A Study on Flood Susceptibility Mapping in the Poyang Lake Basin Based on Machine Learning Model Comparison and SHapley Additive exPlanations Interpretation

Abstract

Floods are among the most destructive natural disasters, and accurate flood susceptibility mapping (FSM) is crucial for disaster prevention and mitigation amid climate change. The Poyang Lake basin, characterized by complex flood formation mechanisms and high spatial heterogeneity, poses challenges for the application of FSM models. Currently, the use of machine learning models in this field faces several bottlenecks, including unclear model applicability, limited sample quality, and insufficient machine interpretation. To address these issues, we take the 2020 Poyang Lake flood as a case study and establish a high-precision flood inundation sample database. After feature screening, the performance of three hybrid models optimized by Particle Swarm Optimization (PSO)—Random Forest (RF), Extreme Gradient Boosting (XGBoost), and Convolutional Neural Network (CNN) is compared. Furthermore, the Shapley Additive exPlanations (SHAP) framework is employed to interpret the contributions and interaction effects of the driving factors. The results demonstrate that the ensemble learning models exhibit superior performance, indicating their greater applicability for flood susceptibility mapping in complex basins such as Poyang Lake. The RF model has the best predictive performance, achieving an area under the receiver operating characteristic curve (AUC) value of 0.9536. Elevation is the most important global driving factor, while SHAP local interpretation reveals that the driving mechanism has significant spatial heterogeneity, and the susceptibility of local depressions is mainly controlled by the terrain moisture index. A nonlinear phenomenon is observed where the SHAP value was negative under extremely high late rainfall, which is preliminarily attributed to the “spatial transfer that is prone to occurrence” mechanism triggered by the backwater effect, highlighting the complex nonlinear interactions among factors. The proposed “high-precision sampling, model comparison, SHAP explanation” framework effectively improves the accuracy and interpretability of FSM. These research findings can provide a scientific basis for smart flood control and precise flood risk management in basins.

1. Introduction

Floods are one of the most frequent natural disasters worldwide, posing a constant threat to human life and property due to their sudden onset, widespread impact, and significant chain reaction of disasters [1]. According to the United Nations Office for Disaster Risk Reduction (UNDRR) 2023 Global Assessment Report on Disaster Risk Reduction [2], between 1990 and 2022, floods caused more than 2.3 million people to be affected globally, with direct economic losses reaching as high as 1.8 trillion US dollars. China, as a major flood-prone region, also experiences frequent flood disasters. In the context of global climate change, extreme precipitation events are becoming more frequent and intense, further increasing the probability of flooding. Poyang Lake, as a key water regulation hub in the middle reaches of the Yangtze River basin, is constrained by its unique topographical features of being surrounded by mountains on three sides and sloping northward into the river. Combined with the frequent occurrence of extreme precipitation caused by global climate change and the decline in the lake’s water regulation capacity due to human activities (such as land reclamation and water conservancy projects), it has become one of the regions in China most frequently affected by floods and with the most severe disaster situations [3]. Therefore, it is crucial to compile flood susceptibility mapping, identify flood susceptibility in each region, and take targeted measures to respond to it. This will not only effectively prevent floods but also mitigate the damage they cause.

Flood Susceptibility Mapping (FSM) is a key foundational task for understanding and supporting regional flood susceptibility management. It aims to identify the spatial distribution of flood susceptibility areas and provide core support for decision making in flood prevention and disaster mitigation. Early FSM research relied heavily on traditional statistical models and physical process models. Statistical models sought to uncover the relationship between factors that cause disasters and flood events. For example, the Analytic Hierarchy Process (AHP) quantifies weights by constructing a comparison matrix [4]. The Frequency Ratio model (FR) identifies key factors by calculating the correlation between factor status and flood probability [5]. There is also an integrated weighting method (WoE) for evidence [6], as well as hydrological and hydrodynamic models based on physical processes (HEC-RAS) [7]. However, these methods generally suffer from inherent limitations, such as strong subjectivity, difficulty in effectively capturing complex nonlinear relationships, high requirements for sample quantity and quality, and limited accuracy [8,9]. With the development of machine learning (ML) algorithms, their powerful nonlinear modeling and high-dimensional data processing capabilities have been widely applied in FSM research, significantly improving model prediction performance [10], algorithms such as Random Forest (RF), Support Vector Machine (SVM), Convolutional Neural Network (CNN), and Extreme Gradient Boosting Tree (XGBoost) are widely applied in FSM research. By integrating multi-source data, including topography, hydrology, geology, meteorology, and environment, and exploring the complex correlations between these factors and flood occurrence, these algorithms have become the mainstream methods for generating high-precision flood susceptibility mapping [11,12,13].

Nevertheless, current ML-based FSM research still faces several important challenges: first, there is a lack of systematic evaluation of model applicability and performance. Different ML algorithms exhibit significant performance differences under varying geographical conditions and data constraints, particularly in complex river and lake systems like Poyang Lake, where factors such as flood backwater effects from the Yangtze River, heterogeneous surface conditions, and intense human activities interact. The robustness and predictive accuracy of which model is most suitable for such complex basins remain poorly understood [14]. Secondly, sample quality limits model performance. Building a high-quality flood inundation sample set is a prerequisite for training robust ML. General data sources, such as radar images, limit the accuracy of flood sample extraction [15]. Current sample extraction primarily relies on optical and synthetic aperture radar (SAR) remote sensing imagery. While optical remote sensing can directly reflect surface features, it is often constrained by weather conditions. In situations such as dense cloud cover, this can lead to missing data on flood events. For example, during the extreme rainstorms at Poyang Lake in 2020, optical imagery could not be obtained due to excessive cloud cover [16]. In comparison, SAR imagery, with its ability to actively transmits electromagnetic waves, can penetrate cloud cover to enable continuous, around-the-clock monitoring and is also highly sensitive to water bodies, making it particularly suitable for flood-prone areas prone to floods that experience heavy rainfall (such as the Poyang Lake basin) [17,18]. However, SAR data is susceptible to interference from factors such as wind speed and terrain distortion, leading to reduced accuracy in flood boundary extraction. Commonly used automated flood extraction methods, such as traditional extraction methods based on thresholds (Otsu algorithm), often confuse flood areas as moist soil, stable water bodies, shadows, or still water surfaces in complex scenarios. This results in a significant amount of mislabeled data being introduced into the dataset, introducing notable noise and thereby reducing the reliability of the model [19,20]. Thirdly, insufficient model interpretability hinders the understanding of disaster mechanisms. Although ML models have high predictive accuracy, their decision-making processes for decision making often lack transparency. Traditional global feature importance rankings are unable to reveal the spatial heterogeneity of driving factors and the complex nonlinear interactions [21]. However, understanding this heterogeneity and nonlinearity is crucial for analyzing their impact on unique hydrological processes, such as river backwater, and for formulating precise flood control strategies based on zoning and classification.

