Introduction

Sustainable development and disaster risk are closely interconnected at various levels, with natural hazards serving as significant barriers to sustainable development1. In 2020 alone, disasters associated with natural hazards affected approximately 100 million people, with global economic losses estimated at 190 billion USD and 15,082 fatalities2. In recent years, factors such as climate change and rapid urbanization have become major drivers of the increasing frequency and intensity of disasters from natural onsets, exacerbating the vulnerabilities of communities at large3,4. Among various types of natural hazards, river floods are one of the most devastating, causing significant human and economic losses on a global scale each year5,6. The United Nations Office for Disaster Risk Reduction (UNDRR) reports that of the 61.77 million people affected by natural hazards in 2018, nearly half—approximately 35.38 million—were impacted by floods7. Furthermore, flooding is directly linked to human resilience deterioration, leading to issues such as food insecurity and poverty expansion, making flood risk management an urgent and critical issue8,9.

On the other hand, rivers and associated floodplains are crucial areas that benefit humans, as supply of resources and foundation for food production10,11. Most human societies have historically settled and thrived along the major rivers across the world12,13. Ancient settlements were often located in places that facilitated access to diverse resources, while simultaneously mitigating the risks of floods and other natural hazards14,15. These communities developed by balancing the environments’ natural risks with derived benefits15,16. However, since the modernization, the development of engineering flood control technologies, such as continuous levees and flood regulation dams, has led to a decrease in the frequency of floods in floodplains17,18. As a result, settlement in areas historically prone to frequent flooding has increased. Notably, the phenomenon of settlement in flood-prone areas, such as floodplains, due to the construction of levees is known as the “levee effect.” This effect has been observed worldwide, where populations continue to increase even after flood events19,20. In addition, the improvement of flood protection levels through the construction of continuous levees has reduced flooding damage in floodplains, which were originally natural flood retention areas with significant inundation. The reduction of flood risks through engineering flood control measures has made it possible for people to settle in floodplains. However, in the event of large-scale floods exceeding the designed level with levee breaches, there is a high potential for significant damage. Traditional flood control methods, which adjusted flood defence levels according to land use by varying the presence and height of levees along the riverbanks or between upstream and downstream sections, have been largely abandoned18. The continuous construction of levees with uniform safety levels has led to a form of “Russian roulette” in terms of flood occurrence, where flooding risks are randomly distributed across different locations. Moreover, the settlement of people in areas with high potential flood risks has made it increasingly difficult to manage flood events that exceed the scale of planned flood protection measures, particularly those involving levee breaches and large-scale inundation21,22.

The increase in extreme weather events due to climate change has raised concerns about the growing risk of flooding23. To reduce the risks of increasingly severe water-related disasters driven by climate change, relying solely on structural flood control measures has certain limitations. Therefore, implementing basin-wide strategies that include the development of flood retention areas, such as floodplains, and promoting land-use planning that allows for flooding in certain areas is essential24. Identifying areas prone to flooding within a basin is crucial for integrated flood management at the basin level. To achieve this, it is necessary to accumulate knowledge not only from hydrological and hydraulic studies but also from geoscientific perspectives, including those focused on basin formation processes and landscape dynamics. Landscape dynamics refers to any alteration in the physical, biological, and cognitive components of a landscape. These dynamics are influenced by external disturbances, including eruptions, earthquakes, erosion, extreme weather events, fires, and human activities25. Landscape dynamics, being the driving force behind landform formation and a trigger for various disasters26,27, are considered to govern the various geomorphic processes, both large and small, that influence the formation of flood-prone areas.

