Impact of land cover, rainfall and topography on flood risk in West Java

Flooding represents around 32% of total disasters in Indonesia and disproportionately affects the poorest of communities. The objective of this study was to determine significant statistical differences, in terms of river catchment characteristics, between regions in West Java that reported suffering from flood disasters and those that did not. Catchment characteristics considered included various statistical measures of topography, land-use, soil-type, meteorology and river flow rates. West Java comprises 154 level 9 HydroSHEDS sub-basin regions. We split these regions into those where flood disasters were reported and those where they were not, for the period of 2009 to 2013. Rainfall statistics were derived using the CHIRPS gridded precipitation data package. Statistical estimates of river flow rates, applicable to ungauged catchments, were derived from regionalisation relationships obtained by stepwise linear regression with river flow data from 70 West Javanese gauging stations. We used Kolmogorov–Smirnov tests to identify catchment characteristics that exhibit significant statistical differences between the two sets of regions. Median annual maximum river flow rate (AMRFR) was found to be positively correlated with plantation cover. Reducing plantation land cover from 20 to 10% was found to lead to a modelled 38% reduction in median AMRFR. AMRFR with return periods greater than 10 years were found to be negatively correlated with wetland farming land cover, suggesting that rice paddies play an important role in attenuating extreme river flow events. Nevertheless, the Kolmogorov–Smirnov tests revealed that built land cover is the most important factor defining whether or not an area is likely to report flood disasters in West Java. This is presumably because the more built land cover, the more people available to experience and report flood disasters. Our findings also suggest that more research is needed to understand the important role of plantation cover in aggravating median annual maximum river flow rates and wetland farming cover in mitigating extreme river flow events.