To address the aforementioned research gaps, this study focuses on the Poyang Lake basin as the study area and develops a comprehensive FSM research framework that integrates sample construction with high precision, multi-model performance comparison, and SHAP interpretation. The main innovations of this study include: (1) This study integrates Sentinel-1 SAR time-series analysis across multiple temporal dimensions and high-resolution optical imagery (Google Earth) for manual interpretation, compares imagery features before, during, and after the event, and supplements the analysis with hydrological and meteorological data to accurately identify and exclude mountain shadows, building scattering, and stable water bodies—thereby constructing a flood inundation sample database with high precision and high reliability and fundamentally resolving the mislabeling issues commonly associated with automatic sample extraction based on SAR. (2) For the first time, a systematic comparison of the performance of three representative machine learning (ML) algorithms (RF, XGBoost, CNN) in FSM was conducted in the Poyang Lake basin, and particle swarm optimization (PSO) was introduced for hyperparameter tuning. This aimed to identify the optimal flood susceptibility prediction model for the region, thereby enhancing the model’s applicability and reliability. (3) To overcome the linear limitations of traditional importance ranking, this study innovatively integrates the SHAP (SHapley Additive exPlanations) interpretable framework, which not only quantifies the global contributions of each factor but also provides a deeper explanation of the local effects of key flood factors and the interaction effects between two variables, significantly enhancing our understanding of the mechanisms of flood susceptibility in Poyang Lake.

The overarching goal of this study is to improve the accuracy and interpretability of FSM in the complex Poyang Lake basin. To this end, we pursue three specific objectives:

(1)

To develop a flood inventory with high fidelity by integrating remote sensing data from multiple sources with manual interpretation to overcome the limitations of automated SAR extraction.

(2)

To determine the most robust predictive model for this region by systematically comparing the performance of RF, XGBoost, and CNN algorithms optimized by PSO.

(3)

To decipher the decision-making mechanisms behind flood susceptibility using the SHAP framework, elucidating both global importance and local nonlinear interactions among driving factors, this objective addresses the interpretability gap.

The remainder of this paper is structured as follows:

(1)

Section 2 describes the study area and the methodologies employed.

(2)

Section 3 presents the results of the model comparisons and the SHAP interpretation.

(3)

Section 4 discusses the implications of the findings.

(4)

Section 5 concludes with the main findings and future research directions.

2. Materials and Methods

2.1. Study Area

The Poyang Lake Basin is located in northern Jiangxi Province, south of the Yangtze River, east of Mount Lu. Its main area roughly coincides with the administrative boundaries of Jiangxi Province, located between 28°22′–29°45′ north latitude and 115°47′–116°45′ east longitude. The terrain is primarily composed of plains and hills, with elevations ranging from −187 m to 1445 m within the study area (Figure 1). Poyang Lake serves as a major tributary lake in the middle reaches of the Yangtze River and is also China’s largest freshwater lake [22]. The river is 173 km long from north to south, with a maximum width of 74 km from east to west and an average width of 16.9 km. Its drainage basin covers an area of approximately 162,200 square kilometers. The basin has a water system that is well developed, receiving water from five major tributaries—the Gan River, Fu River, Xin River, Rao River, and Xiu River—as well as smaller rivers such as the Boyang River and Xi River. It connects to the Yangtze River via Hukou [23], the complex hydrological system is characterized by interconnected rivers and alternating periods of high and low water levels. The terrain consists of three parts, including mountainous areas along the edges, hilly areas in the south and central region, and the Poyang Lake Plain. The entire basin has a wide range of altitudes, and its unique topography allows precipitation to quickly flow into the lake area.

The basin has a subtropical humid monsoon climate, with warm and humid weather and an average annual precipitation of 1662 mm. Precipitation is unevenly distributed in terms of time and space. From April to June, the region is affected by the season before the monsoon in South China, and during the rainy season from March to September, heavy rainfall accounts for about 80% of the region’s annual precipitation [24]. From July to September, due to rainfall and snowmelt, coupled with the terrain of the lake area that is low lying, flood disasters frequently occur in the basin. Among these, the severe flood disasters of 1954, 1998, and 2020 caused significant social and economic losses [25].

Given the unique geographical scope and climatic conditions of the Poyang Lake basin, its flood formation mechanism is distinctive, characterized by a composite flood resulting from the interaction between the “five rivers’ inflow” and “Yangtze River backwater effect”. During the main flood season (April to June), floods are primarily formed by concentrated heavy rainfall in the basins of the Gan, Fu, Xin, Rao, and Xiu rivers. At this time, the lake water level rises, and lake water flows into the river. However, during the late flood season from July to September, frequent flood peaks occur along the Yangtze River mainstem. When the Yangtze River water level exceeds that of Poyang Lake, severe “backwater effect” or even “river water backflow” phenomena occur, significantly obstructing Poyang Lake’s outflow, resulting in prolonged high water levels, expanded inundation areas, and exacerbated disaster conditions [26]. This unique hydrological situation means that the region’s susceptibility to flooding depends not only on local rainfall, but also on complex external hydrological factors related to the Yangtze River, resulting in significant non-linearity and spatial heterogeneity [23].

