Flood Hazard Assessment in Australian Tropical Cyclone-Prone Regions

Abstract

This study investigated tropical cyclone (TC)-induced flooding in coastal regions of Australia due to the impact of TC Debbie in 2017 utilising a differential evolution-optimised random forest to model flood susceptibility in the region of Bowen, Airlie Beach, and Mackay in North Queensland. Model performance was evaluated using a receiver operating characteristic curve, which showed an area under the curve of 0.925 and an overall accuracy score of 80%. The important flood-influencing factors (FIFs) were investigated using both feature importance scores and the SHapely Additive exPlanations method (SHAP), creating a flood hazard map of the region and a map of SHAP contributions. It was found that the elevation, slope, and normalised difference vegetation index were the most important FIFs overall. However, in some regions, the distance to the river and the stream power index dominated for a similar flood hazard susceptibility outcome. Validation using SHAP to test the physical reasoning of the model confirmed the reliability of the flood hazard map. This study shows that explainable artificial intelligence allows for improved interpretation of model predictions, assisting decision-makers in better understanding machine learning-based flood hazard assessments and ultimately aiding in mitigating adverse impacts of flooding in coastal regions affected by TCs.

1. Introduction

Tropical cyclones (TCs) are highly devastating severe weather phenomena, with potential impacts of high mortality, major economic loss, and infrastructure destruction. Between 1970 and 2019, TCs were responsible for a third of all weather-, water-, and climate-related deaths and economic losses worldwide [1]. The main hazards associated with TCs are destructive winds, storm surges, and torrential rain, which can cause flooding and landslides. With the adoption of the Sendai Framework for Disaster Risk Reduction, there has been increased focus on understanding the risks of hazards potentially leading to disasters, including TCs [2].

Risk assessments are crucial for resource and funding allocations in risk mitigation efforts against severe weather events such as TCs. The prevailing risk assessment framework within the literature is the risk triangle, which includes three aspects of risk: hazard, exposure, and vulnerability [3,4]. Hazard is a physical process that can cause harm and is defined by the United Nations Office for Disaster Risk Reduction (UNDRR) as “a process, phenomenon or human activity that may cause loss of life, injury or other health impacts, property damage, social and economic disruption or environmental degradation” [5].

Compound hazard events occur when a location experiences more than one hazard at once, which has a larger impact compared with single hazard events combined. In the case of TCs, in coastal regions, the compound flooding from extreme rainfall and storm surges makes the regions particularly vulnerable to TC impacts [6]. TC-related extreme rainfall can cause pluvial (flash) flooding and create downstream riverine (fluvial) flooding; both can cause deaths and damage to infrastructure and the environment [7,8].

With anthropogenic climate change, there are projections of a global decrease in TC frequency but an increase in TC intensity [4]. Compounded with a projected increase in TC precipitation rates globally, this will result in increased exposure to TC-related hazards, and thus TC impacts can be expected to be more severe in the future [9,10]. With the immense impact on Australian TC-prone regions, it is crucial to advance our understanding of TC-related hazards. This study focused on flood hazard assessment and mapping in the TC context.

Modern flood hazard assessment and mapping methods can be divided into two types: physically based flood models and empirical models. In physical models, also called hydrodynamic models, the flood is modelled mathematically utilising known physical processes, whereas empirical models utilise past flood location data to create a model to estimate flood susceptibility [11]. For the purposes of this study, flood susceptibility is the predisposition of a location due to its physical attributes that determine its propensity to flood.

Physically based models, also called rainfall–runoff models or hydrological models, use physical equations based on momentum, conservation of energy, and conservation of mass to model flow [11]. They can function in one, two, or three dimensions, and they tend to have increased accuracy compared with statistical (non-physical) models [12]. Whilst they have a high data input requirement, they are able to take into account spatial variability within a catchment [13]. The downfalls of physically based models are that they are computationally intensive and complex and require expertise to implement [14]. R An earlier study by Antwi-Agyakwa et al. [15] described these models as being unusable in many locations, with disadvantaged regions particularly affected by a lack of readily available data. One example of this is in central Australia, where the observation network is sparse, and thus parameters for these models can be inaccurate [16].

Contrary to modelling the physical processes of an actual flooding event, empirical flood hazard methods utilise past flooding events and associated data to create an inferential model. Within empirical flood hazard methodology, there are traditional statistical models, such as logistic regression and frequency ratio [17]. However, these methods commonly assume linearity, whereas flooding follows a non-linear pattern.

Another method utilized is multi-criteria decision-making (MCDM), which assigns weights to flood-influencing factors (FIFs) [11]. MCDM methods, such as the analytical hierarchy process [18], fuzzy logic [19,20], and analytic network process [21], are used to rank FIFs’ propensity for flooding. Whilst MCDM is useful for making decisions within the flood hazard assessment, with MCDM specifically created to handle the uncertainty of complex flood hazards, the methods can result in biased outcomes [11]. Quite commonly, an expert opinion is used to judge the accuracy of the weightings of FIFs [18,22,23].

In recent years, empirical pattern recognition/machine learning (ML) models have become popular for mapping flood hazards [24,25,26]. These models utilise past data and geographically specific attributes, also referred to as FIFs, to train an algorithm to infer if a location may be flood-prone. Widely used ML models are artificial neural networks (ANNs), support vector machines (SVMs), decision trees (DTs), and the adaptive neuro fuzzy inference system (ANFIS) [27,28,29].