Introduction
Flooding represents around 32% of total disasters in Indonesia (Yulianto et al. 2015) and disproportionately affects the poorest of communities (Kim and Gim 2020). The impact of fluvial flooding is expected to increase due to increasing rainfall intensity, land subsidence, sedimentation of river channels, rapid population growth and land-use change (Marfai et al. 2015). Historical changes in both land-use and climate have collectively led to increased river flow rates in the region of West Java, with land-use thought to have played a stronger role (Julian Poerbandono and Ward 2014).
Since as early as the 1920s, agricultural activity has represented over 70% of total land cover in Java (Verburg and Bouma 1999). Land cover in West Java can be broadly categorised into six land-use types including water body, built land, dryland farming (mostly agriculture with low irrigation requirements), wetland farming (mostly rice paddies and fish farms), plantation (mostly tea plantation) and forest (including both natural and managed forest) (consider Siswanto and Francés 2019; Ridwansyah et al. 2020;Yulianto et al. 2022). Indonesian spatial planning law prescribes that 30% forest cover is needed in all local government districts to ensure adequate hydrological forest functions (Suprayogo et al. 2020).
Forest cover, as compared to other land cover, provides higher rates of evapotranspiration (including canopy interception loss), higher rates of infiltration and lower rates of surface runoff. A plot scale study in East Java measured runoff coefficients for production forest (both old and young) and arable land to be 0.03 and 0.41, respectively (Suprayogo et al. 2020). Interestingly, hydrological models for West Java often assume runoff coefficients for primary forest, secondary forest and plantation of 0.01, 0.03 and 0.40, respectively (Yulianto et al. 2022).
Hydrological modelling in the Upper Citarum, West Java, suggests that reductions in forest cover, combined with increases in settlement cover, during the period of 1994 to 2014, led to a 15% increase in water yield (Siswanto and Francés 2019). Hydrological modelling in the Cimanuk catchment, West Java, suggests that transformation of bush and plantation to dryland agriculture, during the period of 2011 to 2017, led to a 60% increase in surface runoff (Ridwansyah et al. 2020). However, findings from such studies are strongly dependent on the assumptions and parameter values that define the underlying hydrological models used (Beven and Binley 1992;Beven 2006;Morán-Tejeda et al. 2015;Ekblad and Herman 2021). Furthermore, whether or not increases in river flow lead to problematic flood events is dependent on the damage and disruption likely to be caused (Ferreira and Ghimire 2012). An alternative approach is to directly study historical reports of flood disaster occurrence. By studying variations in country level "large flood event" reporting data from 56 countries, Bradshaw et al. (2007) concluded that a 10% reduction in natural forest area can lead to between 4 and 28% increase in flood disaster frequency. They derived their results using multiple linear regression considering country area, median average annual precipitation, average slope, total degraded area, natural forest cover, natural forest loss and non-natural forest cover. Their flood frequency data was obtained from the Dartmouth Flood Observatory, which is mostly derived from news and government sources (Brakenridge 2010).
However, revisiting the Bradshaw et al. (2007) analysis, van Dijk et al. (2009) found that the dependence on natural forest cover reduced to a negligible level when population density was additionally considered. Furthermore, they found that population density explained 83% of the variation in reported flood occurrences. Similar country scale findings were also observed by Ferreira and Ghimire (2012). Indeed, flood events are much more likely to be reported in areas where there are more people to experience and report the events.
Nevertheless, Bhattacharjee and Behera (2018) applied a similar methodology to identify key variables associated with reported flood occurrence at a district level. Specifically they looked at 13 districts in Western Bengal, India, and found that the frequency of reported flood events was strongly negatively correlated with forest cover. Important differences between the study of Bhattacharjee and Behera (2018) and those of Bradshaw et al. (2007), van Dijk et al. (2009) and Ferreira and Ghimire (2012) were that they considered a finer resolution (district as opposed to country scale) and had access to arguably more reliable flood reporting data (obtained directly from the Department of Disaster Management, government of West Bengal, Kolkata). A similar district-scale study in China revealed that frequency of reported flood events was negatively correlated with broadleaf and mixed-tree forest cover but independent of coniferous tree cover (Tembata et al. 2020). The validity of such studies is important because they are likely to affect policy making decisions relating to the preserving and/or extending of forest cover to reduce flood damage in downstream urban environments (Xiao et al. 2022).
A disadvantage of using countries or districts to define individual study areas is that the land within a district does not necessarily represent the hydrological watershed upstream of a given flood event. It would be better to study hydrological units as opposed to district level units, so as to better capture the hydrological contribution area of the flood events being studied.
Another problem concerns the strong correlation between flood reporting frequency and population density. Certainly flood reporting frequency will increase if there are more people to report the floods (van Dijk et al. 2009;Ferreira and Ghimire 2012). An alternative approach is to study statistical differences between areas that report flood disasters and areas that do not over a designated multi-year period. The advantage is that a flood disaster reported at one location in a small village will have the same weight as a flood disaster reported at multiple locations in a much larger town or city. It should then be possible to investigate differences between hydrological regions that report flood disasters and those that do not.
Unfortunately, this kind of dichotomous data is unsuitable for conventional linear regression techniques. However, parameter sensitivity can be explored using a two sample Kolmogorov-Smirnov test (KST). For example, suppose we have forest cover data for 200 river catchments. We can split the catchments into those that reported flood disasters and those that did not. We can then use the KST to determine whether there is a statistically significant difference between forest cover data describing the flood disaster catchments and the remainder of the population. Should such a difference exist, we can use the Kolmogorov-Smirnov statistic (KSS) as a measure of parameter sensitivity. Now suppose we study a whole range of different factors (forest cover, mean annual rainfall, median annual maximum river flow rate etc.). The KST can be used to identify which parameters show statistically significant differences. Those statistically significant parameters can then be ranked in terms of their KSS. Furthermore, the CDFs for the most important parameters can be directly studied to gain further insights concerning the difference between catchments that report flood disasters and those that did not.
The objective of this study is to identify catchment characteristics that exhibit statistically significant differences in CDFs for river catchments that report flood disasters in West Java, Indonesia (including the government districts of West Java, Jakarta and Banten). Improved understanding about the spatial pattern of flood risk will help inform land use planning at the national scale, hopefully leading to disaster risk reduction and improved resilience to extreme hydro-meteorological events.
During the period from 2009 to 2013, 601 flood disasters in West Java were reported to the Indonesian National Board for Disaster Management (BNPB 2021). West Java comprises 154 level 9 HydroSHEDS sub-basin regions (Lehner et al. 2008). We split these regions into those where flood disasters were reported and those where they were not. We derive statistical estimates of river flow rates from regionalisation relationships obtained by stepwise linear regression with river flow data from 70 West Javanese river flow gauging stations. We then use KSTs to identify catchment characteristics that exhibit significant statistical differences between the two sets of regions. Catchment characteristics considered include various statistical measures of topography, land-use, soil-type, meteorology and river flow.
The outline of this article is as follows. The various data sources are explained. The methods used for flood frequency analysis, stepwise linear regression and Kolmogorov-Smirnov testing are described. Results from the stepwise linear regression and Kolmogorov-Smirnov testing are presented and discussed. Relevant conclusions are then drawn.