In 2020, Poyang Lake experienced an unprecedented flood. From late June to early July, the basin experienced continuous heavy rainfall. On 8–9 July, it was hit by a historic torrential downpour, with an average rainfall of 108 mm across the province, breaking the historical record since 1961. This flood event submerged approximately 1961.95 square kilometers of the Poyang Lake region. The most severely affected areas of Poyang Lake impacted over 600,000 people, with approximately 335 square kilometers of crops damaged, resulting in direct economic losses of 31.33 billion yuan [27,28]. This has also caused ecological damage, such as the submersion of wetland vegetation and the fragmentation of aquatic habitats, resulting in a significant decline in biodiversity. Given that this event not only caused significant impacts but also vividly demonstrated the composite flood characteristics resulting from the combined effects of the “five rivers’ inflow” and the “Yangtze River backwater effect”, it effectively reflecting the nonlinearity and spatial heterogeneity of flood susceptibility in this region. Therefore, this study selected the 2020 flood event as the basis for sample collection for the Flood Simulation Model (FSM).

2.2. Data Source

The data in this study encompass multiple influencing factors with multiple dimensions. Topographic data consist of the ASTER GDEM Digital Elevation Model (DEM) with a spatial resolution of 30 m. Provided by the Geospatial Data Cloud Platform (Computer Network Information Center, Chinese Academy of Sciences, Beijing, China), which was jointly developed by NASA (Washington, DC, USA) and Japan’s Ministry of Economy, Trade, and Industry (METI, Tokyo, Japan). This dataset covers terrestrial areas globally between 83° north and south latitude globally. For hydrological factors, river data originated from OpenStreetMap (OSM) vector river data, while water level data for Hukou Station—officially named “Hukou Hydrological Station of the Yangtze River Water Resources Commission,” with a precise geographic coordinate of 29°45′ N, 116°13′ E, specifically located in Hukou County, Jiujiang City, Jiangxi Province, China, at the confluence of Poyang Lake and the Yangtze River (i.e., the estuary where Poyang Lake flows into the Yangtze River)—from 23 June to 27 August 2020, were sourced from the Yangtze River Water Resources Commission’s Water and Rainfall Conditions Report column. Precipitation factor data derived from the NOAA (National Oceanic and Atmospheric Administration) National Environmental Information Center (Asheville, NC, USA)’s 1929–2024 global site daily precipitation data. Vegetation coverage data utilized the 2020 Normalized Difference Vegetation Index (NDVI) from the National Ecological Data Center Resource Sharing Service Platform (Beijing, China). This study employed the open source dataset from Zenodo (CERN, Geneva, Switzerland) for land use type data. Lithological data were based on regional geological survey data processed via the open source toolkit from GitCode (Shenzhen, Guangdong, China). Road data is sourced from OSM 2020 vector road network data. Sentinel-1A C-band SAR imagery (VV/VH polarization, 10-m spatial resolution) from 26 June and 14 July 2020, used for flood point extraction, is provided by the Alaska Satellite Facility (University of Alaska Fairbanks, Fairbanks, AK, USA), see

Table 1. Data Source Table.

Data	Data Source	Type	Spatial Resolution
Elevation	Geospatial Data Cloud	Raster	30 m
River	https://www.openstreetmap.org/ (OpenStreetMap Foundation, London, UK) (accessed on 15 November 2024)	Vector
Precipitation Factors	https://www.ncei.noaa.gov/data/global-summary-of-the-day/archive (NOAA National Centers for Environmental Information, Asheville, NC, USA) (accessed on 19 January 2025)	Raster	1000 m
NDVI	http://www.nesdc.org.cn (accessed on 17 February 2025)	Raster	30 m
Land Use	https://zenodo.org/records/12779975 (accessed on 19 February 2025)	Raster	30 m
Lithology	https://gitcode.com/open-source-toolkit/53e1c/?utm_source=tools_gitcode&index=bottom&type=card&&isLogin=1 (accessed on 8 April 2025)	Vector
Road	https://www.openstreetmap.org/ (OpenStreetMap Foundation, London, UK) (accessed on 15 November 2024)	Vector
Sentinel-1	Alaska Satellite Facility (ASF)	Raster	5 m × 20 m
Water Level at Hukou Station	http://www.cjh.com.cn/swyb_syqbg.html (accessed on 29 March 2025)

.

2.3. Flood Inventory

Building a high-precision flood sample database is the foundation for the successful training and validation of machine learning models [29]. This study leverages the 2020 catastrophic flooding event in the Poyang Lake basin. By manually interpreting multiple temporal Sentinel-1 SAR imagery (before and after the disaster) and supplementing it with cross-validation using concurrent high-resolution optical imagery (e.g., Google Earth), we precisely delineated the flood inundation areas. The operational workflow proceeded as follows, starting with the use of 26 June (early flood stage) imagery for preliminary flood extent delineation. Subsequently, the 14 July (peak flood stage) imagery was overlaid to identify expanded inundation areas, with particular emphasis on removing misclassified pixels caused by mountain shadows, building reflections, and stable water bodies. To precisely determine stable water bodies, this study introduced two types of water system data from the 2020 basin for comparative analysis, including OSM water system polygon data and water system data from land use data. The comparison revealed that the water area in the land use data was smaller and more consistent with the actual characteristics of the Poyang Lake basin (where seasonal hydrological changes cause significant interannual fluctuations in water surface area). Therefore, it was more suitable as the basis for defining stable water bodies. Accordingly, water bodies from the land use data were converted into vector water body data and defined as “stable water bodies”. Finally, the polygons representing the finalized flood inundation areas were converted into centroid points, serving as flood sample points (totaling 3720). Non-flood sample points were extracted through random sampling from non-inundated areas using post-disaster imagery, ensuring their quantity was comparable to that of flood sample points.