ML models typically output the probability of a place being flood-prone according to the training dataset [30]. Commonly, 70% of the input dataset is used for training, leaving 30% for validation [31,32]. Regarding limitations, ML models are only as accurate as their input data, which may contain measurement or other errors, or the dataset may be an inappropriate size, resulting in the model being incorrectly trained [27]. Additionally, a dataset with a limited size can cause ML models to struggle to extrapolate and predict values outside or on the spatial border of their training range [25,33]. However, both [17] and [34] found that ML approaches were more accurate than traditional statistical and MCDM methods.

Data scarcity is one of the largest limitations of flood hazard assessments. With extreme events occurring rarely, and with in situ data collection during a flooding event being exposed to the hazard itself, flood inundation data can be challenging to survey [25]. Remote sensing and data from aerial spectral imagers have become prevalent in mapping floods. However, remote sensing only provides a flood or no flood value at any point, giving no indication of flood depth or severity [31,32,35].

Current flood modelling literature prioritises creating “cheaper, faster and more accurate models [25]” to increase their usability and decrease their drawbacks. These different model architectures can be a combination of ML with statistical techniques, MCDM techniques, optimisation algorithms, and other ML algorithms [36]. These techniques aim to increase the efficiency and accuracy of flood susceptibility mapping models [37,38].

Used in both physical and ML modelling scenarios, an ensemble structure decreases the drawbacks of single models. Ensembles aggregate outputs of multiple independent models run in parallel [35,36]. These ensemble models act to increase accuracy and reduce variance compared with a singular model [39]. A well-established ensemble model in flood susceptibility mapping is the random forest (RF) model, an ensemble model of decision tree ML algorithms. RF is considered robust and able to deal with complex data [26,36]. A particular advantage of RF is its ability to replace missing predictor values and maintain accuracy [40].

With data selection and availability being key factors in effective ML models, FIF selection must have a physical connection to flood susceptibility. While different studies use different FIF predictors, it should be noted that elevation and slope are always used [27,36]. An earlier study by Tehrany et al. [41] found that elevation, the terrain ruggedness index (TRI), and the slope and stream power index (SPI) were selected by their decision tree model as the most impactful, whereas [27] found the normalised difference vegetation index (NDVI), distance to stream, elevation, and lithology to be the most impactful. The range of impacts of FIFs is due to the different conditions of flooding in each study area, the flood inventory data, and the different models utilised. In this study, datasets that highlight the topographical, hydrological, and land-use features were desired.

The ML studies described all lack a significant component: model explainability. ML and AI algorithms are referred to as “black box” models, as the factors contributing to their outputs are not always intuitively understood. The opacity of model decision-making has reduced decision-makers’ trust in the outputs of such models and has led to hesitancy in their use [33]. Explainable AI (XAI) is a way to explain the impact of features within each model on the model’s output. XAI aims to create a way to explain the inner decision-making of specific outputs of an ML model. The current XAI algorithms utilised are local interpretable model agnostic explanation [42], neural-backed decision trees [43] and Shapely Additive exPlanations (SHAP) [44].

SHAP, a game theory-based algorithm, was developed to determine the contribution of an individual player in a collaborative game to an outcome [45]. It explains the approximate contribution of each feature to the final predicted outcome of the model. It is model-agnostic and has been used in recent years to describe earthquake damage models [46] and drought prediction models [47].

Explaining the inner workings of ML models is paramount to trust in model outputs and critical assessments of its predictions [33]. To the best of our knowledge, [48] is the only flood hazard mapping study that has utilised SHAP for an XAI model. In the present study, SHAP was employed to differentiate decisions over the large study region to identify significant features in different locations utilising a novel method of comprehensively mapping SHAP values. This can inform decision-makers and add to the robustness of models.

Additionally, within the Australian context, few ML methods have been utilised for flood hazard modelling [11]. This study investigated flood susceptibility in a location that has not been analysed with ML methods utilising uniquely comprehensive remotely sensed flood mapping data.

The aim of this study was to assess the flood hazard susceptibility in Australian TC-prone coastal regions and to create a flood hazard map. To achieve this, an RF ML model was trained, optimised, and validated on the flooding event during TC Debbie using relevant FIFs to identify the flood susceptibility of the location. Using the SHAP method, the impacts of the FIFs were used to explain the flood hazard map and critically assess the flood hazard assessment.

This paper is organized as follows. Section 2 introduces the study area, data sources, and methodology of this study. Section 3 describes the results of the flood susceptibility model, and Section 4 explains and describes the impact of the findings and the implications for further study.

2. Materials and Methods

2.1. Study Area

On average, the Australian region experiences 11 TCs during TC season, which typically lasts from November to April, with approximately four crossing the coast [49]. Landfalling TCs are particularly dangerous and destructive to coastal regions. On 28 March 2017, TC Debbie made landfall over Airlie Beach, causing costs of up to AUD 3.5 billion in damages and fourteen deaths, most of which were flood-related [50]. TC Debbie was one of the most dangerous cyclones to impact Australia since TC Tracy in 1974 [51]. The study area selected for this research encompassed the coastal region in Queensland impacted by the landfall of TC Debbie (Figure 1).

The major populated areas within the selected region were Mackay, Airlie Beach, Proserpine, and Bowen. In these areas, flooding was reported and mapped, with nearly 100% power outage in Proserpine and Bowen and many roads closed in the region. The region is characterised by the Coral Sea bordering the east and the Great Dividing Range to the west. It encompasses the Proserpine River Basin in the centre, portions of the Don River Basin in the north, and the Pioneer River Basin in the south. The study area had an area of 14,001 km², a maximum elevation of 1235 m, and a minimum elevation of −18 m.