Until now, the advancement of numerical simulation and their application for identifying flood inundation areas and quantifying damages have been the mainstream in hydraulic engineering research28. However, there are challenges when conducting flood inundation calculations, such as uncertainties in the setting of levee breach points29 and the inability to account for upstream flooding when calculating inundation caused by flows exceeding the flood conveyance capacity of the river channel30. Additionally, most of these hydraulic engineering studies target floods with occurrence probabilities over time scales of several decades to a hundred years, typical of flood control planning; however, there has been insufficient examination of large-scale flood disasters with extremely low occurrence frequencies. This requires elucidating long-term basin formation and geological processes that influence them (Fig. 1a). Furthermore, flood management measures in each river basin should be based on an understanding of the long-term basin formation processes and their impact on the formation of flood-prone areas. For this, the current time scales of several decades to a century, which are considered in existing hydrological and hydraulic flood management practices, are insufficient. It is necessary to clarify the relationship between long-term geomorphological processes and flooding. The landscapes we observe are formed by a combination of internal factors (such as volcanic activity and tectonic movements involving the subsurface movement of formations caused by the Earth’s natural forces), external factors (such as erosion and sedimentation), and human influences. The complex processes involved in their formation give rise to the unique disaster characteristics of each basin. However, traditional disaster research lacks the perspective of long-term geological processes, and the relationship between landscapes formation across various timescales and the corresponding disaster characteristics has not been systematically explored.

Fig. 1
figure 1

Relationship between time and spatial scales of earth phenomena (Modified from Kaizuka (1990)67 and Buffington (2012)68).

Full size image

In this study, we aimed to develop a method for identifying locations within a river basin that are prone to flood inundation using flood data from the 17th century onward. The findings of this study clarify the relationship between flood events and long-term geological processes occurring on geological timescales, which is essential for effective flood management.

Location and flood characteristics of the study site

This study focuses on the Japanese archipelago, a region where intense orogeny results in highly active sediment production31,32. Moreover, climate change has a significant impact on sediment production, and fluctuations in sea levels lead to substantial changes in deposition phenomena33. Therefore, Japan can be considered as an ideal location for investigating the influence of long-term geological processes on the formation of flood-prone areas. Additionally, the findings of this research are expected to be applicable to other regions with active orogeny. The study focused on the alluvial plains of 11 rivers in the Japanese Archipelago, for which sufficient data on flood inundation and long-term geological processes have been accumulated (Fig. 2a). Information on flood inundation was gathered by reviewing historical and local documents, which provided data on 432 flood inundation locations that occurred between 1625 and 2000 in the 11 target rivers, where the positions of these events could be identified (Fig. 2b). Using a dataset of 432 collected flood events, statistical analysis was conducted to examine the non-random distribution of flood occurrence locations and to investigate the topographical factors influencing these locations.

Fig. 2: Location of the study sites and temporal distribution of flood inundation event.
figure 2

a The study focuses on 11 rivers across Japan. These rivers range from the Tone River, which has the largest catchment area in Japan (16,840 km²), to the Toyo River, with a catchment area of 724 km², representing a variety of river catchment sizes. b A total of 432 flood inundation records were extracted from the literature, including 42 records with unspecified years. The average number of flood records per river is 38.2 ± 11.6. The highest frequency of flood events occurred between 1850 and 1900, while the lowest frequency was observed after 1950, following the implementation of modern river management measures.

Full size image

The importance of long-term geological processes in predicting flood inundation locations

In this study, indicators of long-term geological processes within the watershed were investigated, including the land-sea boundary during the maximum transgression period, the intersection of the Last Glacial River Long-Profile (LGRP) and Present River Long-Profile (PRP), and geomorphological change points such as mountain ranges, terraces, alluvial fans, meander belts, and deltas. The maximum transgression period in the Japanese Archipelago, which was targeted in this study, occurred approximately 7000 years ago, when sea levels were 2–3 meters higher than those at present. As sea levels increased, the scouring force in the downstream sections decreased, leading to increased sediment deposition. Additionally, due to the variation in shear forces between terrestrial and marine environments, the land-sea boundary is characterized by a change in the riverbed profile, which results from differences in sedimentation rates. Therefore, although the area is currently terrestrial, it is believed that at locations corresponding to the land-sea boundary during the maximum transgression period, sharp changes in the riverbed profile would have been present. During the last glacial sea-level lowstand, factors such as an increase in rock debris production due to freeze-thaw processes, lowering of the forest limit, and enhanced scouring of alluvial plains due to sea-level regression34 contributed to the deposition of basal gravel layers, which form the substrate for overlying alluvial deposits35. Therefore, it is believed that near the intersection of the PRP and the LGRP, changes in the composition of materials that make up the alluvial plains and differences in the propagation characteristics of flood flows make flooding more likely. These factors were included as indicators of long-term geological processes (Fig. 3).