For terrain data, ArcGIS 10.7 software (Esri, Redlands, CA, USA) was used to extract slope, aspect, terrain undulation, calculated using a 5 × 5 pixel window, Terrain Wetness Index (TWI—the uphill catchment area is calculated based on the D8 algorithm, with slope expressed in radians), as well as planar curvature, profile curvature, and composite curvature from ASTER GDEM. For hydrological factor processing, the DEM underwent fill processing first. Subsequently, hydrological analysis was conducted using the D8 single flow direction algorithm on the filled DEM to obtain cumulative flow volume and flow direction data. Concurrently, Euclidean distance analysis was applied to OSM vector river data to derive a “distance to river” raster dataset. Precipitation factor data undergoes inverse distance weighting (IDW) interpolation (with variable search radius and power index set to 2). Combined with administrative boundary masking and bilinear interpolation resampling, this generates 30-m resolution grids for segmented indicators, including early period (19 June–25 June5), late period (8 July–14 July), peak (13 July), and full period (19 June–14 July) rainfall. Lithological data were reclassified based on permeability and erosion resistance into five categories: “Hard Impermeable Erosion-Resistant Rock Class,” “Dense Impermeable Erosion-Prone Rock Class,” “Soluble Rock Class,” “Other Special Rock Types Class,” and “Loose Permeable Erosion-Prone Rock Class.” Road data were derived from Euclidean distance analysis of OSM 2020 vector road network data, yielding a “distance from roads” raster dataset. For flood point extraction, the European Space Agency (ESA) SNAP 11.0.0 software (ESA, Paris, France) was used to perform orbit correction, radiometric calibration, pulse band noise removal, polarization matrix C₂ generation, enhanced Lee filtering (3 × 3 window to suppress speckle noise), multiple view processing (4 views in azimuth + 1 view in range), and terrain correction (based on 30-m resolution SRTM DEM). The final image was projected into the WGS1984-UTM Zone 50N coordinate system. All data were ultimately unified to a 30 m resolution and the WGS1984 UTM Zone 50N coordinate system.

Finally, all factors and flood sample points were integrated into a geospatial database. All sample points (including flood and non-flood) were randomly divided into a training set (2976 samples) and a test set (744 samples) at a ratio of 8:2, respectively, used for model training and performance validation.

2.4. Flood Influencing Factors

The occurrence of floods is a process influenced by multiple factors (Figure 2). Numerous scholars have conducted extensive research on the factors related to floods, providing important evidence for the accurate assessment of areas prone to floods. Based on previous research and the characteristics of the study area, this paper selected 20 factors from four major categories—topography and geology, hydrology and meteorology, land cover, and human activities. Among the topography and geology factors, topography plays a crucial role in determining flood inundation areas [30]. Elevation directly determines the topography of an area. Areas with terrain that is low lying and minimal elevation changes are more susceptible to flooding [31]. Slope has a significant impact on water flow velocity and runoff time. The gentler the slope, the smaller the surface undulation, the slower the water flow velocity, the longer the runoff time, and the more likely it is to cause flooding during heavy rains [32]. Slope orientation affects sunlight and precipitation, which in turn affects soil moisture and runoff conditions [33]. Curvature reflects the characteristics of changes in terrain undulations and directly affects the convergence and dispersion trends of water flow [34]. Rock type affects the permeability and erosion capacity of rocks [35]. The Topographic Wetness Index (TWI) measures the effect of topography on water distribution. Higher values of the index indicate that the region is more water retentive, increasing the susceptibility of the region to surface runoff and flooding. It is important to note that this index differs from the previously mentioned “low lying terrain with minimal elevation changes being prone to flooding” (i.e., the direct impact of topography on flooding). It does not solely focus on the elevation or undulation of the terrain itself. Instead, it comprehensively evaluates the “upstream catchment area” (how much upstream water the region can collect) and “slope” (the ease or difficulty of water retention or runoff on the surface) to determine the terrain’s potential for water accumulation. Simply put, TWI reflects terrain’s influence on flooding through its “water convergence and retention capacity,” rather than relying directly on topographic features [36]. Hydrological and meteorological conditions are the direct driving forces behind the formation of floods. The cumulative flow volume reflects the intensity of water convergence within a watershed, the greater the cumulative volume, the higher the likelihood of flooding [37]. The flow direction determines the direction of water flow and influences the path of flood propagation [38]. The distance from a river is an important indicator for assessing the flood risk level of an area. Generally, the closer the distance to the river, the higher the flood risk. However, in the case of the Poyang Lake flood, the influence of this factor may be relatively low [39]. Rainfall was the key factor in triggering this flood. Early cumulative rainfall (19–25 June) increased soil saturation, while later rainfall (8–14 July) enhanced the flood’s persistence. Peak Rainfall (13 July) triggered a sudden rise in water levels, and total cumulative rainfall (19 June–13 July) determined the scale of the flood [40,41]. Land cover influences flood events by affecting rainfall interception, infiltration, and runoff processes. The Normalized Difference Vegetation Index (NDVI) is a measure of vegetation cover, with higher values indicating better vegetation cover. Plant surfaces can intercept precipitation, thereby decreasing surface runoff and lowering the likelihood of flooding [42]. Different land uses, such as farmland, forest land, and construction land, have different effects on rainwater infiltration and runoff [43]. Different soil types have varying degrees of porosity, permeability, and water retention capacity, which means they have different effects on runoff generation and flood formation [44]. Human activities influence flood susceptibility by altering landforms and hydrological systems. Road construction changes landforms and drainage systems, and areas close to roads may become more susceptible to flooding due to poor drainage [45].