2.2. Data

2.2.1. Maps of Flooded Areas

In any case in which ML is used, robust data about the target variable, in this case flooding, are required to effectively train and test the model. Traditionally, flood maps during TC events are challenging to obtain, as most remote sensing instruments have difficulties in observing the terrain through dense TC clouds. This study utilised data from Copernicus Emergency Management Service Activation EMSR200 [52] rapid flood mapping using synthetic aperture radar (SAR) on 29 March 2017. A benefit of SAR is its utilisation of microwaves at spectra that show little absorption by water vapour, and therefore, SAR can obtain images of terrain under dense cloud cover. These mappings were from data from RADARSAT-2 and COSMO-SkyMed satellites at a scale of 1:80,000. The landfall of TC Debbie occurred on 28 March 2017; however, major rainfall was recorded between the 26th and the 29th of March in this region [53]. The mapping data utilised in this study coincided with peak water heights at gauges within the study area. The flood data and the mapped regions are described in Figure 2, and the flooded area covered a large portion of the coastal study region. These floods caused widespread damage to property, infrastructure, and agriculture and caused road closures in the region [54].

2.2.2. Flood-Influencing Factors

The flood-influencing factors (FIFs) in this study were selected on the basis of the literature and their availability within the study region. Maps of each FIF are presented in Figure 3, and the data sources and original resolutions can be found in

Table 1. Data sources.

Flood Influencing Factor	Dataset	Source	Original Resolution	Year
Elevation	SRTM-derived 1 Second Digital Elevation Model [55]	Geoscience Australia	1 s (30 m)	2009
Slope	Hydrologic Derivatives for Modelling and Analysis [56]	USGS	1 s (30 m)	2017
Terrain Ruggedness Index	SRTM-derived 1 Second Digital Elevation Model [55]	Geoscience Australia	1 s (30 m)	2009
Stream Power Index	Hydrologic Derivatives for Modelling and Analysis [56]	USGS	1 s (30 m)	2017
Normalised Difference Vegetation Index	AVHRR derived daily NDVI CDR	NOAA	0.05 degrees (5 km)	2017
Distance to River	Major watercourse lines	Queensland Open Data Portal	1 s (30 m)	2022
Topographical Wetness Index	Topographic Wetness Index	CSIRO	1 s (30 m)	2016
Soil Moisture	Australian Water Outlook Historical Root Zone	Bureau of Meteorology	5 km	2017
Land Use-Land Cover	Land use of Australia 2015–2016	Australian Bureau of Agricultural and Resource Economics and Sciences	250 m	2016

.

• Elevation. Typically, a lower elevation creates conditions for greater flood susceptibility, as water will flow downhill and pool in lower elevation areas. Elevation is a frequently used FIF, and within this study area, there is a large swath of coastal low-lying regions.
• Slope angle. A lower slope angle implies flatter ground, allowing water to collect and pool, increasing flood susceptibility compared with steeper ground where water flows downhill. Typically, areas at a low elevation with a low slope are prone to riverine flooding [41].
• Stream power index (SPI). SPI is a measure of the erosive strength of a stream. This can indicate flow paths over terrain and flow accumulation areas. The SPI is calculated with Equation (1):

  $$SPI=A\tan B$$

  (1)

  where A is the flow accumulation into each cell and B is the slope angle [57].
• Topographical wetness index (TWI). The TWI is a measure of the relative accumulation of water within an area in the context of the whole catchment. The TWI is used to quantify the topographical effect on hydrological processes and is a predictor of water accumulation at a location [41]. Developed by [58], the TWI is described by Equation (2):

  $$TWI={\log }_e\frac{A}{S}$$

  (2)

  where A is the specific catchment area and S is the local slope angle.
• Terrain ruggedness index (TRI). The TRI is the difference in elevation between a central cell and the surrounding eight grid cells. These values can show cells in which water may pool. The TRI was calculated using the QGIS ruggedness function, which utilises the algorithm from [59] using elevation data.
• Distance to river (DtR). Regions in closer proximity to rivers are usually prone to flooding. When there is extreme rainfall upstream or in the vicinity of the river, the banks of the river overflow. This is called riverine flooding. Major river systems were rasterised, and then a distance-to-river matrix was created using QGIS, with the raster only clipped after the distance-to-river matrix was created to ensure that river systems outside the bounds of the study area were also included.
• Soil moisture (SM). The SM data in for this study comprised the absolute root zone (1 m surface depth) soil moisture in per cent volume one week prior to the landfall of TC Debbie (22 March 2017). Soil moisture is not commonly used as an FIF in ML models; however, it is known that an increase in antecedent soil moisture increases the severity of flooding [60]. The susceptibility of locations increases because the soil is already close to saturation, and thus less rainfall is needed to fully saturate it, at which point overland flow and flooding occur.
• Normalized difference vegetation index (NDVI). The NDVI is a remote sensing analysis of the greenery of an area and can indicate the density of vegetation in an area. Some studies have found an inverse relationship between vegetation density and flooding; in bare lands with low vegetation, there is no control over the rapid flow of water over the ground [61]. Conversely, some studies have found that the NDVI has the opposite impact, indicating that greener regions will have higher rainfall and thus tend to have higher flood susceptibility [62]. The NDVI is described by Equation (3):

  $$NDVI=\frac{NIR-R}{NIR+R}$$

  (3)

  where NIR is light reflected in the near-infrared spectrum and R is light reflected in the red range of the spectrum. For the purposes of this study, the daily derived NDVI was used from 21 March 2017.