HydroSHEDS data
Digital elevation data, drainage direction data, river network shape-files and level 9 HydroSHEDS sub-basins (standard) shape-files were acquired at 15 arc-second resolution (approximately 500 m), for the West Java region, from Lehner et al. (2008). The digital elevation data, river network shape-files and sub-basin shape-files are presented in Fig. 1a.
HydroSHEDS uses the Pfafstetter coding system (PCS) to generate twelve nested levels of sub-basins (Lehner and Grill 2013), each depicting consistently sized sub-basin polygons at scales ranging from millions (level 1) to tens of square kilometers (level 12) (Linke et al. 2019). Level 9 was selected for this study because it provides catchments of similar size to gauged river catchments in West Java.
The PCS splits each catchment into a set of hydrological regions based on where tributaries intersect with the main river channel (Lehner 2014). The PCS distinguishes between inter-basin regions and tributary basins. Tributary basins collect water only from the within their associated area. Inter-basin regions collect water from within their associated area and also any part of the wider river catchment upstream of the inter-basin region.
The polygons provided by the HydroSHEDS sub-basin package include tributary basins and inter-basin regions. Both sets of polygons are hereafter referred to as "regions". Each region has two areas identified with it, the area of the region and the area of catchment that contributes to the river outlet of the region. For tributary basins, the catchment area is equal to the region area. For inter-basin regions, the catchment area is greater than the region area.
HydroSHEDS delineates West Java into 154 level 9 HydroSHEDS sub-basins. The polygon areas associated with each region range from 57.2 to 879 km 2 with a median value of 230 km 2 . The catchment areas associated with each region range from 58.0 to 6710 km 2 with a median value of 298 km 2 . For each region we determine the number of flood disasters reported within the area associated with the region and catchment characteristics associated with the catchment area upstream of the river outlet of the region.
In addition to the level 9 sub-basins (hereafter referred to as the HydroSHEDS catchments), the catchment areas for the river flow gauging stations were derived from the HydroSHEDS drainage direction data using the D8 algorithm (Jenson and Domingue 1988).
For each of the gauged catchments and HydroSHEDS catchments, we determined the catchment area (km 2 ), circularity ratio ( = 4 catchment area ÷ square of catchment  Maps of West Java, Indonesia. a Shows digital elevation data, river network data and level 9 basins, all acquired from HydroSHEDS (Lehner et al. 2008). b Shows mean annual rainfall from CHIRPS (Funk et al. 2015) along with the locations of the river flow gauging stations and their associated catchments perimeter), drainage density ( = length of river channel network ÷ catchment area) (km −1 ), P10 elevation (mASL), P50 elevation (mASL), P90 elevation (mASL), P10 slope (m km −1 ), P50 slope (m km −1 ) and P90 slope (m km −1 ). The range and median for the above set of catchment characteristics are given for both the gauged catchments and the HydroSHEDS catchments in Table 1.
The terms P10, P50 and P90 refer to values with a 10%, 50% and 90% probability of non-exceedance within a given catchment. Length of river channel network was obtained from the 15 arc-second HydroSHEDS river network shape-files. The term mASL stands for meters above sea level.

Flood disaster reports
Records were acquired from all flood disasters reported to the Indonesian National Board for Disaster Management (BNPB 2021) for West Java during the period of 2009 to 2013. This dataset included 601 individual river flooding events, the locations of which are shown on Fig. 2a. Each of these events represents anything from an unplanned flood inundation to a flood event with large-scale property damage and multiple fatalities. The vast majority of events can be seen to be located in the high population areas of Jakarta and Bandung. Nevertheless, many additional events are spread out through many rural regions of West Java. But interestingly, some HydroSHEDS regions reported no flood events during the period of study (compare Figs. 1a and 2a).

Land-use data
Land-use data in West Java for 2011 were acquired at 1:250,000 from a web-resource managed by the Indonesia Ministry of Environment and Forestry (IMEF 2021). These data were derived from Landsat multispectral satellite data combined with ground check field survey (IMEF 2020). The web-resource distinguishes between fifteen different classifications including water body, built land, dryland farming, dryland farming mixed with shrubs, garden shrubs, mining area, open land, rice field, aquaculture pond, swamps, plantation, primary dryland forest, secondary dryland forest, mangrove forest, industrial forest. The classifications used were broadly derived from those previously described by Di Gregorio et al. (1998).
For our study, we aggregated these fifteen classifications into six simplified land-use types (see Table 2) perceived to have common hydrological functioning attributes, defined as follows: 1. Water body includes rivers, reservoirs and lakes. 2. Built land includes urban and rural settlements, factories and other built-up infrastructure. 3. Dryland farming includes cropland areas requiring low levels of irrigation mainly used for seasonal crops. Here we also include shrubs, mining area and open land. Mining area and open land represent 0.01% and 0.7% of total land cover in West Java, respectively, and are mostly covered by grass and shrubs. The hydrological functioning is therefore thought to be very similar to other dryland farming areas. 4. Wetland farming includes agricultural areas that experience both permanent and periodic inundation. These areas are mostly comprised of rice paddies. However, we have chosen to also include aquaculture ponds and swamp, the latter of which represents < 0.001 % of total land cover in West Java. Aquaculture ponds mostly include fish and shrimp ponds, which are commonly integrated within the rice farming system in Java (Nurhayati et al. 2016;Fatimahet al. 2020) and are therefore better considered with rice as a single landuse type from a hydrological functioning perspective. 5. Plantation includes dryland farming like practices where crops have an operational life greater than two years, mostly tea plantation. 6. Forest includes all dryland and mangrove forest, both natural and managed. Although mangrove forest is likely to have quite different hydrological functioning, it represents < 1 % of total land cover in West Java and was therefore not considered as a separate classification in this context. Although mangrove forest is very important for mitigating coastal flooding (Menendez et al. 2018), its role on river flooding is thought to be much less significant.  Figure 2a shows a map for the six land-use classifications in West Java alongside the location of the aforementioned reported flood events. It can be seen that wetland farming is concentrated on the north coast but is also widely distributed within the valleys of the central highlands. Forest and plantation mostly reside in the highland mountain regions.
The percentage land cover for each of the six land-use classifications was determined for each of the gauged catchments as well as the 154 level 9 HydroSHEDS catchments. The range and median values for each classification are given in Table 1.
Land-use distribution varies widely from one hydrological catchment to another. Figure 3 shows the cumulative distribution functions (CDF) for the different land-use classifications, in terms of percentage of land cover, for each of the 154 level 9 HydroSHEDS catchments. Water body represents a very small fraction of land cover for all the catchments studied. Dryland farming predominantly represents the largest fraction. Interestingly, less than 20% of the catchments in West Java satisfy the 30% forest cover government requirement, referred to by Suprayogo et al. (2020).