2.5. Methods

The technical workflow of this study is shown in Figure 3 and primarily consists of the following six steps:

(1)

Geographic Spatial Database Construction: First, based on Sentinel-1 SAR and optical imagery with multiple temporal dimensions, flood inundation areas were manually interpreted and converted into 3720 flood point data points. An equal number of non-flood points were randomly generated in non-inundated areas to form the model training dataset. Next, using data from multiple sources (DEM, precipitation, geology, remote sensing, etc.), 20 preliminary environmental factors were calculated and extracted, including terrain, hydrology, meteorology, and other geological and human activity indicators. Finally, all data underwent unified geospatial registration and resampling processing.

(2)

Feature Selection and Optimization: To eliminate multicollinearity and screen out core driving factors, this paper uses VIF to perform a quantitative analysis of the initial factors, with VIF > 10 as the criterion for determining the existence of severe collinearity [46], through an iterative process, factors with the highest VIF values are eliminated until all remaining factors have VIF values ≤ 5. To further avoid overfitting in subsequent models, the Recursive Feature Elimination (RFE) method is used to select key factors based on the purified factor set [47]. By iteratively evaluating the contribution of features to the explanatory power of the model, features with low contribution to the model are gradually removed, and the optimal feature subset is ultimately retained for subsequent modeling.

(3)

Machine Learning Model Construction and Hyperparameter Optimization: Three representative algorithms—RF, XGBoost, and CNN—were selected to construct flood susceptibility prediction models. The PSO algorithm was introduced to automatically optimize the core hyperparameters of each model, with the optimization objective being the area under the curve (AUC) value on the validation set.

(4)

Model Performance Evaluation and Validation: We quantitatively evaluated and compared the three optimized models on the test set using the AUC and the confusion matrix. Additionally, we used the Jenks Natural Breaks method to classify the flood susceptibility probabilities output by the models into five risk levels: extremely high, high, medium, low, and extremely low, and created a flood susceptibility distribution map.

(5)

Explanatory Analysis of Driving Mechanisms: This study employs the SHAP explanatory model to interpret the contribution of each factor, quantifying their average marginal contributions from a global perspective and ranking them accordingly. Additionally, it analyzes the prediction decision-making process for individual samples from a local perspective, thereby revealing the complex relationships among key factors that are not linear and have interactivity.

Figure 3. Flow chart.

Figure 3. Flow chart.

2.5.1. Random Forest (RF)

Random Forest is an ensemble learning algorithm proposed by Breiman in 2001 [48], whose core principle is to enhance the model’s generalization ability by constructing multiple uncorrelated decision trees. The algorithm first generates multiple independent subsamples from the original training set using bootstrap sampling, with each subsample used to train a Classification and Regression Tree (CART) tree. Additionally, during node splitting in each tree, a random selection of features is used in the splitting rule calculations to reduce inter-tree correlation. The construction of a single decision tree is guided by minimizing the Gini index (for classification tasks) or the mean squared error (for regression tasks) [49], with the tree structure generated through recursive splitting without pruning. During final prediction, classification problems use the voting results of all trees, while regression problems take the average. Compared to a single decision tree, random forests effectively reduce overfitting risk through ensemble learning and demonstrate stronger adaptability to high-dimensional data and noise.

2.5.2. Extreme Gradient Boosting Tree (XGBoost)

XGBoost is a boosting ensemble learning method proposed by Chen et al. [50] in 2016. Its core idea is to gradually optimize model error by serially constructing CART trees. The algorithm uses CART trees as its base learners. During tree construction, the optimal split point is determined by calculating the reduction in the loss function (gain) before and after the split. The decision tree model is iteratively constructed to fit the residuals of the previous iteration’s model, thereby reducing the prediction model’s error, and its predicted value is calculated by:

$$Y_i={\displaystyle \sum _{t=1}^T{f}_t}(x_i).$$

(1)

In Equation (1), $Y_i$ is the predicted value for sample $i$; $f_t(x_i)$ denotes the prediction result of decision tree $T$ for sample $i$; $T$ is the total number of decision trees in the ensemble; and $x_i$ is the feature vector of sample $i$.

The XGBoost objective function controls the complexity of the tree through the loss function and regularization terms (such as the number of leaf nodes and node weight penalties). During the iteration process, each new tree aims to minimize the cumulative loss of all previous trees. The leaf node weights are quickly solved using the Taylor expansion of the second order of the loss function, enabling the model to optimize along the gradient descent direction. The expression of the objective function is:

$$L(\phi )={\displaystyle \sum _{i=1}^{n}l}(y_i,Y_i^t)+{\displaystyle \sum _{k=1}^t\mathrm{\Omega }}(f_k).$$

(2)

In Equation (2), $L(\phi )$ represents the objective function of XGBoost; $n$ denotes the total number of samples in the dataset; $y_i$ is the true label of the $i$-th sample (a categorical label for classification tasks or a continuous value for regression tasks); $l(y_i,Y_i^t)$ is the loss function, measuring the error between the true value $y_i$ of the $i$-th sample and the predicted value $Y_i^t$ at iteration $t$; $t$ denotes the current iteration count (i.e., t decision trees have been constructed); $\mathrm{\Omega }(f_k)$ is the regularization term, measuring the complexity of the $k$-th decision tree (including penalties for the number of leaf nodes and node weights to prevent overfitting); $f_k$ denotes the $k$-th decision tree.

2.5.3. Convolutional Neural Network (CNN)

Convolutional neural networks (CNNs) are a type of deep learning model suitable for processing data with a grid structure [51]. Their core advantage stems from the synergistic effect of local receptive fields and weight sharing mechanisms, which can efficiently extract spatial sequence structural features from the data. A hierarchical modular design enables feature abstraction from low to high levels.