Fig. 3: Changes in the riverbed longitudinal profile and general soil structure of alluvial plains due to sea level fluctuations.
figure 3

During the Ice Age, sea-level lowering increased river scouring power, leading to the deposition of basal gravel. With global warming, rising sea levels and increased precipitation have led to an increase in transported sediment, resulting in the deposition of floodplain deposits in alluvial plains (modified from Kaizuka69, and Yamamoto70).

Full size image

Results

A QQ plot was constructed to test the hypothesis of the non-randomness of flood inundation occurrence points, with the standardized expected values for the distance from the river mouth (theoretical values based on a normal distribution) on the x-axis and the actual distances from the river mouth to the flood inundation locations on the y-axis. The plot revealed that specific locations along the rivers exhibited “jumps,” indicating high irregularity in flood inundation occurrences (Fig. 4a, b). The results of the Shapiro–Wilk test for testing randomness showed that the null hypothesis of normality was rejected at the significance level of p < 0.05 for 10 of the 11 rivers (Table 1). This indicates that the occurrence of flood inundations is not random, but tends to concentrate in specific locations. Based on this, a relationship analysis was conducted between the occurrence of flood inundation and geomorphological factors.

Fig. 4: Distribution of flood inundation occurrence points and characteristics of geomorphological factors.
figure 4

a QQ Plot of flood inundation locations for rivers with a target reach of less than 60 km among the 11 rivers. b QQ plot of flood inundation locations for rivers with a target reach of 60 km or more among the 11 rivers. c Differences in geomorphological factors based on the presence or absence of flood inundation in 1 km sections. Of the total 648 km of river length across the 11 rivers studied, 216 km experienced flood inundation, while 432 km did not. The significance of the difference in mean values was tested using a t-test.

Full size image
Table 1 Results of the Shapiro–Wilk test for testing the randomness of flood inundation locations
Full size table

To identify the impact of long-term geomorphological changes on flood damage, we included independent variables such as the distance from the maximum transgression boundary, LGRP and PRP intersection, and geomorphological change points in our analysis. Additionally, several geomorphological factors considered related to flood inundation, such as the distance from the inflection points of the riverbed profile, sinuosity, braiding index, river width variability, and the presence of microtopography (e.g., old river channels, floodplains, and natural levees), were included. Significant differences in these geomorphological factors were noted between segments with and without flood inundation except for meander (Fig. 4c).

The results of the random forest analysis showed that the out-of-bag (OOB) error rate for the 11 rivers ranged from 14.6 to 41.5% for the river with the lowest OOB to the river with the highest OOB, with a mean ± standard deviation of 31.2 ± 7.9%. Additionally, the area under the Receiver Operating Characteristic (ROC) curve, which represents the area enclosed by the curve and the line with True Positive Rate (TPR) = 0 and False Positive Rate (FPR) = 1, had minimum and maximum values of 0.50 and 1.00 and a mean ± standard deviation of 0.78 ± 0.15 (Fig. 5), respectively. The factors identified as the most important influences on the number of flood inundation locations were the distance from the maximum transgression boundary during the transgression period (mean rank 2.10 ± 1.14) and the distance from the boundary between LGRP and PRP (mean rank 2.50 ± 1.89). In contrast, river channel morphological indicators, such as the braiding index (mean rank 7.82 ± 1.54) and sinuosity (mean rank 5.73 ± 2.15), were determined to have lower importance as explanatory variables (Fig. 5). Therefore, the occurrence of flood inundation locations may be more strongly influenced by geoscientific time scales, such as the position of the sea level during the maximum transgression period and the distance from the intersection of the present alluvial plain and the glacial-period alluvial plain, rather than by river channel characteristics.

