Assessment of TRMM rainfall data for flood modelling in three contrasting catchments in Java, Indonesia

Flooding is defined as inundated water on the ground that causes material and non-material damage (Hammond et al. 2015; Cigler 2017). Floods also cause considerable losses to various life sectors, especially the economic and environmental sectors (Jongman et al. 2012; Carrera et al. 2015; Vojtek & Vojteková 2016).

The island of Java in Indonesia has a population of 151.6 million, which is 56% of the total Indonesian population (Hasil Sensus Penduduk 2020). Due to the intense tropical rainfall, many areas in Java are susceptible to river flooding (Marfai et al. 2008; Yulianto et al. 2019). One of the cities that are most prone to river flooding is the capital of Indonesia, Jakarta, with a population of over 10 million people (Budiyono et al. 2016). Every year, floods in the Jakarta region result in economic losses of more than 1 billion US dollars (Sagala et al. 2013; Wijayanti et al. 2017; Januriyadi et al. 2018). The National Disaster Management Agency (BNPB) has noted that during 1994–2019, the frequencies of flooding in Jakarta have increased from year to year. This increasing frequency of flood events is caused by a combination of land-use changes and subsidence within Jakarta (Abidin et al. 2011; Moe et al. 2017). Throughout Java, the increased population has caused an increase in urbanisation, and it is recognized that increasing urbanisation reduces infiltration and increases surface runoff, which triggers flooding (Ogden et al. 2011; Giang et al. 2014; Guan et al. 2015; Kumar Mishra & Herath 2015; Ng 2016; Saraswat et al. 2016). Further deforestation and urbanisation on the island of Java, together with climate change, have the potential to further increase the flood risk on the island (Rafiei Emam et al. 2016; Marhaento et al. 2018). Therefore, mitigation efforts in the form of countermeasures are necessary to reduce the impact and the risk of flooding (Aerts et al. 2013; Muis et al. 2015; Rafiei Emam et al. 2016; Haer et al. 2017; Nofal & van de Lindt 2020).

Hydrological modelling is an important tool in flood forecasting (Praskievicz & Chang 2009; Jain et al. 2018) and also valuable in understanding the effect of land-use and climate change on river flows (Zhang et al. 2018). In hydrological models, the quality of the rainfall data and their spatial and temporal resolution are important in the accuracy of any model prediction (Bardossy et al. 2022). But in developing countries, such as Indonesia, there are large areas with no rain gauges, and when the data exist, they are often incomplete and of variable quality. With remote sensing technology, the availability of rainfall data with complete time series and good spatial and temporal resolution is possible. Hence, rainfall data derived from the Tropical Rainfall Measurement Mission's (TRMM) satellite and the more recent Global Precipitation Measurement (GPM) satellite have been officially used to build flood early warning systems in Indonesia.

A number of studies around the world have considered the accuracy of TRMM rainfall data (Maggioni et al. 2016) and other satellite rainfall products such as CMORPH and PERSIANN (Bitew & Gebremichael 2011). In some cases when carrying out hydrological modelling, the TRMM data produce a similar quality of results compared to rain gauge data (Xue et al. 2013). But the quality of TRMM rainfall data is less good in the tropical regions of Asia, with Mohd Zad et al. (2018) suggesting that in Malaysia it would need prior correction for hydrological modelling. Wang et al. (2017) found that in the Mekong basin, it overestimates the amount of rain when the rain is light and underestimates when the rain is heavy. In the Java region of Indonesia, Sekaranom et al. (2018) also found that TRMM rainfall data are lower than rain gauge values during the heavy rains that occur in the wet season. Senjaya et al. (2020) corrected TRMM rainfall data using measured rain gauges for a hydrological analysis of the upper Bengawan Solo River catchment in Java.

The aim of this study is to test whether the use of rain gauge data or TRMM rainfall data on the island of Java produces a better hydrological model and so assess which is better for predicting flooding events. It is the first analysis of its type on the use of TRMM rainfall data to predict flooding in Indonesia. Three contrasting catchments are considered with hydrological modelling carried out using rain gauge data and TRMM data. The modelled discharges using both sources of rainfall data are compared to the measured discharges with conclusions made about which rainfall data source produces the best results. In addition, the success of flood warnings using both rainfall data sources is considered.

