than MLR.
Regarding flood events prediction, the RF algorithm also performed better than the other algorithms. The algorithm could predict ten out of seventeen events with a RMSE of 0.11 m. On the other hand, MLR and SVM can only predict six and eight events with RMSE of 0.19 m and 0.18 m, respectively. All of these algorithms, however, also predicted four false-positive events, i.e., flood events that never occurred, during the testing phase. 190

Discussion
The two main issues that have been tackled in this study are data scarcity and low computational resources for building flood forecasting model in developing countries (Brocca et al., 2020;Singh et al., 2021). Here, we showed that using an approach that combines hydrodynamic and ML models is promising to obtain a reliable and robust model. We succeeded in building and evaluating ML models trained by the hydrodynamic model output; hence, they did not require extensive observational 195 data in their training phase and did not need high computational cost in its implementation. Therefore, the proposed model is reliable for areas where observational data are scarce and computational resources are limited.
Since the proposed model can accurately forecast water levels, local water management agencies can rely on the model outputs for flood forecasting. Since the machine-learning did not required high computational resources, the limited computational resources will not prevent them from assessing and mitigating their compound flooding hazards. Using the model, they can 200 re-assess their compound flood events and predict the future events. Moreover, once they have more observation data, they can fuse the data to re-adjust the proposed model or to build a more robust one (Muñoz et al., 2021).
Next, we found that the RF algorithm is the best ML algorithm to assess compound flooding in our area of interest. In general, the performances of all tested ML algorithms for water level prediction are reasonable and acceptable. However, considering the NSE values in all testing phases, the number of flood events that are accurately predicted, and how close the predicted 205 water level is during the events, the RF performs clearly better than other algorithms. The superiority of the RF algorithm in predicting water levels has also been shown in previous studies in the Upo Wetland (Choi et al., 2020) and the Poyang Lake (Li et al., 2016). Therefore, we proposed a machine learning model with the RF algorithm as the most appropriate model to be implemented in the study area.
In addition, we also found that the tidal elevation measured one-hour prior at the river mouth is the main parameter controlling 210 the river water level at Pontianak. Even though the city of Pontianak is located 20 km from the river mouth, the tidal dynamics still strongly affects the river water level changes in the city. This result confirms those of a previous study, which revealed that the tide propagation on the Kapuas River still impacts the water level in Sanggau, a city located 285 km from the river mouth (Kästner et al., 2019).
However, the machine learning model proposed here also has some limitations. Since the rainfall impact on river water level 215 is minor compared with the other parameters, the model could not optimally capture the urban flooding over the city due to excessive rainfall. Based on the field observation, a short inundation occurs in the city if excessive rain occurs for a few hours.
This inundation could be due to the poor quality of the urban drainage system. Unfortunately, this phenomenon is not directly Commented [JS34]: We modified and split this paragraph into: Regarding flood events prediction, the RF algorithm also performed better than the other algorithms. It could predict eleven out of seventeen events (65% accuracy). On the other hand, MLR and SVM could only predict six and ten events (35% and 59% accuracy, respectively). Therefore, we know that the RF is the most accurate machine-learning algorithm to predict floods for our test case. Unfortunately, these three algorithms also predicted false-positive events, i.e., flood events that never occurred during implementation (Table 3). While the RF predicted four false events, the MLR and the SVM predicted three false events. This false event prediction is the shortcoming of the algorithm, which should be addressed in future studies.

Commented [JS37]: Changed to: agencies
Commented [JS38]: Modified to: Next, we found that the RF algorithm is the best ML algorithm to predict water level as a proxy for compound flooding in the area of interest.

Commented [JS39]:
Modified to: it could be concluded that the RF performs better than other algorithms.

Commented [JS41]:
We modified this sentence: This result confirms previous studies, revealing that the tide propagation on the Kapuas River dominantly controls the river water level up to 30 km upstream (Sampurno et al., 2021), and still impacts up to 285 km from the river mouth (Kästner et al., 2019). captured by the water level observation located within the river. The increase of the river water level due to the heavy rain is minor. Next, the model relies on the predicted input parameters such as weather parameters and river discharges to predict the 220 future water level. Consequently, the more biased the predictors, the higher the uncertainty in water-level prediction.
Overall, our integrated approach can provide a model to assess compound flooding driven by the interaction of tide, storm surge from the oceanward, and high discharge from the river upstream. However, the model still cannot optimally assess flooding caused by heavy rain over the city due to the limitation of the chosen indicator capability to capture the events.
Therefore, we will enhance the model capability in future studies by considering more indicators representing flooding events 225 due to excessive rainfall. Moreover, we will also try to reduce the number of predictors to minimize the model output's uncertainty. Mean sea-level rise due to climate change will also be evaluated to broaden the model implementation and create better flood mitigation.

Conclusion
This study shows that an integrated approach between the hydrodynamic and the machine learning models successfully 230 overcomes modeling river water-level and predicting compound flooding in a data-scarce environment with limited computational resources. Therefore, the approach is suitable for local water management agencies in developing countries whose encountered those issues. Regarding the implementation in Pontianak, we found that the machine learning model with the RF algorithm has the most accurate output compared to the other algorithms. In addition, the tidal elevation, measured one hour before the current time, is the main predictor for river water level modeling in the study area. 235

Author contribution
JS, VV, and EH conceptualized the research; JS and RA curated the data; JS, VV, and EH analyzed the data; JS wrote the manuscript draft; JS and EH reviewed and edited the manuscript.

Competing interests
The authors declare that they have no conflict of interest. However, the integrated model proposed in this study also has some limitations. Firstly, the accuracy of the machine learning model built depends on the accuracy of the hydrodynamic model. The more accurate the hydrodynamic model in predicting observational floods, the better the machine learning model will perform. Therefore, we need to tune the hydrodynamic model as accurately as possible. Next, since the rainfall impact on river water level is minor compared to other parameters, the model could not optimally capture urban flooding due to excessive rainfall. Based on the field observation, the city is shortly inundated if rain falls excessively for a few hours. This inundation could be due to the poor quality of the urban drainage system. Unfortunately, this phenomenon is not directly captured by the water level observation located within the river. The increase in the river water level due to the heavy rain is minor. Furthermore, the model relies on the predicted input parameters such as weather parameters and river discharges to predict the future water level. Consequently, the more biased the predictors, the higher the uncertainty in the water-level prediction.

Commented [JS43]: Modified to:
Regarding the limitation of the chosen indicator's capability to capture flood events, we will look for more data and indicators to enhance the model capability in future studies. Moreover, we will reduce the number of predictors to minimize the model output's uncertainty.

Commented [JS44]:
Modified to: Moreover, we will reduce the number of predictors to minimize the model output's uncertainty. We will also evaluate mean sea-level rise due to climate change to broaden the model implementation and create better flood mitigation.

Commented [JS45]:
Modified to: that are faced with these issues.

Commented [JS46]:
We added new sentences here: However, the accuracy of the machine learning model depends on the accuracy of the hydrodynamic model. If the hydrodynamic model is inaccurate in predicting real-life floods, the machine learning model's accuracy will also be lower. Besides, it has not yet optimally captured the urban flooding due to excessive rainfall. Considering more indicators representing this kind of flooding is essential to enhance the model's capability in the future. Region.

Commented [JS49]:
We changed this caption to: Figure 3: The hydrodynamic model domain is discretized with an unstructured mesh whose resolution is set to 50 m along the riverbanks, 400 m along the coast near the estuary, 1 km over the rest of the coastline, and 5 km offshore. The bathymetry of the model domain ranges from ~100 m depth offshore to 1 m in the river mouth.