Fig. 5: Results of random forest with flood inundation locations as the dependent variable and geomorphological factors as explanatory variables.
figure 5

The figure shows the OOB (Out of Bag) error for each model, the AUC (Area Under the Curve) from the ROC curve, and the ranking of each variable’s importance based on the mean decrease in Gini. DLSBT Distance from the land-sea boundary during the maximum transgression period, DILP Distance from the intersection of the LGRP and the PRP, DGC Distance from the geomorphological change point, DCBS Distance from the change point in bed slope, S Sinuosity, B Braiding index, CRW Changes in river width, FRC Former river channel, FP Flood plain, NL Natural levee.

Full size image

As relatively good predictive accuracy was achieved using the factors for the target rivers (OOB: 31.2 ± 7.9%, AUC: 0.78 ± 0.15), a prediction model for flood inundation locations was constructed using multiple logistic regression analysis for all sections of the targeted 11 rivers. We found that the greater the values of the braiding index, the percentage of old river channels, and the percentage of natural levees in the hinterland, the higher the flood inundation risk. Additionally, the smaller the distances from the inflection points of the riverbed slope, the maximum transgression sea level, and the intersection of the PRP and LGRP, the higher the flood inundation risk. These results suggest that not only long-term geological processes, but also river channel characteristics which are an indicator of topography on a smaller spatial scale, are necessary for predicting flood inundation locations (Table 2). To improve predictive accuracy, it is necessary to consider the influence of factors not addressed in this study, such as an understanding of the intrinsic geomorphological processes of the watershed and human-induced flood control measures. The average value of the balanced accuracy, calculated from the average sensitivity and specificity of the prediction model, was 0.628 for the entire length of the rivers. However, there was variation across individual rivers, with values ranging between 0.320–0.845 (Table 3). This suggests that the factors influencing flood inundation vary by river, and for rivers with high accuracy, the factors considered in this analysis were sufficient to predict flood inundation locations. In contrast, for rivers with lower accuracy, further consideration of additional factors is necessary.

Table 2 Results of multiple logistic regression analysis
Full size table
Table 3 Prediction model accuracy evaluation derived from logistic regression analysis
Full size table

Discussion

Studies on flood inundation characteristics have primarily focused on the interaction between hydrological factors and river morphology36. In particular, regarding river morphology, research has highlighted factors influencing flood inundation, such as the meandering characteristics of river channels37,38, characteristics of braided channels in alluvial fans39, natural micro-topography behind levees18,40, and the backwater effect caused by confluences41,42. In contrast, here we conducted an analysis that included not only traditional local and short-term scale geomorphological factors but also long-term geological processes. The results indicated that factors related to long-term geological events, such as the distance from the transgression maximum coastline and the distance from the intersection of LGRP and PRP, were stronger influencing factors than topographical features such as meandering or braiding index (Fig. 5).

Among the factors identified, the distance from the transgression maximum coastline during the Holocene is considered a key influence. The development of alluvial lowlands is widely recognized as strongly influenced by sea-level fluctuations since the late Pleistocene, including the Holocene43,44. In the case of the Japanese Archipelago, similar sea-level changes to the global average have been observed, with sea-level fluctuations affecting the geomorphological formation of alluvial lowlands. A typical soil structure in Japan’s alluvial plains consists of gravel layers transported by rivers during low sea levels of the Last Glacial Maximum (LGM), overlaid by floodplain deposits. The transgression maximum coastline, which marks the boundary between alluvial deposits and basal gravel, is frequently located in these regions, and the differences in sediment types, as well as the resulting changes in the longitudinal gradient of the alluvial surface, are believed to influence the occurrence of flood inundation.