The Citarum river is the longest in West Java with the outflow into the Java Sea at the eastern edge of the Jakarta metropolis. This work considers only the upper part of the catchment (Figure 2) to the gauging station at Nanjung (1,753 km²). This catchment contains the city of Bandung which is surrounded to the north, east and south by volcanoes. In the flatter areas of the catchment (outside the city), the land use is predominantly agriculture, and in the steeper areas, it is shrubland and forest. Previous hydrological models of this catchment include those by Harlan et al. (2010), Harlan et al. (2018) and Siregar (2018).

The Bengawan Solo is the longest river in Java. The main river channel has its source near the mountains at the southern edge of Java and flows northwards between the volcanoes that dominate the centre of the island, before heading in a general north-easterly direction to its outlet into the Java sea. This work considers the river catchment to the gauging station at Babat (14,485 km²) just upstream of where the river becomes tidal (Figure 2), and it is the largest of the three catchments and has the lowest annual precipitation. The catchment is predominantly agricultural, although there is some shrubland and forest on the steeper and mountainous parts of the catchment and a number of built-up areas including the city of Surakarta.

The digital elevation model (DEM) data are obtained from the Shuttle Radar Topography Mission (SRTM), which were in turn obtained by the United States National Aeronautics and Space Agency (NASA) with a spatial resolution of 90 m × 90 m (http://srtm.csi.cgiar.org/). The DEM data were also used to define the catchment boundaries.

Land cover data use satellite imagery from the Moderate Resolution Imaging Spectroradiometer (MODIS) (Friedl & Sulla-Menashe 2019). MODIS observes the entire surface of the earth every 1–2 days and collects its data in 36 spectral bands or several groups of wavelengths. The data consist of 17 classes divided according to a scheme defined by the International Geosphere-Biosphere Program (IGBP). We use the MODIS land-use data from 2005.

Soil data were obtained from the Harmonized World Soil Database (HWSD) produced by the Food and Agriculture Organization (FAO) (FAO/IIASA/ISRIC/ISS-CAS/JRC 2012). This combines regional and national updates of soil information worldwide. It provides details of the soil type and the textures of the top (0–30 cm) and deeper soil profiles (30–100 cm). Also included are details of the drainage, available water storage capacity and reference depth.

Two sources of rainfall data are used in this work, measured rain gauge data and TRMM satellite data. For the Ciliwung catchment, hourly rain gauge data were obtained at Citeko, located within the catchment (Figure 2(a)), from BMKG (the Meteorological, Climatological, and Geophysical Agency of Indonesia) for the period from 2002 to 2008. For the Citarum and Bengawan Solo catchments, daily rain gauge data from the Indonesian Ministry of Public Works were used. These data are of variable quality (Remondi et al. 2016; Sekaranom et al. 2018) and needed significant quality control (Faybishenko et al. 2022). This yielded rainfall data from 2005 to 2008 for four locations of the Citarum at Nanjung catchment (Figure 2(d)) and from 2006 to 2009 for eight locations of the Bengawan Solo at Babat (Figure 2(g)). In both cases, the locations were well distributed around the catchment, and the small gaps present in the records were infilled using data from nearby rainfall stations.

Satellite data from the Tropical Rainfall Measuring Mission (TRMM) data source (Huffman et al. 2007; Tropical Rainfall Measuring Mission 2011) were used. This was a joint space project between NASA and the Japan Aerospace Exploration Agency (JAXA), which monitored rainfall in the tropics from 1997 to 2015. Its successor the GPM satellite was launched in 2014. The Integrated Multi-satellitE Retrievals for GPM (IMERG) algorithm fuses precipitation estimates from the TRMM and GPM satellites. In this study, the TRMM 3B42V7 algorithm is used with three hour totals available. There are four stages in the algorithm to produce the totals. Firstly, microwave precipitation estimates are calibrated and combined; secondly, infrared precipitation estimates are created using the calibrated microwave precipitation; thirdly, the microwave and infrared estimates are combined; and fourthly, rain gauge data are incorporated. A single TRMM location was used for the Ciliwung at Katulampa catchment (Figure 2(a)), five locations for the Citarum at Nanjung catchment (Figure 2(d)) and 23 locations for the Bengawan Solo at Babat catchment (Figure 2(g)).

Potential evaporation was obtained from the Global Land Evaporation Amsterdam Model (GLEAM) for the year 2002–2008 (Miralles et al. 2011). The potential evaporation is calculated based on observations of surface net radiation and near-surface air temperature.