Soil-type data
Soil-type data were acquired at 1:250,000 from the Soil Research Centre, Ministry of Agroforestry, Indonesia. Soils in West Java are dominated by alluvial clays, alluvial soils (comprising sand and clay), Andosols, Latosols, Mediterranean soils, podzolic soils, Regosols and soil complexes (a mixture of more than two types of the previously mentioned soils) (Suhardjo and Soepraptohardjo 1982).
The alluvial clays and podzolic soils have high clay content and can be considered relatively low in permeability. The Andosols and Regosols have high sand content and can be considered relatively high in permeability. Other soils should be expected to be more intermediate in permeability. From a hydrological perspective we expect to see higher runoff coefficients in catchments dominated by low permeability soils.
A map of soil-type across West Java is shown in Fig. 2b. Alluvial clays dominate the main flood plains on the North coast. Alluvial soils (which have a higher sand content than the alluvial clays) dominate the main river channels. Andosols and Regosols dominate the highlands around Bandung and Bogor (a smaller city to the south of Jakarta).
The percentage land cover for each of the eight soil-type classifications was determined for each of the gauged catchments as well as the 154 level 9 HydroSHEDS catchments. The range and median values for each classification are given in Table 1.

Weather data
Gridded daily mean temperature, wind speed, relative humidity and incoming shortwave radiation were acquired at 0.25 • resolution from the AgMERRA data package (Ruane et al. 2015). These data were used to calculate reference crop potential evapotranspiration   according to FAO56 (Allen et al. 1998). Gridded daily precipitation data were acquired at 0.05 • resolution from the CHIRPS data package (Funk et al. 2015). The AgMERRA data package was chosen because it currently provides the most comprehensive gridded meteorological dataset (in terms of providing temperature, wind speed, humidity and shortwave radiation) for Southeast Asia. AgMERRA combines reanalysis data, gauged data and satellite date (Ruane et al. 2015). AgMERRA also provides precipitation data. However, the CHIRPS precipitation data package was chosen instead due to its higher spatial resolution. CHIRPS combines gauged data and satellite data (Funk et al. 2015) and has a significant track record of use in Indonesia (e.g., Narulita and Ningrum 2018;Auliyani and Wahyuningrum 2021;Wahyuni et al. 2021).
A map of mean annual rainfall for West Java is shown in Fig. 1b. Rainfall is not strongly controlled by topography and is mostly driven by monsoon and convection processes. Significantly elevated rainfall is observed on the West side, upstream of Jakarta, and this rainfall can be linked to the large cluster of reported flood hazards in Jakarta seen in Fig. 2a. Nevertheless, many reported hazards can also be seen around Bandung, where less rainfall is observed (see Fig. 2a).

River flow data
Observed daily mean river flow rate data was acquired for 70 river flow gauging stations, scattered across West Java, from the Center for Hydrology and Water Management, which is part of the Indonesia Ministry of Public Works (Balai Besar Wilayah Sungai). Additional details about measurement and quality protocol are presented and discussed by Yuningsih (2019).
The locations of each gauging station along with their associated catchment areas are shown in Fig. 1b (both the open and solid circular markers). Each gauge station record includes at least 10 complete (although not necessarily consecutive) years of daily mean flow rates from within the period of 1980 to 2016.