2.2.3. FIF Pre-Processing

Working with datasets of different resolutions and extents, each dataset needed to be pre-processed to have the same meta-attributes. All rasters were reprojected to the EPSG:4283 projection. To align all the FIFs, all rasters were resampled to a 1 s resolution using a bilinear process (except LULC, which utilised a nearest-neighbour process due to its discrete nature) and were clipped to the study area.

2.2.4. Selection of Data Points

Data points were selected with the assumption that non-flood mapped regions did not have flooding. A total of 1360 training data points were randomly selected, with 686 flooded and 674 non-flooded locations. Utilising a commonly used ratio of 70:30 [63], 575 testing points were sampled at least one kilometre away from the training points, with 296 flooded and 281 non-flooded locations selected. The flooded data points are shown in Figure 4; refer to Appendix A for further maps of the sampled points.

2.3. Method

2.3.1. Multicollinearity Feature Selection

Whilst many ML models are built to assume correlation between features, correlated data reduce the efficacy of feature importance measures, as when one feature shares information with another, one may be chosen to describe the flooding and the other may be ignored. Collinear data are data that correlate with each other. For this, a variance inflation factor (VIF) test was completed. VIF describes the amount of the variance of one feature that can be described by the other features together. A VIF of 1 implies no correlation, and a VIF above 10 is usually considered a high correlation; thus, data with a VIF of 10 or above were removed [64]. Tolerance and VIF are described by Equations (4) and (5) respectively:

$$TOL=1-R^2$$

(4)

$$VIF=\frac{1}{TOL}$$

(5)

where R² is the coefficient of determination in a multiple regression of all the other features.

Whilst correlated data do not affect model predictions, they reduced the efficacy of the explainable aspects of this study. This is due to the algorithm learning that it can use one feature to describe the same thing another correlated feature describes, thus consistently choosing to use just that one feature and ignoring the other one, which skews the overall feature contribution assessment. The VIF was calculated using the statsmodels library in Python.

2.3.2. Random Forest

Random forest (RF) is a widely used ML algorithm [65]. RF is an ensemble algorithm of decision trees (DTs); as described by [66], through bootstrapped sampling (sampling with replacement) for the training set of each DT and random feature selection to decide the split of samples at each node, the aim is to create the decision trees so that they are not correlated with each other. This reduces overfitting and allows for the modelling of complex relationships. When trained, an input of FIFs of a single location traverses each tree, resulting in a classification from each DT, and then the average of each tree’s output is the RF’s prediction. The RF algorithm and all metrics were created utilising the scikit-learn Python library.

2.3.3. Model Evaluation

Model metrics are utilised to measure a model’s ability to learn from training data and predict outcomes in unseen data. To ensure that the model was robust in this study, the overall accuracy (OA) and receiver operating characteristic (ROC) curve were used as metrics for performance measures. OA is described by Equation (6):

$$Overall Accuracy \left(OA\right)=\frac{Number of correct predictions}{Total number of predictions}=\frac{TP+TN}{TP+FP+TN+FN}$$

(6)

where true positive (TP), false positive (FP), true negative (TN), and false negative (FN) describe the classification of each point in the dataset as 1 or 0, corresponding to correct or incorrect.

An ROC curve is a curve plotting the TP rate against the FP rate at different thresholds of classification. The area under the ROC (AUC) is a commonly used measure of the model’s ability to discern between flooded and non-flooded classes. This metric indicates the model’s overall ability to ordinally rank flooded and non-flooded points, and thus it is a good indication of the model’s ability to classify the study region in this context.

2.3.4. Differential Evolution Hyperparameter Optimisation

Differential evolution (DE), developed by [67], has been used in the past as an effective strategy for hyperparameter optimisation for RF algorithms, balancing the exploration and exploitation of the search space. As described in [68], the DE optimisation works by creating a population of “models”, training them, and then assessing them on the basis of an objective scoring function. Then the population is iterated each generation. The RF parameters of the previous generation are combined and randomised to explore better parameters whilst keeping aspects of strongly performing models. In this study, the objective scoring function was the mean of a 5-fold stratified cross-validation on the training dataset of each RF created. Fivefold cross-validation is the splitting of the training set into 5 parts and iteratively training on 4 of the 5, and testing on the fifth was left out of the training. This metric assesses the ability of the model to learn across the training dataset.

The differential evolution class in the SciPy library was utilised with the parameters of maxiter = 200, popsize = 20, recombination = 0.7, and mutation = (0.5,1). The hyperparameters that were optimized, with respective search spaces, were max_depth from 2–10, n_estimators from 10–1000, min_samples_split from 2–10, min_samples_leaf from 1–10, and max_features from 1–8.

2.3.5. Shapely Additive Explanations

SHAP values are the average marginal contribution by each FIF over every possible permutation of the features. Equation (7) describes the overall SHAP contribution of each feature in the model:

$${\mathsf{\varphi }}_i=\frac{1}{N!}{\displaystyle \sum }_{S\subseteq N\setminus i}\frac{\left|S\right|!\cdot \left(N-\left|S\right|-1\right)!}{N!}\left(v\left(S{\displaystyle \cup }i\right)-v\left(S\right)\right)$$

(7)

where ϕ_i is the contribution of FIF i, N is the set of all features, n is the number of features in N, S is the subset of N containing feature i, and v(N) is the base value, i.e., the predicted outcome for each feature in N without knowing the feature values [48].