Obtaining high-quality discharge data for river flows for catchments in Java proved to be quite a challenge. Due to the steep topography, erosive climate and the type of soils, which are the result of recent volcanic activity, there is a large amount of sediment transported in rivers in Java (Valentin et al. 2008; Verbist et al. 2010). The large amount of sediment means the stage–discharge relationship is not stable and requires a number of new curves every year (Pizarro et al. 2020). In many cases, new stage–discharge relationships have not been defined, so the measured discharges are not reliable.

In addition to standard quality checks on the discharge data (Crochemore et al. 2020), three extra checks were made. Firstly, manual checks of the hydrograph were done to check for obvious jumps in the measured discharge, as a result of removing sediment or a change in the stage–discharge relationship. Secondly, manual checks were also done for periods where a constant discharge was measured, suggesting errors in the measurements of the stage. The 2006 discharge data at Katulampa were rejected for this reason. Thirdly, the annual water balance of each catchment was considered, and the discharge data were rejected when the measured discharge data were producing annual totals outside the range of possible values (taking into account that there is little annual variation in stored water or losses to other catchments).

Reasonable quality daily discharge data were obtained for three locations: Katulampa on the Ciliwung river (2002–2005 and 2007–2008), Nanjung on the Citarum river (2005–2008) and Babat on the Bengawan Solo (2007–2009). In addition, hourly stage and discharge data were obtained at Katulampa in 2004–2005 and 2007–2008. Although the timings of the peaks in these datasets are correct, there is still considerable uncertainty in the measured discharge values. There is also a gap in the data collected at Katulampa from 3 February 2007 at 18:00 to 5 February 2007 at 9:00 during a catastrophic flood, which has an estimated return period of 150–300 years (Liu et al. 2015).

In this study, we used the Shetran hydrological model to model the three catchments. Shetran (http://research.ncl.ac.uk/shetran/) is a physically based distributed modelling system for water flow, sediment and solute transport in river basins (Ewen et al. 2000; Birkinshaw et al. 2010). It has been used for simulating water flow in a number of catchments in developing countries with low-quality input data (Op de Hipt et al. 2017; Sreedevi et al. 2019).

The most convenient way of visualizing the Shetran model structure is as a set of vertical columns with each column divided into 34 finite-difference cells. The lower cells contain aquifer materials and groundwater, higher cells contain soil and soil water and the uppermost cells contain surface waters and the vegetation canopy. River channels are specified around the edge of the finite-difference columns, and the location and elevations of these channels were calculated automatically depending on the upstream contributing area (Birkinshaw 2010). The grid resolution was set at 500 m for the Ciliwung catchment, 1,000 m for the Citarum catchment and 2,000 m for the Bengawan Solo catchment. The Shetran grid network, elevations and river channels for the three catchments can be seen in Figure 2(b,e,h).

The 17 land cover categories derived from MODIS were simplified to 7 classes of land covers in Shetran (Figure 2(c,f,i)). The vegetation parameters used standard Shetran libraries (Lewis et al. 2018). Shetran soil parameters (porosity, residual moisture content and van Genuchten parameters) for the different HWSD soil types were based on the soil textures and the Hypresv2.0 database (Wösten et al. 1999). Calibration of the saturated hydraulic conductivities and soil depths was carried out taking into account the deep volcanic ash soils (Van Ranst et al. 2002) with high porosities and conductivities (Dahlgren et al. 2004) found on Java.

The aim of this study is to test whether rainfall from the rain gauges or the TRMM data produced better simulation results, i.e. which is best able to simulate potential flooding issues downstream. Therefore, the catchments used the same temporal resolution for the measured rain gauge and TRMM data. For the Ciliwung catchment, 3-h measured rain gauge data and 3-h TRMM data were used. For the Citarum and Bengawan Solo catchments, daily measured rain gauge data and daily TRMM data were used.

A split sample calibration/validation approach was used (Klemeš 1986). Separate calibrations were carried out using the measured rain gauges and the TRMM data. In all three catchments, the same soil parameters were found to produce the optimum calibration, but the actual/potential evaporation parameter was higher for the TRMM rainfall, which was needed to take into account the higher annual rainfall totals produced by this dataset (Table 1).