Flood frequency analysis
Flood frequency analysis provides a useful method for aggregating the daily river flow time-series into characteristic statistics that describe the extreme nature of the river flows in the gauged catchments. Following common flood frequency analysis practice (e.g., Lim and Lye 2003;Kjeldsen and Jones 2006;Mulyantari et al. 2011;Badyalina et al. 2021), we determine the median annual maximum flow rates, Q med (m 3 s −1 ), along with associated t and t 3 sample L-moment ratios (see, Hosking and Wallis 1997, p.28), denoted hereafter as Q t (-) and Q t 3 (-), respectively. The three river flow statistics, Q med , Q t and Q t 3 can be used to parametrise a generalised logistic distribution function, which describes the entire flood frequency curve of the designated river catchment (see, Kjeldsen and Jones 2006). Median values are preferred to mean values in this context to minimize the potential impact of outliers (Kjeldsen and Jones 2006).

Weather data statistics
Catchment averaged daily potential evapotranspiration and daily rainfall time-series were compiled for each gauged catchment and HydroSHEDS catchment studied. These were then used to derive corresponding values for mean annual potential evapotranspiration, mean annual rainfall, median annual maximum daily rainfall and median annual maximum number of consecutive rainfall days. The sample L-moment ratios, t (similar to coefficient of variation) and t 3 (similar to skewness) were also derived for both annual maximum daily rainfall and annual maximum number of consecutive wet days using equations provided by Hosking and Wallis (1997), p.28. The range and median for the above catchment characteristics are given for both the gauged catchments and the HydroSHEDS catchments in Table 1.

Step-wise linear regression
The analytical method described in Sect. 3.1 provides values for the river flow statistics, Q med , Q t and Q t 3 , for each of the 70 river flow gauging stations. However, to explore the importance of flood frequency data in determining whether or not a HydroSHED region reports flood disasters, we need to extrapolate the river flow statistics to the 154 level 9 HydroSHEDS sub-basins being studied. This can be achieved by using step-wise linear regression to derive regionalisation relationships, relating river flow statistics to relevant catchment characteristics for the gauged catchments.
Here we adopt a step-wise linear regression approach, previously applied by Mathias et al. (2016) and Ye et al. (2014), whereby additional parameters are incorporated until the so-called Bayesian Information Criterion (BIC) is minimised (i.e., Eq. (12) of Ye et al. (2014)). The procedure can be described as follows: 1. Determine the correlation coefficients of each catchment characteristic listed in Table 1 (with the exception of median annual maximum flow rate) with the flood statistic of concern ( Q med , Q t or Q t 3 ).

3
2. Select the catchment characteristic with the highest absolute correlation coefficient and check that the relationship passes a significance test (i.e., that the probability of getting a correlation as large as the observed value by random chance is < 0.05).

Develop a linear regression relationship between the selected catchment characteristic
plus any previously selected catchment characteristics and the flood statistic of concern. 4. Calculate the BIC. 5. Determine the correlation coefficients of the remaining catchment characteristics with the residuals between the developing regionalization relationship and the flood statistic of concern. 6. Repeat Steps 2 to 4. 7. If the new BIC is less than the previous BIC and the significance test from Step 2 passes, repeat Steps 5 to 7, otherwise consider the previous form of the regionalization relationship to be optimal.

Kolmogorov Smirnov testing
We use two-sample Kolmogorov-Smirnov tests (KST) to test the null hypothesis that individual catchment characteristics, describing the catchment areas contributing to HydroSHEDS regions with and without reported flood disasters, are from the same statistical distribution. The KST also provides the Kolmogorov-Smirnov statistic (KSS), which measures the maximum absolute difference between the cumulative distribution functions (CDF) for the two data populations. The higher the KSS, the more different the CDFs. When the KST fails, the difference between the CDFs, for the catchment characteristic under consideration, from the regions where flood disasters were reported and those where they were not, can be considered to be statistically significant. The KSS can then be used to measure the difference between the two CDFs, which represents a relative measure of parameter sensitivity.
We use the KST to identify statistically significant catchment characteristics, in the context of flood disaster occurrence. We then use the KSS to rank the statistically significant catchment characteristics in terms of their sensitivity (high KSS corresponds to high sensitivity) to identify the most important parameters.

Gauges with at least 10 years of data
The regionalisation relationships obtained by stepwise linear regression for the river flow statistics, Q med , Q t and Q t 3 , using river flow data from all 70 river flow gauging stations are listed as follows: (1) Q med = 0.02655x 1.078 1 exp 0.03195x 2 − 0.0006444x 3 (2) Q t = 0.3683x −0.06852 1 exp 0.9058x 4 where x 1 is catchment area (km 2 ), x 2 is median annual maximum daily rainfall (mm), x 3 is P10 elevation (mASL), x 4 is t 3 for annual maximum daily rainfall (-) and x 5 is percentage land covered by alluvial clays (%).
Plots of observed data against modelled data (using Eqs. (1) to (3)) for each of the 70 gauged stations studied are shown as red open markers in Fig. 4. The correlation between the modelled and observed data for Q med is very high with a correlation coefficient of 0.944. However the correlations for the sample L-moment ratios, Q t and Q t 3 , are much Results for sample L-moment ratio, t, for annual maximum flow rates, Q t (-). c Results for sample L-moment ratio, t 3 , for annual maximum flow rates, Q t3 (-) lower with correlation coefficients of 0.334 and 0.237, respectively, rendering these relationships of little value. Note that each of the 70 river flow gauging stations had at least 10 complete years of daily river flow data. However, only 21 of the river flow gauging stations had more than 20 complete years of data. The reason for the higher order statistical relationships (Eqs. (2) to (3)) performing so poorly is arguably due to 10 years of data being insufficient to observe these kind of higher-order phenomena.