SHAP values are helpful for the contextual analysis of feature contributions for pixel-level predictions. SHAP values were utilised to create force plots, pixel-level transparent explanations of the models’ outputs. Additionally, SHAP values were mapped over the study region at 10-cell horizontal intervals and then duplicated to encompass the subsequent 10 cells because of major computational constraints due to the large study area. As SHAP is calculated over the permutation of features, the computation increases at a rate of N!, where N is the number of FIFs. SHAP values were created utilising the shap Python library using specifically the TreeSHAP explainer, an algorithm that explains tree-based models, such as RF, efficiently. The additional explainers are out of scope and are described in [69].

2.3.6. Flood Hazard Mapping

The flood hazard map was created using the RF predict function on each cell in the study area using the FIF datasets. Each cell was predicted as a number between 0 and 1 and was then split into four classes, as described in

Table 2. Flood hazard classification.

Class	Value
Low	0.0–0.25
Moderate	0.25–0.5
High	0.5–0.75
Severe	0.75–1.0

.

The methods described in Section 2.2 and Section 2.3 are outlined in Figure 5.

3. Results

3.1. Multicollinearity Feature Selection

The VIF measures the correlation of each FIF in the training dataset. As shown in

Table 3. VIF results.

Features	VIF	Tolerance
Irrigated Agriculture	7.62	0.13
Infrastructure	3.01	0.33
Intensive Agriculture	1.14	0.87
Estuary	2.73	0.37
Marsh	5.29	0.19
Dryland Agriculture	22.29	0.04
Forestry	1.84	0.54
Natural Environment	27.25	0.04
Elevation	2.10	0.48
Slope	10.13	0.1
TWI	2.01	0.5
SM	1.20	0.83
NDVI	1.24	0.81
TRI	9.95	0.10
DtR	1.33	0.75
SPI	1.02	0.98

, the LULC classes of natural environment and dryland agriculture had VIF values of 22 and 27, respectively, indicating a high correlation with other features. Additionally, the natural environment LULC class was also observed to be a proxy indicator of urban areas, and thus, it gave an indication of regions that were mapped for flooding (urban areas and their surroundings).

Lastly, the major increase in cardinality (dimensions) of the dataset by one-hot-encoding the categorical LULC reduces the efficacy of the explainable SHAP aspect of this method, as the computational power needed increases exponentially with dimensionality. Thus, LULC was removed from the model. Whilst the slope and TRI were correlated, they include contextual topographical information about each location and were not assumed to be independent.

3.2. Hyperparameter Optimisation

The optimisation of hyperparameters is a balance of ensuring the model learns as much as possible from the training data (stopping underfitting) whilst ensuring that the model does not overfit and lose its generalisability. The DE algorithm was run as per the specifications in Section 2.3.4. The outcome was n_trees = 127, max_features = 5, maximum_depth = 7, min_samples_per_split = 5, and min_samples_per_leaf = 3, with the best score of 0.935 fivefold cross-validated on the training dataset. This means that within the RF model, 127 trees were created, and each node in each tree used at most five features to split the data at a maximum depth of seven nodes. The hyperparameters were assessed to be appropriate and were used for the final model.

3.3. Model Evaluation

The scoring of the model allowed for the assessment of its ability to learn the rules of flooding within the study area whilst ignoring the noise within the data. On the testing set, the model had an OA of 0.802, indicating a high level of accuracy on the unseen data. The ROC curve in Figure 6 shows high skill at differentiating between flooded and non-flooded points. On the curve, the FP rate is plotted on the x-axis, and the TP rate is plotted on the y-axis at each classification threshold. The AUC of 0.925 is high and shows the robustness of the model.

3.4. Feature Importance

Conventional feature importance values are measured through the mean decrease in impurity feature importance, a measure of the impact of each feature over the RF model’s trees. As shown in Figure 7, in the model, the most important feature was elevation. The low-lying coastal regions were the most flood-prone, and the most flooding occurred during the TC along the coast. The slope, NDVI, SM, and DtR were also significant contributions. The TWI, SPI, and TRI appeared to have minimal overall contributions.

3.5. Flood Hazard Mapping

The final flood hazard map is shown in Figure 8. The flood hazard map of the study area was divided into four classes, as indicated in

Table 4. Flood hazard classification area.

Hazard Classification	Area (km2)	Percent of Total Area
Low	11,195.4	78.9
Moderate	1198.6	9.0
High	890.4	6.7
Severe	717.0	5.4

: low, moderate, high, and severe, with 78.9%, 9%, 6.7%, and 5.4% of the area, respectively, classed as such. Along the low-lying coastal areas of the study area, the flood susceptibility was the highest. There was a trend of higher susceptibility along rivers, as can be seen in the centre and in the south. The susceptibility decreased quickly to low in the mountainous regions in the west of the study area. The confidence in the high and severe flooded class was extremely high, with a high TP rate and low FP rate, as indicated in Appendix B.