To test the quality of the simulations, the standard Nash Sutcliffe Efficiency (NSE) was used (Moriasi et al. 2007) for comparing simulated and measured discharge time series. NSE value ranges between −∞ and 1, where a value close to 1 indicates that the performance of a model is good. A value of 0 signifies that the model is as good as the mean of the measured time series. The test is sensitive to extreme values and hence is particularly useful for flooding issues. PBias (percent bias) was also calculated, and this is a measure of the average tendency of the simulated discharge to be larger or smaller than the measured values. A value of 0 signifies that there is no bias in the simulated discharge.

Due to the short record lengths, a threshold event with a return period of approximately half a year was selected (Norbiato et al. 2008) which corresponds to 36 m³/s in the Ciliwung catchment, 318 m³/s in the Citarum catchment and 1,820 m³/s in the Bengawan Solo catchment. In addition, for each of these events, the day of peak discharge predicted by the model was compared to the day of the peak discharge in the measured discharge, and this was done for both the measured rain gauge data and TRMM rainfall data.

Figure 4 shows that for measured rain gauges, only 20–30 km apart, the cross-correlation is in the range of 0.15–0.56 with a typical value of 0.3. For the same separation between locations, the TRMM values are much higher and in the range of 0.65–0.85 with a typical value of 0.72. Cross-correlations for rain gauges in regions with more frontal rainfall and less intense convection rainfall also have much higher cross-correlation values for the same separation between locations (Berndtsson 1987; Bacchi & Kottegoda 1995; Burton et al. 2013). For example, Burton et al. (2013) found that daily rainfall in July (summer events) in the United Kingdom has a cross-correlation value of around 0.7–0.8 if the separation between rain gauges is 20–30 km. A cross-correlation value of 0.3 would only be expected if the rain gauges are 250 km apart. This analysis shows how localised rainfall events are in Java, and hence building a good hydrological model is difficult as some of the rainfall events will be missed by the measured rain gauges and some events captured by the rain gauges will be applied to parts of the catchment where they did not occur.

The correspondence between the observed discharge and the simulated discharge for the calibration and validation periods for the three catchments and for both the measured rain gauge data and the TRMM data is shown in Table 2. The rain gauge data are producing consistently better results than the TRMM rainfall data. Considering mean daily discharge, the Nash Sutcliffe efficiency (NSE) values for the rain gauges for the three catchments for the calibration period were 0.75, 0.70 and 0.85 and for the TRMM rainfall data were 0.44, 0.44 and 0.75. The NSE values were 0.71, 0.62 and 0.89 for the validation period using rain gauges and 0.26, 0.61 and 0.58 for the TRMM data. The relatively small differences in NSE values between calibration and validation samples for the rain gauge-driven simulations give reassurance to the Shetran model performance. The measured rainfall data are also producing better results for the 3-h rainfall in the Ciliwung and also for the monthly discharge data, apart from for the validation period for the Citarum catchment where the TRMM rainfall data produce slightly better results.

As expected, due to averaging effects, in all three catchments, the NSE values are best for the monthly discharge, followed by the daily discharge and least good for the 3-h discharge. For the monthly discharge, capturing the pattern of the wet and dry seasons is important, whereas for the 3-h discharge, the individual events are important and these localised, short-duration, high-intensity events are hard to capture correctly. This is also the reason why overall the TRMM rainfall data results are only slightly less good than the measured rain gauge data for the monthly discharges, as this rainfall dataset captures the monthly variation but is unable to capture variations in the 3-h and daily rainfall (Figures 3 and 5). Similarly, the results from the much larger Bengawan Solo catchment are better than those of the other two catchments, as the discharge hydrograph is much more attenuated so measuring the localised event correctly is less important.

The ability of the models to correctly generate flood warnings for the three catchments is presented in Table 3. In the Ciliwung at Katulampa catchment using the rain gauge data, 10 measured events greater than the threshold value of 36 m³/s were correctly simulated as being greater than this value (hit). Four measured events greater than 36 m³/s were simulated as being less than 36 m³/s (miss), and three events were simulated as being greater than 36 m³/s when the measured discharge was less than 36 m³/s (false). The overall CSI value (Equation (1)) was 0.59. In the case of the TRMM data, there were 2 hit events, 12 miss events and 2 false events and a much lower CSI value of 0.13. For the other two catchments, the results also show that the measured rain gauge data produce better results than the TRMM rainfall data. In the Citarum catchment, the CSI value is 0.38 for the rain gauge data and 0.08 for the TRMM data, whilst in the case of the Bengawan Solo, the rain gauge data have a CSI value of 0.63 and the TRMM data have a CSI value of 0.50.