Its core components include: (1) Convolution layer: This layer is the core of feature extraction, performing local scans on input data using multiple sets of learnable convolution kernels (or filters). Each convolution kernel moves across the input in a sliding window of fixed size, generating feature responses by calculating the weighted sum of the input within the window and the convolution kernel, edge information can be preserved through padding mechanisms. The convolution results are processed through an activation function (such as ReLU) to introduce nonlinear transformations, enabling the network to learn complex features. Lower-level convolution layers typically capture basic visual features (such as edges and textures), while higher-level layers gradually learn more abstract semantic features (such as object components) as the network deepens [52]. (2) Pooling layers follow convolution layers to perform nonlinear downsampling, reducing computational load by compressing feature map sizes, aggregating local information to enhance the model’s robustness to input translation, rotation, and scale changes, and suppressing overfitting risks. Common methods include max pooling (taking local maxima) and average pooling, with max pooling being more effective at retaining salient features [53]. (3) After multiple rounds of convolution, activation, and pooling, the high-level feature maps are flattened into vectors of one-dimension vectors and fed into the fully connected layer. This layer integrates all high-level features through weighted connections and ultimately outputs the prediction results through the output layer. Dropout regularization is often introduced to suppress overfitting, enabling learning from end to end from raw data to decision-making [54]. The value of CNN lies in its automated feature learning capability, which enables it to mine deep spatial patterns from raw data without the need for manual feature engineering.

2.5.4. Particle Swarm Optimization (PSO)

The particle swarm optimization algorithm was proposed by Kennedy & Eberhart (1995) [55]. Used to optimize ML model hyperparameters. Efficiently searches the solution space by simulating the collaborative foraging behavior of flocks of birds. Each particle represents a potential solution to the optimization problem and updates its own state by tracking the historical optimal position $p_{best,i}$ of the individual and the global optimal position $g_{best}$ of the group. The state of the particle is described by the position vector $x_i$ and the velocity vector $v_i$. The velocity update formula is [56]:

$$v_i^{k+1}=w\cdot v_i^k+c_1\cdot r_1(p_{best,i}^k-x_i^k)+c_2\cdot r_2(g_{best}^k-x_i^k).$$

(3)

Among these, $v_i^{k+1}$ denotes the velocity vector of particle $i$ at iteration $k+1$; represents the inertia weight, used to balance the particle’s historical velocity with the current search trend; $v_i^k$ denotes the velocity vector of particle $i$ at iteration $k$; $c_1$ and $c_2$ are learning factors (typically set around 2.0), controlling the strength of learning toward the particle’s individual best position and the global best position, respectively; $r_1$ and $r_2$ are random numbers following a uniform distribution $U$ (0, 1), introducing randomness into the search process; $p_{best,i}^k$ denotes the individual historical best position vector of particle $i$ at iteration $k$; $x_i^k$ denotes the position vector of particle $i$ at iteration $k$; $g_{best}^k$ represents the global optimal position vector of the entire particle swarm at iteration $k$; the superscript $k$ indicates the current iteration number, while $k+1$ notes the next iteration number. Based on velocity updates, particle position updates are achieved via Equation (4):

$$x_i^{k+1}=x_i^k+v_i^{k+1}.$$

(4)

Here, $x_i^{k+1}$ denotes the position vector of particle $i$ at iteration $k+1$; $x_i^k$ denotes the position vector of particle $i$ at iteration $k$; and $v_i^{k+1}$ denotes the velocity vector of particle $i$ at iteration $k+1$ after updating via Equation (3).

2.5.5. SHAP Algorithm

This study uses the SHAP (SHapley Additive exPlanations) method to explain the model prediction mechanism. This method is based on the SHapley value theory in game theory [57]. By assigning contribution metrics to the features of each sample, we can quantitatively decompose the prediction results of nonlinear models. For the predicted value $f(x)$ of any sample, it can be decomposed into the sum of the baseline value and feature contributions via SHAP values, and the formula is:

$$f(x)=f(\varnothing )+{\sum }_{i=1}^n{\varphi }_i(x).$$

(5)

Among them, $f(\varnothing )$ is the model baseline value (the global mean of all training sample predictions), ${\varphi }_i(x)$ is the SHAP value of feature $i$ on sample $x$, quantifying the marginal contribution of that feature to the current prediction, with its absolute value representing the strength of the influence. This expression satisfies the Shapley value’s local unbiasedness and global consistency axioms, ensuring the fairness of feature contribution decomposition. The specific implementation involves constructing a feature importance ranking by calculating the mean absolute value of the SHAP values across the entire sample, and using a dependency graph to illustrate the nonlinear influence patterns of key features on prediction results across different value intervals (such as positive/negative effects, saturation effects, etc.), thereby providing traceable explanatory evidence for the model’s prediction logic.

3. Results

3.1. Collinearity and Actor Screening

To ensure the stability of the model, the VIF was first used to assess multicollinearity among the feature factors (Figure 4). The results showed that the factors derived from the terrain (composite curvature: VIF = 123.07, profile curvature: VIF = 50.75, planar curvature: VIF = 36.63, terrain undulation: VIF = 16.10, slope: VIF = 14.98) all exhibit severe multicollinearity (VIF > 10), which may distort the assessment of model parameters and importance ranking [58]. Given the high collinearity between composite curvature and terrain undulation, this paper prioritizes the removal of both factors. After this operation, the VIF values of all retained factors are reduced to below 5, indicating that the collinearity issue has been effectively resolved.

3.2. Feature Selection

To optimize the feature space, improve model efficiency, and enhance interpretability, this study employs RFE combined with a Random Forest classifier for feature selection, as shown in Figure 5. By progressively reducing the number of features and evaluating model performance metrics such as the AUC value and accuracy rate, it was found that when the number of features was reduced to 9, the model performance reached its peak (AUC = 0.9599, accuracy rate 0.8958). Further reduction in the number of features resulted in a slight decline in model performance. Therefore, 9 features were determined to constitute the optimal input set, specifically including hydrological factors (peak rainfall, late rainfall, total rainfall, early rainfall, distance to river), topographical factors (elevation, TWI), vegetation cover (NDVI), and human activity index (distance from roads).