3.6. SHAP Analysis

Zooming in on specific points, the SHAP force plots demonstrated the different contributions of features to local cell-level predictions. As shown in Figure 9, the force plots were read by starting at the base value—the expected output of the model, which was approximately the average of the training data—and then additively “forcing” the base value up (down), increasing (decreasing) the flood susceptibility. The SHAP values and the base value were added together, resulting in the predicted value, indicated by f(x). This relationship is also described by Equation (8):

$$f\left(x\right)=b+{\displaystyle \sum }SHAP\left(x_i\right)$$

(8)

where f(x) is the model’s output, b is the base value, and SHAP(x_i) is the SHAP contribution of each feature i.

The force plots in Figure 9 are labelled with the corresponding reference points in Figure 8. The base value of the model was 0.505. The size of the arrow below the number line is the SHAP value, with red contributing to increased flood susceptibility and blue contributing to a decrease in flood susceptibility. For example, at point A, described by the top row of Figure 9, the low elevation of 3 metres had a large contribution, increasing the susceptibility of the location, whereas the location 0.27 degrees (28 km) away from a river reduced the susceptibility. The addition of each feature’s contribution yielded a prediction indicated by f(x) of 0.88, a susceptibility in the severe class. At point B, the second row of Figure 9, a location near the town of Proserpine, there was a moderate impact of the elevation, slope, DtR, and NDVI (vegetation density), yielding a severe susceptibility of 0.85. In the force plot for points C and D, which were further inland, DtR dominated. Inland in the south of the study area, elevation stopped dominating as the key contributing factor to flood susceptibility, and a low DtR and a high SPI took on a larger role near the flow paths. This contrasted with Figure 7’s overall permutation contribution scores compared with cell-specific predictions.

SHAP can also be used to explain overall trends of feature importance within the study area. Similar to the force plots, Figure 10 shows a map of the elevation, DtR, NDVI, and SM SHAP contributions to susceptibility but over the entire study area, with their subsequent test dataset dependence plots. The accompanying dependence plots show the SHAP value for each respective feature at selected flooded and non-flooded points, allowing for a contextual analysis of the maps. The SHAP value’s spread in these plots is due to the interaction effects with other FIFs.

Elevation was seen as a significant factor in the model’s prediction across most of the study area, excluding, specifically, inland in the south and many locations in the 20–100 m elevation indicated in the dependence plot. The dependence plot shows a sharp linear decrease in SHAP contribution between 0 and 100 m and then flattens to a range of contributions between −0.2 and −0.4. DtR had a major impact along the major rivers, and, as demonstrated in force plot D, the SPI values were high near rivers and could also contribute significantly. There was a non-linear relationship between DtR and its SHAP contribution: as DtR increased, there was a threshold where the contribution started to significantly decrease. It was also clear that in the north, DtR had a minimal contribution, as shown in force plot A. The dependence plot shows clustering at a distance below approximately 0.07 degrees (7 km), above which there is a spread of SHAP contributions that decrease susceptibility.

NDVI also had a significant contribution to reducing susceptibility in the south coast of the study area, a region with a low NDVI. The NDVI dependence plot has a hyperbolic-shaped relationship with its SHAP values, with a spread between NDVI values of 0 and 0.1, and a variance between −0.05 and 0.05 for higher NDVI values. SM had a marginal impact overall, with a significant impact reducing susceptibility in the south and in the north of the study area. The dependence plot shows a consistent variance around −0.1 to 0.1, with a clear maximum at around the 0.7 SM value. Some specific interactions caused it to have a major effect on decreasing the prediction, as can be seen by a few major outliers.

4. Discussion

In this study, the flood susceptibility of the landfalling region of TC Debbie was modelled using the flood data of the event. Using a DE-optimised RF model, nine FIFs and 988 flooded data points were utilised for the development of this model. This resulted in a fivefold cross-validation OA on the training dataset of 93.5%, a testing OA of 80.2%, and an AUC-ROC of 0.925, demonstrating this model’s robustness. The model was explained using the SHAP method, allowing for transparency of the model decision process. The plots introduced were the SHAP force plot and SHAP map, a map of SHAP values of a specific feature within the study area. These allowed for the analysis of the significant influencing factors on a pixel-level scale and improved the transparency of the model’s decision-making process.

Placing this study in the Australian context, past studies in the region have achieved comparable results. Mapping flood hazard along the Brisbane River, [70] used an artificial neural network, a deep learning neural network, and a particle swarm optimisation–deep learning neural network and found that elevation and DtR were the most significant influencing factors. An earlier study by Tehrany et al. [41] also used DT and SVM along the Brisbane Catchment to analyse significant FIFs and found that adding additional factors did not increase accuracy, and they found that the elevation, TRI, slope, and SPI were the most significant factors. This study’s findings that elevation, slope, and DtR were significant factors align with past literature within the region.

The differences in the important factors found are likely to be due to differences in the study areas, particularly as previous studies used the Brisbane River. As will be further discussed in Section 4.3, SPI does have a significant impact on susceptibility in certain circumstances; however, the impact of TRI is extremely limited.

To the best of our knowledge, the intricacies of feature importance on a pixel level using XAI have not been analysed in any other studies in Australia, and the only known study in which flood susceptibility is explained in an ML context is [48]’s study in Republic of Korea.

4.1. Flood Hazard during a Tropical Cyclone Event

The flood hazard map of the study area is a core contribution produced by this study. It showed a large portion of the low-lying coastal regions having a high or severe susceptibility to flooding, and some regions close to rivers also had high flood susceptibility.