From a flood warning perspective, it is also important that the models are able to correctly predict the timing of the events, to prepare for flooding and potentially save lives. For the same events, Table 4 shows the number of times in each catchment where the peak was predicted on the correct day and the number of times where it was predicted to be 1 day or 2 days different. Predictions of the timing are better in the Ciliwung catchment as 3-h rainfall data are being used, and there is a shorter lag time and peakier discharges. In the case of rain gauge data at Ciliwung, 11 out of 14 events were correctly predicted as being on the correct day, whereas in the case of the TRMM data, only 4 out of 14 were correctly predicted. For the other two catchments, results are slightly better using the rain gauge data compared to the TRMM data.

There are regular flooding events on the island of Java in Indonesia, and with a population of over 150 million people, the social and economic costs are huge. In the Jakarta region alone, annual economic losses are more than 1 billion US dollars, and the National Disaster Management Agency of Indonesia says that this issue is getting worse, with more flooding events in recent years. Any methods that can improve predictions of flooding have the potential to save lives and also save a considerable amount of economic damage. One method to predict flooding is early warning systems triggered by high water levels, such as at Katulampa on the Ciliwung river (Syafalni et al. 2015), which provided approximate warning before 10 h in the centre of Jakarta.

In combination with these early warning systems triggered by high water levels, hydrological modelling has the potential to improve flooding predictions. Hydrological modelling also has the potential to improve flooding predictions on many other rivers without such systems. To successfully carry out hydrological modelling, high-quality rainfall data with a good spatial resolution is needed. This research has tested which measured rain gauge data or TRMM satellite rainfall data when used in a hydrological model are better at simulating the discharge and producing accurate flood warnings. The results of the Shetran hydrological modelling for three contrasting river catchment (Ciliwung at Katulampa, Citarum at Nanjung and Bengawan Solo at Babat) show that the measured rain gauge data produced substantially better-simulated discharges than those using the TRMM rainfall data. This was for 3-h, daily, and monthly discharges in all three rivers, apart from monthly discharge on the Citarum river. Flood warnings above a threshold discharge level in each of the three catchments were improved using the rain gauge data compared with that using the TRMM data, with higher CSI values. The timings of the flood peaks in each of the three catchments were also improved using the rain gauge data compared with the TRMM data. It is important to note that the measured rain gauge data consistently produced better results than the TRMM data across three contrasting catchments with areas from 150 to 14,485 km² and from the west to the east of Java. It is therefore suggested that measured rain gauges rather than TRMM or GPM satellite rainfall data are used for flood warning systems on the island of Java.

The research has also highlighted that accurately predicting flooding on the three rivers is difficult. Cross-correlation between the rain gauges shows the rainfall is localised with high-intensity, short-duration events. Localised events can be missed by all of the rain gauges, which can lead to under prediction of the simulated discharge, or a very localised event captured by a rain gauge can produce an over prediction of discharge. A recent work has highlighted how errors in capturing the spatial distribution of rainfall can cause 50% of the errors in the model results (Bardossy et al. 2022). So it is recommended that additional high-quality rain gauges with hourly or sub-hourly temporal resolution are installed, and these will further improve the reliability and accuracy of the rainfall data and so the quality of the simulated discharge from using the measured rain gauge network. It will improve both the modelled peak discharges and the timing of the peak discharges, which are both important from a flood warning perspective.

There are also many issues with obtaining high-quality discharge data, which are necessary to calibrate the models. The steep volcanic soils and high-intensity rainfall produce a lot of sediment which significantly affects the quality of the discharge measurements. So it is recommended that more investment is carried out in the quality of discharge measurement and in particular improving the stage–discharge relationship with regular measurements of the discharge. This will improve both the early warning system for flooding based on water level data and the hydrological simulations.

Improvements in both rainfall and discharge measurements will enable better predictions to be made of potential flooding issues on the island of Java. Investment in more and better data will be less expensive than the potential damage caused by flooding.

All relevant data are included in the paper or its Supplementary Information.

The authors declare there is no conflict.