The distance from the LGRP and PRP intersection was also identified as an important factor in predicting flood inundation locations (Fig. 5). During the LGM, when sea levels were low, riverbeds had steep gradients and carried enough flow power to transport gravel, resulting in surface composed of gravel layers35,45. In contrast, the surface of the present riverbed is formed by river deposits, consisting of finer particles than those of the Last Glacial surface (Fig. 3). Therefore, the intersection of the LGRP and PRP is believed to correspond to a point of change in the grain size of riverbed sediments, with the difference between them likely creating a sharp change in bed roughness. Generally, abrupt changes in bed roughness lead to complex flow characteristics46,47, which can influence sediment transport during floods and potentially act as a discontinuity in water surface profiles. Hence, the intersection of these two profiles is considered a front of geomorphological change, marking a transition in riverbed materials and water surface profiles during floods, making it a location prone to flood inundation. It is believed that there is a temporal and spatial hierarchy between the local and short-term scale geomorphological factors, such as river channel characteristics, and long-term geological processes. Therefore, analysing their relationship is necessary. This analysis has clearly demonstrated that long-term geological processes have a clear impact on flood inundation, highlighting the significance of introducing a new perspective into flood management.

Furthermore, the results from the multiple logistic regression analysis revealed rivers for which flood inundation locations could be predicted using the factors considered in this analysis, and those for which this was not the case (Table 2). A comprehensive examination of the impact of long-term geological processes on river morphology is needed for each watershed. For example, in the case of the Kiso River, a fault is located upstream of the transgression maximum coastline, where the river curves to the south and marks the boundary between the braided channel and meandering belt, resulting in enhanced flood inundation (Fig. 6). In this watershed, such geomorphological changes caused by faults are as significant as sea-level fluctuations. Research on river morphological changes caused by faults has been progressing in the fields of geology and geomorphology48,49. However, its application to disaster prevention engineering is of critical importance. Understanding the impact of geomorphological changes in alluvial plains, as well as how these changes influence disaster characteristics, is essential for understanding the unique flood characteristics of each watershed.

Fig. 6: Distribution of flood inundation locations and topography and faults in the kiso river alluvial plain.
figure 6

The locations of faults are based on Aichi prefecture (https://www.pref.aichi.jp/bousai/atlas.html) and while the topographic distribution is based on Sakamoto et al. 71 and was created by the author.

Full size image

Human influences on river morphology cannot be disregarded. Since ancient times, humans have expanded their settlements while battling flood disasters14,50,51. Human activities such as channel modifications and embankments construction have altered the natural flood inundation characteristics. The results of this study showed that changes in the current riverbed long profile are of low importance as a variable explaining the distribution of flood inundation locations (average rank 5.71 ± 2.43) (Fig. 5). In principle, changes in the riverbed long profile are important geomorphological change points and typically represent discontinuities in the water surface profile during floods, making them locations where flooding is more likely to occur52. However, the riverbed long profile has diverged from its natural state due to factors such as changes in sediment dynamics caused by dam construction, reduced flood opportunities due to embankment construction, and gravel extraction associated with rapid economic growth52,53. This divergence is considered a key factor for the estimation of the flood inundation sites. Particularly, as the breach data used in this study predominantly predate modern river engineering practices, the riverbed profiles at the time of the breaches were likely different from the current profile, which is why these changes were not selected as important indicators. In Japan, which is the focus of this study, flood control measures have been implemented primarily through the construction of continuous embankments, especially in alluvial lowlands where population and assets are concentrated, since the enactment of the River Act in 1896. These measures have significantly reduced the frequency of flooding in floodplains. However, this has led to the non-manifestation of flood-prone areas inherent to rivers and watersheds due to long-term geological processes. Furthermore, the potential for disasters has increased due to factors other than the latent disaster characteristics, such as the probability of embankment failure. Including the impact of such human influences on how flood characteristics in the watershed have changed is crucial for improving the accuracy of flood inundation predictions.