Using TC Debbie as a case study, flooding occurred across the study area, with a particularly high inundation region in the centre of the study area, near Proserpine, and in the north near Bowen. The map appeared to capture most of these flooded regions, indicating that the low-lying regions along the coast and near major rivers are susceptible to flooding. Some relatively smaller flooded regions were not assessed as susceptible regions, indicating that the low flood class still holds a level of susceptibility; however, there were no major flooded regions observed in the low class.

Along the coast between Mackay and Airlie Beach, there were regions that were not flood-mapped but were still classed as having severe flood susceptibility, indicating that there may have been some unrecorded flooding in these regions.

The regions around the town of Proserpine had the most areal inundation and were in the most susceptible locations in the flood hazard map. During this event, the catchment had an event total rainfall of 621 mm and set a record for the highest daily March total on record [53]. This was an extreme event in the catchment, and the extent of the flooding was clear within the flood inventory.

Interestingly, no flood levels were observed along the Proserpine River, but flood extents were observed in the remote sensing dataset utilised in this study. This region had the highest inundated area, aligning with the map’s assessment of the region being extremely susceptible. In the north, around Bowen, major flooding heights were observed on the ground at the Don River and through remote sensing. These findings align with the map’s assessment of the region. In the south, flooding was observed at river gauges along the Pioneer River further inland, but no flooding was observed near the coast near Mackay. These findings align with the map’s assessment of the region, as the further inland regions had a higher susceptibility than the coastal region.

The flood hazard map also indicated regions that may be susceptible in more extreme events. A large portion of the high- and severe-classed regions did not actually flood during this event, but their classification indicates that they have conditions that are conducive to flooding. Given that TC events are projected to become more intense, in a future event with similar antecedent conditions and a greater or more intense amount of rainfall, a greater flooding extent covering these high and severe classes may be expected compared with the TC Debbie event.

4.2. Model Validation

The cross-validation of models is a crucial part of assessing a model’s robustness to new data. Within the academic community, there is a conversation around the cross-validation of datasets when creating maps. Some researchers contend that training and validation data should ensure that the data are not spatially autocorrelated [71,72]. Points that are closer to each other are typically more similar than points that are distant. However, there are also those who propose that spatial cross-validation is grossly pessimistic about model performance. As described by [73], there is a paradox between excluding data that are spatially autocorrelated and avoiding extrapolation within the geographical extent and the flood susceptibility model. Whilst there appears to be merit to both using spatially independent and non-spatially independent validation data, in this study, it was decided to use random cross-validation techniques with a 70:30 balanced dataset to assess the mapping model’s accuracy, similar to other studies of its kind. This may have contributed to an overoptimistic accuracy score; thus, a final validation dataset with a ratio of 1:50 (flooded:non-flooded) dataset, the proposed optimal imbalance described by [63], was selected for further validation of the model. This resulted in an OA of 88.9% and an AUC of 0.973, indicating that the model was robust to the imbalanced data. It is noted that this validation dataset cannot be assumed to be independent of the training data. The imbalanced dataset’s ROC curve can be seen in Appendix D.

4.3. Explaining the Flood Hazard Assessment

The increased use of AI and ML in flood hazard assessment has shown increased accuracy whilst also reducing the transparency of such models, reducing decision-makers’ trust in outcomes [33]. This study utilised the novel SHAP explainable AI algorithm using specifically a force plot figure and a map of SHAP values for specific features for the analysis of decisions and their spatial distribution. These figures showed the overall contributions of features throughout the study area, allowing for more transparent ML modelling, debugging, and pattern recognition.

SHAP values and force plots can be utilised for pattern recognition and show the diverse flooding conditions across the study region. At point A (Figure 9 and Figure 10), low elevation had a significant contribution to increased susceptibility within the primarily coastal regions. Specifically, the far northern part of the study area is susceptible due to its low elevation. Water flows from high to low elevation, and this is in line with past research. Similarly, at point B (Figure 9), there was a contribution from elevation; however, the higher elevation reduced susceptibility, while the flat slope and high vegetation density indicated that the inland region near Proserpine is highly susceptible. This region is generally flat, and the high vegetation implies that the region may already be wet and have a high root zone SM. Conversely, in the south of the study area, there was a clear indication that flooding inland is more riverine in nature. Both points C and D (Figure 10) had DtR as a significant contributor, with elevation becoming less relevant as the river was followed upstream, as indicated at C. TWI, SPI, and NDVI all became significant contributing factors in regions where elevation became a worse predictor of flooding, indicating the significant impact they can have on flood susceptibility. As expected, higher slope values did have a strong effect on reducing flood susceptibility, with lower slope values increasing the susceptibility in the vicinity of point C. Further information about these analyses is supported by the dependence plots between features and SHAP values in Appendix E.

These SHAP map and force plot observations can be verified using reported flooding during the event. In the Proserpine River Catchment, high river levels causing flooding were not reported, indicating that the flooding was pluvial (flash) in nature, aligning with the SHAP maps in Figure 10. Conversely, in the south of the study area, major river flooding was recorded, with the Sarich stream gauge near point C observing record flooding river levels [53]. This supports the model’s observations of different flooding conditions occurring across the study area and affirms the ability of the model to discern flood-susceptible regions.

Given the impact that DtR has on a certain location’s flood susceptibility, the simplification of the DtR feature may underestimate other flood-susceptible regions. DtR was the distance from the major river in each catchment in the study area. This excludes any smaller waterways, which may have similar flooding properties. The SPI appeared to fill this gap in the data in the model; however, the SPI does not include a distance component, and locations close to high SPI values were not classed as susceptible within this model. Thus, there may be susceptible regions further inland along riverbanks and floodplains that were not included within the DtR feature. Given the SPI’s high impact in small spatial extents only along flow paths, other studies may have overlooked the impact of the feature and its strength as a flooding predictor.