Gauges with at least 20 years of data
To gain further insight, we repeated the stepwise linear regression exercise using only river flow gauging stations that had at least 20 complete years of flow data. This included 21 gauging stations in total, the locations of which are shown as black solid circular markers on Fig. 1b. The resulting regression relationships are listed below: where x 1 and x 2 are as before, x 6 is percentage land cover used by plantation (%), x 7 is catchment circularity (-) and x 8 is percentage land cover used by wetland farming (%).
Plots of observed data against modelled data (using Eqs. (4) to (6)) for these 21 gauging stations are shown as black solid circular markers in Fig. 4. The correlation coefficients comparing the modelled and observed data for all three flood statistics are much higher using Eqs. (4) to (6) as compared to when using Eqs. (1) to (3) (compare the correlation coefficients given in the legends for Fig. 4). To some extent this is because the models are being forced to predict less data. However, the marked improvement in correlation for the two sample L-moment ratios is arguably also due to the longer data records used. Also of interest is that whereas Eqs. (1) to (3) show no dependence on land-use, Eqs. (4) to (6) show dependence on plantation land cover and wetland farming land cover.
Nevertheless, Eq. (1) should be considered a more reliable method for estimating median annual maximum river flow rate for the HydroSHEDS catchments because it utilises a much larger set of observations. Range and median values for the river flow statistics, Q med , Q t and Q t 3 , are shown in Table 1 Table 3 shows the KSS values for all catchment characteristics for which the KST (as described in Sect. 3.4) failed when considering all 154 of the HydroSHEDS regions studied. Note that catchment characteristics for which the KST passed are not shown in Table 3,

All catchments
because a passed KST implies that significant statistical differences are not observed when comparing regions with reported flood disasters and regions with none.
The parameters listed in Table 3 have been ranked in order of their KSS values. The higher the KSS value the more sensitive reported flood disaster occurrence is to that parameter. The most important parameter in this context turns out to be built land cover.
Also shown in Table 3 are rank correlation coefficients between each parameter and built land cover (RCCBL). Note that RCCBL is listed as zero where a correlation fails a significance test. Only catchment area, Latosols coverage, and median annual maximum river flow rate are found not to be strongly correlated with built land cover (i.e., with absolute values of RRCCBL < 0.30 ). CDFs, for both HydroSHEDS regions that reported flood disasters and those that did not, for these three parameters along with built land cover are shown in Fig. 5.

Rural catchments
In an attempt to take out the strong control of built land cover on flood disaster reporting, the KSTs were repeated using only HydroSHEDS regions contributed by catchments with less than 3% built land cover, a threshold commonly used to describe rural catchments  Table 4 shows the KSS values for all catchment characteristics where the KST failed when considering the 55 HydroSHEDS regions considered to be draining rural catchments. The most important parameter in this context turns out to be catchment area followed by median annual maximum river flow rate. RCCBL values for the six identified parameters are also shown in Table 4, from which it can be seen that all of these (with the exception of built land cover) are not strongly correlated with built land cover. Corresponding CDFs for these six parameters are shown in Fig. 6.