It is clear that factors influencing the flood inundation characteristics of each watershed are related to the geological processes and tectonic movements. Identifying these factors is the first step in flood management. Furthermore, in response to the watershed characteristics formed over long time scales, humans have historically maximized the benefits of these features while minimizing flood damage21. However, in recent times, especially in developed countries, the temporary success of flood control through engineering measures has led to an increased risk of large-scale disasters21,54. As exemplified by the phenomenon known as the levee effect19,20,55, humans technology and resource deployment to mitigate natural threats may increase rather than decrease damage56. This highlights the need for a more comprehensive and resilient approach that does not rely solely on structural solutions. The approach proposed in this study, which incorporates long-term geological processes to elucidate the flood inundation characteristics of river basins, can contribute to minimizing flood damage by identifying areas with high flood potential. We can better manage flood risks by utilizing designated flood zones and guiding land use accordingly. The results of this analysis demonstrate that flood breaches do not occur randomly but at specific locations (Fig. 4a, b). However, flood defence measures, such as the continuous construction of levees implemented in many countries, conceal the river’s inherent potential flood-prone areas, and may actually increase the randomness of flood events. Identifying episodic geomorphological process related to the flood inundation characteristics of each river basin and clarifying their relationship with society makes it possible to move beyond the current flood control measures that rely on engineering solutions focused on specific time scales. This shift could lead to flood management strategies that ensure long-term sustainability.

Furthermore, elucidating the long-term geomorphological and geological events that contribute to the formation of a river basin is essential not only for flood control measures but also for the formation of the basin’s biota. Understanding these processes is critical for ecosystem conservation and restoration. Human activities have led to the rapid biodiversity loss, and it has been highlighted that existing policy frameworks make conservation and restoration difficult57. The decline in biodiversity and the expansion of flood scale and frequency pose significant threats to human sustainability, and the importance of measures addressing both issues is being increasingly recognized58. Long-term geomorphological processes in river basins investigated in this study are events that also relate to the movement and dispersion of biota59. Therefore, these processes significantly contribute to understanding biota formation processes in the target river basins and identifying biodiversity hotspots, making it possible to set goals for the conservation and restoration of river basin ecosystems. Therefore, understanding the long-term geomorphological processes of river basins is the key to bridging disaster management and biodiversity conservation. Based on these insights, the realization of flood prevention engineering that incorporates the natural form of rivers is expected to become a disruptor in building sustainable river basins. Future studies should aim to achieve a deeper understanding of river-specific geomorphological characteristics, including various long-term geological processes such as tectonic movements and volcanic activity to refine and complete a robust methodology.

Furthermore, by elucidating the impact of the identified geomorphological development processes on the formation of disaster-prone areas, it is believed that it will be possible to identify locations with high disaster potential in rivers and watersheds where disaster data is insufficiently organized. Additionally, by identifying areas with high disaster risks, it becomes feasible to implement land-use-based measures, such as utilizing these areas for flood retention. In particular, Japan, the subject of this study, is entering a society with a declining population, and there is ongoing discussion about which areas should be withdrawn from residential use from a disaster prevention perspective60,61. This study aims to provide fundamental insights for large-scale flood disaster countermeasures, which occur infrequently, and contribute to the solution of societal challenges related to population decline by moving away from relying solely on structural flood control measures.

Methods

Data collection

We collected information from 19 scholarly articles and historical/local records for 11 rivers in the Japanese archipelago, focusing on river floods since the 17th century (Supplementary information). The aim was to identify specific locations of flood inundation events. The reason for focusing on the 17th century is that large-scale civil wars in Japan ended during this period, and the emergence of a centralized feudal state led to a more systematic recording of natural hazards.

Each river was longitudinally divided into 1 km segments, from the river mouth to approximately the upstream end of the floodplain. Based on historical flood damage data, the locations of flood inundation records were plotted within these segments. The presence or absence of flood inundation in each 1 km segment was used as the dependent variable for analysis. Additionally, geomorphological factors, including long-term geological processes, were included in the analysis. Among the indicators related to long-term geological processes, the distance from the current river mouth to the land-sea boundary at the peak of marine transgression, as well as information regarding the intersection of the Long-Term Geological and Post-Regression Processes (LGRP and PRP, respectively), were obtained from Honda (2014)62. Topographical data used to calculate geomorphological change points, as well as riverbed longitudinal slope changes, were obtained from publicly available data provided by river management authorities.