The use of SHAP to explain flooding during the TC event can also indicate reasons as to reduced susceptibility. In Figure 11A, the antecedent SM and NDVI values are quite low, contributing to very low SHAP values, reducing flood susceptibility specifically in this event. This implies that there is a significant reduction in susceptibility with low vegetation density and low antecedent SM. Perhaps this indicates that during such intense, short-burst rainfall events, regions with higher antecedent SM and regions that are predisposed to have higher rainfall indicated by vegetation density struggle to cope with the high infiltration rates needed for water to percolate into the soil, and the soil saturates, creating overland flow and flooding. This finding supports the positive correlation of NDVI with flooding, as found in studies such as [62].

Alternatively, SHAP can also be used to critically assess the model’s outcomes and may reveal decisions that are inaccurate or incorrect. As shown in Figure 11B, the model’s assessment of the southern coastal region may indicate that it overtrained on the background non-flooded data in these regions of the study area and assessed these regions as not being flood-susceptible when they may be susceptible due to topographical factors. In this region, a high SM value (the highest in the study area) significantly reduced susceptibility, in conjunction with low NDVI values also reducing susceptibility.

Along the coastal region, the elevation is still relatively low, and physically, it is expected that a higher SM indicates higher flood potential [74]. During the event, the region was far south from the centre of the TC, and thus it may not have been affected; however, the assessment of this region by the model appeared to be misguided. This analysis forms a proof of concept for map validation and assessment using SHAP values and maps. The use of such analysis can help give stakeholders additional trust in model outputs and flood hazard maps, ensuring that the most accurate and sensible outcomes are used for decision-making.

4.4. Recommendations and Future Research

Whilst the outcomes of this study indicate a robust assessment of flooding in the study area, a simplification of the flooding process was the choice of a single time step that coincided with peak flooding at multiple stream gauges. Given the scope of this study, this was appropriate; however, the additional information included in the time-based flooding data would be of benefit to future modelling efforts. Areas that are flooded for longer are in more danger than those that flood for short periods. For the TC Debbie event, there are three open-source and publicly available SAR time steps, allowing for further study into this event using rare high-resolution flood mapping data. In future research, the use of mixed effects models such as mixed effects random forest (MERF) could be utilised to give a better indication of flooding over multiple time steps and bridge the gap between flood susceptibility and flood hazard.

Additionally, accumulated rainfall was unable to be included, as the variance across the study area made it redundant because rainfall was high across the whole region during the TC event. In a model that takes a temporal dimension into account, it would be expected that accumulated rainfall will vary more, allowing for the modelling of flood susceptibility during the course of the event.

Whilst the broad-brush flood assessment of the study area was successful, small flood locations may have been left out with the flood sampling approach employed in this study. Given that the misclassified points in the test dataset were commonly on the fringe or within relatively small flooded polygons, as shown in Appendix A, a more rigorous sampling technique may assist in identifying the intricacies and edges of the decision boundaries around flooded regions.

Alternatively, given the polygonised nature of the flooded dataset, alternative ML or deep learning algorithms may be more appropriate in future studies. These models may also remedy a major limitation of non-spatial ML models, in which spatial indicators are not included, such as latitude and longitude, so as not to overfit to location, a problem described in [75]. For example, a convolutional neural network, commonly used for image identification, may have success in utilising the flooded datasets to their fullest extent and have the ability to model spatial aspects of floods effectively.

In future, the use of SHAP plots can be utilised for feature contribution measures over large study areas with different flooding characteristics. Whilst the SHAP maps are quite computationally expensive and become exponentially more expensive as more features are included, it is worth noting that the use of force plots is much more computationally feasible. Selected points could be used for the analysis of contributing factors in surrounding regions. The use of SHAP calculations on sampled points may be beneficial for spatial SHAP patterns and observations. Additional plots, such as decision plots, can also be utilised to observe common high flood-susceptibility pathways, allowing for the specific differentiation of flooding events from each other.

5. Conclusions

Flooding during TC events is dangerous and destructive. This study used satellite remote sensing data and a random forest ML model to assess flood susceptibility in Australian TC-prone regions; a case study of the landfall of TC Debbie was examined. The developed model demonstrated an excellent ability to differentiate flood-susceptible regions from non-flood-susceptible regions, as indicated by the evaluation of the model’s performance using metrics such as an area under the receiver operating characteristic curve of 0.925 and an overall accuracy of 80.2%. A flood hazard map was created, and the map was explained using the SHAP method. This was a novel contribution of this work, showing that the flood susceptibility in the study area had different contributing factors, namely low-elevation coastal flooding and upstream riverine flooding. Overall, elevation, slope, and the normalized difference vegetation index were the most important flood-influencing factors. It was also demonstrated that the explainable maps can be utilised to validate the overall flood hazard map against physical understanding, adding to trust in final flood hazard maps. The explainable aspect of this model allows for the unravelling of the “black box” of complex ensemble models, such as RF. The demonstrated differences in the factors contributing to susceptibility allow for tailored decisions to be made to mitigate flooding within these regions, increasing resilience.

The method of this study is replicable with appropriate flood inventory and spatial data. The computational demands of this study allow for its replicability in data-scarce and low-resource regions, further supporting vulnerable populations in preparing for TC-related hazards.