Flood frequency regionalisation relationships
Although the main objective of this article is to explore catchment controls on flood disaster reporting, the flood frequency regionalisation relationships (described in Sect. 4.1) also warrant further discussion.
Of the relationships obtained using all river flow gauges (i.e., Sect. 4.1.1), only Eq. (1) for the median annual maximum flow rate, Q med , was found to have a reasonable correspondence with the observed data. The expressions for the L-moment ratios, Eqs. (2) and (3), should be dismissed because 10 years of river flow data is arguably inadequate to observe this kind of higher-order phenomena.
Equation (1) shows that Q med increases almost linearly with catchment area, increases exponentially with median annual maximum daily rainfall and decreases exponentially with P10 elevation. This dependence on catchment area and rainfall is very common and is also observed in UK regionalisation relationships (Kjeldsen et al. 2008). The effect of P10 elevation is due to the fact that catchments with low P10 elevation are likely to have outlets closer to the coast and therefore represent some of the larger catchments within the West Java region.
The relationships obtained, using only river flow gauges with greater than 20 years of data (i.e., Eqs. (4), (5) and (6) in Sect. 4.1.2), exhibit much better correspondence with the observed data for all three of river flow statistics studied.
The alternative equation for Q med , Eq. (4) is very similar to Eq. (1) except that P10 elevation is swapped out for plantation land cover. Yulianto et al. (2022) previously hypothesised that runoff coefficient for plantation land cover in West Java might be 40 times greater than that for primary forest. Plantation land cover ranges from 0 to 33.4% with a median of 1.74% for the gauged catchments and from 0 to 30.5% with a median of 2.00% for the HydroSHEDS catchments. Equation (4) suggests that reducing plantation land cover from 2 to 1% leads to a 4.6% reduction in median annual maximum river flow rate. Alternatively, reducing plantation land cover from 20 to 10% leads to a modelled 38% reduction in median annual maximum river flow rate.
Equation (5) for the L-moment ratio, Q t , suggests that the coefficient of variation for annual maximum flow rate increases with increasing rainfall and increasing catchment circularity. Assuming spatially uniform rainfall, a more circular catchment will have more pathways of the same length and hence more water reaching the outlet at the same time, leading to higher peak flow rates. Equation (6) for the L-moment ratio, Q t3 , suggests that the skewness for annual maximum flow rates decreases with increasing wetland farming land cover. The impact of this effect is better understood by studying how Q t3 affects flood frequency. Figure 7 shows plots of normalised annual maximum river flow rate (normalised by dividing by the median annual maximum river flow rate) against return period for two catchments; one with 0% wetland farming land cover (river Cikapundung at Gandok) and one with 20% wetland farming land cover (river Cimanuk at Tomo). The circular markers are the observed data plotted using a Gringorten plotting position (Shaw 2005, p.314). The solid lines are median and L-moment matched generalised logistic distribution functions (see, Kjeldsen and Jones 2006). The black lines were obtained using the sample L-moments of the observed data. The coloured lines were obtained by determining the L-moments from Eqs. (5) and (6) using the associated catchment characteristics but with wetland farming land cover as specified in the legend. It can be seen that increasing wetland farming land cover leads to a flattening of the flood frequency curve at higher return periods. An analogous effect was observed in the revitalised UK flood estimation handbook with flood plain extent (Kjeldsen et al. 2008, p.55). This could be because rice paddies are able to attenuate very large flood events better than other land cover types. The flood control value of rice paddies has long been recognised by ecosystem service studies (Natuhara 2013;Saputra and Setiyanto 2021). Rice paddies store rainfall and reduce peak river flows (Huang et al. 2006). They also benefit from being widely distributed across the landscape, offering widespread protection as compared to sparsely located dammed reservoirs (Kim et al. 2006). Asdak and Supian (2018) previously proposed that maintaining wetland farming land cover in West Java is important to help avoid future increases in the magnitude and

All catchments
Following removal of catchment characteristics with strong correlation with built land cover, Kolmogorov-Smirnov testing revealed that only built land, catchment area, Latsols cover and median annual maximum river flow exhibit statistically significant differences in CDFs for river catchments that report flood disasters. The percentage of built land cover on catchments contributing to regions that report flood disasters is generally much larger (see Fig. 5a). A similar finding was observed in Central Java by Handayani et al. (2020). Built land cover is known to lead to higher surface runoff and hence higher river flows. However, this property was not observed in the two stepwise linear regression exercises described in the previous section. A more important factor is probably that the higher the percentage of built land cover, the more people available to experience and report flood disasters (van Dijk et al. 2009;Ferreira and Ghimire 2012).
The second most important parameter is catchment area. Regions that report flood disasters have statistically larger catchment areas (see Fig. 5b). This could be because river flow rate is strongly correlated with catchment area (consider Eqs. (1) and (4)).

b) River Cimanuk at Tomo
Observed L-moments matched 0% wetland farming 20% wetland farming 50% wetland farming 100% wetland farming Fig. 7 Comparison of observed and modelled normalised annual maximum river flow rate plotted against return period for two different catchments. The circular markers are observed data plotted using the Gringorten plotting position. The black solid lines are generalised logistic distributions with their median value and L-moments matched to the observed data. The coloured lines are generalised logistic distributions, with t and t 3 obtained from Eqs. (5) and (6), respectively, using catchment characteristics but with wetland farming land cover specified as shown in the legend. a Shows data for the river Cikapundung at Gandok, which actually has 0% wetland farming land cover. b Shows data for the river Cimanuk at Tomo, which actually has 20% wetland farming land cover Regions that report flood disasters also have statistically higher Latosols coverage (see Fig. 5c). Latosols have only a small correlation with built land cover (see Table 3). However, Latosols are the main soil-type underlying the highly urbanised region between Jakarta and Bogor (see Fig. 2b), which is strongly affected by flooding from the River Ciliwung (Asdak and Supian 2018).
The final parameter concerns median annual maximum river flow rate. However, regions that report flood disasters seem only to experience marginally higher flow rates (see Fig. 5d). Furthermore, median annual maximum river flow rate was calculated from Eq. (1), which has no dependence of land-use cover. Therefore, it can be concluded that land-use cover, with the exception of built land cover, seems not to have an impact on whether HydroSHEDS regions report flood disasters or not, a point supported by van Dijk et al. (2009) and Ferreira and Ghimire (2012).