To further ensure the reliability of the selected features and avoid the potential impact of multicollinearity on model stability, this study calculated the Pearson correlation coefficients between all features and plotted a correlation heat map (Figure 6). Multicollinearity testing is a critical prerequisite for building robust machine learning models, as highly correlated features can distort the true estimation of feature importance in the model and may lead to overfitting and reduced generalization ability [58]. The analysis results show that the absolute values of the correlation coefficients between all features are below 0.7. This result indicates that there is no severe multicollinearity issue within the selected feature subset, laying a solid foundation for the subsequent construction of a stable and reliable predictive model. Although different temporal rainfall factors exhibit relatively high correlations (e.g., the correlation coefficient between Peak Rainfall and Total Rainfall is 0.66, between Peak Rainfall and Early Rainfall is 0.63, and between Total Rainfall and Late Rainfall is 0.60), their values are still far below the threshold for severe multicollinearity (0.8 or 0.9). Therefore, all these factors are retained.

3.3. Comparison of Multiple Models

This study employed three models—RF, XGBoost, and CNN. To ensure a fair comparison and maximize each model’s performance, we first applied the Particle Swarm Optimization (PSO) algorithm for hyperparameter tuning. The PSO process successfully identified distinct optimal configurations for each model (see Supplementary Material Table S1 for full details). Finally, the models with optimized parameters were tested, and multi-dimensional performance evaluation was conducted using ROC curves and confusion matrices, with the results shown in Figure 7.

ROC curve evaluation indicates that all three models demonstrate strong classification performance, but RF has the highest AUC value (0.9536), followed by XGBoost (0.9528) and CNN (0.9355). Further analysis using the confusion matrix and accuracy metrics shows that the RF model correctly classified 692 flood category samples and incorrectly classified 52. 655 correctly classified non-flood category samples and 89 incorrectly classified samples, with an overall accuracy rate of 0.9052. The XGBoost model correctly classified 692 flood category samples and incorrectly classified 52. The non-flood category correctly classified 654 samples and incorrectly classified 90 samples, achieving an accuracy rate of 0.9046, the best performance among the three models. The CNN model correctly classified 651 flood category samples and incorrectly classified 93 samples. The non-flood category correctly classified 614 samples and incorrectly classified 130 samples, with an accuracy rate of 0.8501, slightly lower than the previous two models. Considering the AUC value and accuracy rate, the RF model outperforms the XGBoost model in terms of AUC value and accuracy rate, with differences of 0.0008 and 0.0006, respectively. It demonstrates stronger discrimination capability between flood and non-flood samples and the lowest classification error rate. The XGBoost model performs moderately, while the CNN model is effective, it lags behind the previous two models in terms of performance.

3.4. Flood Susceptibility Zoning Results

To validate the generalization capabilities of different ML models in the Poyang Lake basin after training, this study imported the prediction results of the trained RF, XGBoost, and CNN models into ArcGIS for mapping visualization. Using the natural breakpoint method, the results were divided into five flood susceptibility levels: extremely low susceptibility, low susceptibility, moderate susceptibility, high susceptibility, and extremely high susceptibility. The study selected the severely affected areas of Poyang County, the Donghu District of the main urban area, and the Xingcheng Development Zone of Gongqing City as typical validation zones for the 2022 flood. Flood susceptibility maps were created (Figure 8), and the proportion of each susceptibility level across the entire study area was statistically analyzed and presented in numerical form (

Table 2. Table showing the proportion of susceptibility levels for different model grades.

Model	Class	Pixel Count	Area (km2)	Proportion (%)
RF	Very low	15,331,415	13,798.27	62.36
	Low	3,126,034	2813.43	12.71
	Middle	2,168,700	1951.83	8.82
	High	1,929,767	1736.79	7.85
	Very high	2,030,544	1827.49	8.26
XGBoost	Very low	17,423,811	15,681.43	70.87
	Low	1,958,296	1762.47	7.96
	Middle	1,332,112	1198.9	5.42
	High	1,342,960	1208.66	5.46
	Very high	2,529,281	2276.35	10.29
CNN	Very low	11,266,195	10,139.58	45.83
	Low	3,553,441	3198.1	14.45
	Middle	3,225,684	2903.12	13.12
	High	2,889,495	2600.55	11.75
	Very high	3,651,645	3286.48	14.85

).

Figure 8 visually displays the prediction results of the three models. The RF and XGBoost models exhibit high consistency in spatial distribution patterns (Figure 8a,b), with their prediction results showing significant spatial heterogeneity, clear boundaries between hazard levels, and rich details. In contrast, the CNN model’s prediction results (Figure 8c) exhibit a pronounced spatial smoothing effect, with hazard levels distributed more densely, generally larger patch areas, and weaker reflection of details of micro topography. All three models consistently identified four core aggregation zones with a high risk. These zones include the southern and western lakeshore areas of Duchang County, the downstream floodplain of the Rao River basin, the northern part of Xinjian District and the eastern lowlands of Yongxiu County, and the central and western parts of Hukou County and the western part of Pengze County. However, as shown in the comparison between Figure 8a–c, the CNN model tends to predict higher flood risk levels in areas such as the lakeshore zone of Poyang Lake, Nanchang County, Wenyang County, and the southern part of Xinjian District, upgrading many high-risk areas to extremely high-risk and downgrading low-risk areas to middle-risk. However, in the central part of Yugan County, the CNN model assesses the flood risk level as lower than RF and XGBoost, classifying this high-risk area as a middle-risk zone.