Furthermore, since river channel morphology is a major factor influencing the flooding characteristics of rivers, the braiding index, sinuosity and changes in river width were calculated for each 1 km segment of the targeted river. The braiding index63 was used to assess drainage density, specifically by counting the number of braiding nodes within each 1 km segment. For sinuosity, the ratio of the length of the river centreline in each meandering section to the straight-line distance between the upstream and downstream endpoints of that section was calculated. For changes in river width, the rate of change was calculated by taking the ratio of the river width at the centre of each 1 km section to the river width at the centre of the preceding 1 km section, subtracting one from this ratio and taking the absolute value of the result. The river morphology used for these calculations was obtained from maps provided by the Geospatial Information Authority of Japan (https://maps.gsi.go.jp/).

Moreover, since the microtopography in the hinterland of the river shows traces of flooding and is closely related to flood hazards, it was included as an indicator for predicting flood-prone areas. The microtopography in the hinterland targeted in this study includes seven landform types: landfill, old river channels, backmarsh, floodplains, alluvial fans, natural levees, and terraces. Data for the microtopography of each river were obtained from the Landform Classification Map for Flood Control (https://maps.gsi.go.jp/). Regarding microtopography in the hinterland, the proportion of each landform type along both banks of the river within each 1 km section was calculated.

Statistical analysis

This study is based on the hypothesis that flood-prone areas are formed by long-term geomorphological processes specific to each watershed, which makes it possible to identify locations where floods are likely to occur in each river. Therefore, an analysis was conducted at the beginning of the study to verify whether floods occur randomly. A QQ plot was created with the expected standardized value of the distance from the river mouth to the locations of flood events for each river along the x-axis, and the actual distance from the river mouth to the flood inundation sites along the y-axis. To test for randomness, a Shapiro-Wilk normality test was performed, and if p < 0.05, the occurrence of floods was determined to be random.

Next, to confirm whether there are differences in long-term geomorphological processes and river channel characteristics between sections where flood inundation occurred and sections where it did not, a t-test was conducted to compare the mean values of the 216 km flood-prone sections and the 432 km non-flood-prone sections. The variables tested in the analysis were the ten explanatory variables used in the Random Forest model, which is detailed later for identifying the factors influencing flood inundation locations. These ten variables are: (1) Distance from the land-sea boundary during the maximum transgression period, (2) Distance from the intersection of the LGRP and the PRP, (3) Distance from the geomorphological change point, (4) Distance from the change point in bed slope, (5) Sinuosity, (6) Braiding index, (7) Changes in river width, (8) Former river channel, (9) Flood plain, (10) Natural levee.

To identify the factors influencing flood inundation, a random forest analysis was conducted with the presence or absence of flood inundation in each 1 km river section as the dependent variable and geomorphological factors as independent variables. In random forest, using highly correlated variables as explanatory variables may lead to issues in selecting variable importance. Therefore, the correlation coefficients between the explanatory variables for each river was first calculated and variable selection was performed by ensuring that the absolute correlation coefficient R was less than 0.764. Next, random forest was executed using the selected explanatory variables. The importance of environmental factors was evaluated using the mean decrease Gini65, and the model’s goodness of fit was assessed using the out-of-bag (OOB) error66.

Furthermore, to develop a prediction model for flood inundation across all river sections, multiple logistic regression analysis was performed with the presence or absence of flood inundation as the dependent variable and geomorphological factors as independent variables. Multiple logistic regression is a method used to analyse the effects of multiple independent variables on a binary dependent variable. Through stepwise regression, the set of explanatory variables that minimized the Akaike Information Criterion (AIC) was selected to construct the prediction model. The model’s predictions were then compared to observed values to evaluate its performance. All statistical analyses were conducted using R software (version 4.1.0).