Rural catchments
When we consider only rural catchments (with < 3% built land cover), built land cover is much less correlated with other important parameters identified by the Kolmogorov-Smirnov testing (see Table 4). Similar to for all the catchments, rural catchments that report flood disasters are found to have larger catchment areas, larger median annual maximum river flow and more built land cover (see Figs. 6a and b).
Flood affected (rural) catchments are also found to have more alluvial clay cover and lower t 3 L-moment ratio and median value for annual maximum daily rainfall (see Figs. 6d-f). The importance of clay cover could be due to its low permeability and high runoff coefficient and/or that there is substantial clay cover around the Jakarta and Bandung cities (see Fig. 2b). An explanation concerning the impact of rainfall statistics could be due to rural communities being more resilient where extreme rainfall is more common.
The general trend in parameters remains the same as when all the catchments are studied, with flood disaster reporting more likely in regions contributed by larger catchment areas, larger river flows and more built land. van Dijk et al. (2009) and Ferreira and Ghimire (2012) found that country-scale flood reporting data was strongly correlated with population density (a good proxy for built land cover) and inadequate for investigating the impacts of land-use cover on flood disaster frequency. Bhattacharjee and Behera (2018) and Tembata et al. (2020) observed some success at linking flood disaster reporting to forest cover at a district scale. However, our findings further support the assertion of van Dijk et al. (2009) and Ferreira and Ghimire (2012), even at a sub-catchment scale.

Summary and conclusions
The objective of this article was to determine significant statistical differences, in terms of topography, land-use, soil-type and hydro-meteorology, between river catchments in West Java that reported suffering from flood disasters and those that did not.
Regional analysis using data from river flow gauging stations (focusing on the 21 gauge stations with at least 20 complete years of data) led to stepwise linear regression relationships for median and sample L-moment ratios t (similar to coefficient of variation) and t 3 (similar to similar to skewness) of the annual maximum daily mean river flow rate. Median annual maximum flow rates were found to be positively correlated with catchment area, rainfall and plantation cover. Reducing plantation land cover from 20 to 10% was found to lead to a modelled 38% reduction in median annual maximum river flow rate. This is presumed to be because of the increased runoff coefficient associated with converting forest to plantation.
The t 3 sample L-moment ratios were found to be negatively correlated with wetland farming land cover. Using a generalised logistic function it was shown that increasing wetland farming land cover reduced the extent to which annual maximum flow rates escalate with extreme return periods (i.e., greater than 10 years). This suggests that rice paddies play an important role in attenuating extreme river flow events.
A set of Kolmogorov-Smirnov tests (KST) were performed on 34 topographic, land-use and hydro-meteorological catchment characteristics looking at 154 HydroSHEDS regions across West Java. The aim was to look for statistical differences associated with regions that reported flood disasters during the period from 2009 to 2013 and those that did not. Built land cover was found to be the main factor determining whether a region was likely to report flood disasters. Built land cover was also found to be a significant factor in rural catchments (i.e., with less than 3% built land). Interestingly, built land was not identified as a contributing factor in the aforementioned flood frequency regionalisation study. This would suggest that the importance of built land is more about the fact that the higher the percentage of built land cover, the more people available to experience and report flood disasters.
The KSTs did not identify any other land use factors although annual maximum river flows were found to be slightly larger in flood disaster affected areas, which in turn are partially controlled by plantation and wetland farming cover.
Despite flood frequency (in terms of river flow rate) being strongly controlled by plantation cover and wetland farming cover, our research shows that percentage built land cover is the most important factor defining whether or not an area is likely to report flood disasters in West Java. Our findings also suggest that more research is needed to understand the important role of plantation cover in aggravating median annual maximum river flow rates and wetland farming cover in mitigating extreme river flow events.
Author contributions Conceptualization: RR and SM; Methodology: RR, SM, SR and GV; Formal analysis and investigation: RR and SM; Writing -original draft preparation: RR and SM; Writing -review and editing: RR, SM, SR, GV, RS and AR; Funding acquisition: RR, SM, SR, GV, RS and AR.

Funding Information
The authors are grateful for funding received from the UK Natural Research Council, UK Newton Fund and the Indonesian funding agency, Ristekdikti, awarded through the NERC Hydrometeorological Hazards Programme (grant number NE/S00310X/1). Rahmawati Rahayu is also grateful for their PhD scholarship funded by the Durham University Global Challenges Centre for Doctoral Training, which is in turn funded by the UK Global Challenges Research Fund (GCRF).

Conflict of interest
The authors have no relevant financial or non-financial interests to disclose.
Open Access This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article's Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article's Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/.