By analyzing the risk levels of each model using the statistical data in Table 2, the results of the RF and XGBoost models are similar. In the RF model, the area with extremely low risk accounts for the largest proportion (62.36%), followed by low risk (12.71%), middle-risk (8.82%), extremely high risk (8.26%), and high risk (7.85%). The XGBoost model also shows the highest proportion of extremely low-risk areas (70.87%), followed by extremely high-risk (10.29%), low-risk (7.96%), high-risk (5.46%), and middle-risk (5.42%). The CNN model shows a different pattern in the proportion of high-risk levels, with significant increases in the area proportions of extremely high, high, and middle-risk (14.85%, 11.75%, and 13.12%, respectively). The proportion of low-risk areas is similar to that of the other two models (14.45%), but the proportion of extremely low-risk areas has dropped significantly to 45.84%.

From the selected typical urban areas, Gongqing City, due to its proximity to the Poyang Lake basin, is predicted by all three models to be primarily high-risk and extremely high-risk, with concentrations in the central, southeastern, and northeastern parts of the city. Donghu District is predominantly low-risk in all three models, with only a few scattered middle-risk and high-risk patches in the northern part. The central and southern parts of Poyang County, as the hardest-hit areas, are accurately identified as high-risk regions by the RF and XGBoost models. However, the CNN model generally overestimates the risk level in this area, with risk grades consistently higher than actual conditions (Figure 8c).

3.5. Explanation of the Decision-Making Mechanism for Flood Susceptibility Mapping

3.5.1. Global Factor Importance: Prioritization of Key Disaster-Causing Factors

Using the SHAP interpretability model, a quantitative analysis of the primary drivers of flood occurrence in the Poyang Lake basin can be conducted at both the global and local levels. The analysis results indicate (Figure 9) that Elevation is the primary driving factor for flood susceptibility (0.400), significantly higher than other variables. Among rainfall timing factors, Late Rainfall (0.109) and Peak Rainfall intensity (0.103) have higher contribution levels, followed by Early Rainfall (0.091) and Total Rainfall (0.070). Thus, Elevation and Rainfall meteorological factors are the primary driving conditions. In contrast, TWI, Distance to Road. Distance to River and NDVI have lower explanatory power, all of which are relatively weak (all < 0.07).

As shown in Figure 9b, the horizontal axis indicates the direction and magnitude of each feature’s influence, with a gradient from blue to red reflecting an increase in feature values. Elevation shows a significant negative correlation with flood occurrence probability, meaning that SHAP values increase as elevation decreases, consistent with the physical mechanism that areas that are low lying terrain is prone to water accumulation [59]. Notably, Later Rainfall exhibits an inverse pattern: lower Later Rainfall corresponds to higher SHAP values, while Peak Rainfall is positively correlated with SHAP values. The SHAP values of other factors are densely clustered around zero, indicating their limited role in driving flood susceptibility differentiation.

3.5.2. Global Single-Factor Dependency Analysis

To further quantify the driving contributions of different rainfall sequences to flood susceptibility in the Poyang Lake basin, this study generated dependency plots of a single factor for key rainfall factors using SHAP analysis (Figure 10). Analysis of Peak Rainfall intensity (Figure 10a) shows that this factor also exhibits a significant positive correlation with flood susceptibility. When Peak Rainfall intensity is below 90 mm, SHAP values are primarily negative. In the 90–130 mm range, the samples are highly concentrated, and SHAP values are widely distributed between −2 and 2, reflecting the high sensitivity of flood susceptibility predictions to changes in Peak Rainfall intensity and the variability of model responses in this range. When Peak Rainfall intensity exceeds 130 mm, SHAP values are generally greater than 0.

Analysis of Later Rainfall intensity (Figure 10b) shows that the factor dependency exhibits more complex nonlinear characteristics. Within the approximately 150–160 mm range, SHAP values are mostly positive, while when rainfall is below approximately 145 mm, SHAP values tend to be negative. When Later Rainfall is abnormally high (>200 mm), SHAP values instead show a negative trend, suggesting that this high-value range may be influenced by other geographical or hydrological factors.

Early Rainfall analysis shows (Figure 10c) that the SHAP value exhibits a relationship that is not monotonic and fluctuating with rainfall. When rainfall is below 100 mm, the SHAP value is generally negative. In the 100–120 mm range, SHAP values show an upward trend and approach 0 around 120 mm. Subsequently, SHAP values rise again between 120 and 140 mm, reaching a local positive peak around 130 mm, and then decline back to negative values around 140 mm. In the 140–160 mm range, SHAP values return to positive values. It is worth noting that within the broad range of 100 mm to 160 mm, SHAP values are densely distributed between −1 and 2, with relatively small fluctuations, indicating significant variability in the marginal contribution of Early Rainfall to model predictions within this range. When rainfall exceeds 160 mm, SHAP values are primarily negative.

Analysis of Total Rainfall (Figure 10d) shows a significant positive response trend between rainfall and flood occurrence probability. When rainfall is below 550 mm, SHAP values are generally negative. When rainfall exceeds 650 mm, SHAP values stabilize as positive. In the transitional zone between 550 mm and 650 mm, SHAP values are densely distributed between −1 and 1, indicating that the model predictions within this range exhibit high uncertainty in response to Total Rainfall.

3.5.3. Single Cause Dependency Diagram

Global feature importance analysis reveals the overall influence patterns of factors, but it is insufficient to explain the specific mechanisms that drive floods in local regions. To investigate this spatial heterogeneity and compare the differences between local and global explanations, this study selected typical samples for local SHAP analysis (Figure 11). As shown in the figure, the model’s predicted value for this sample is f(x) = 1.656, indicating an extremely high probability of flooding. The contributions (SHAP values) and directions of various driving factors to the predicted value vary: TWI is the primary promoting factor (SHAP = +1.05). Total Rainfall is the primary inhibiting factor (SHAP = −0.58). Distance to River (SHAP = +0.53) and Distance to Road (SHAP = +0.40) are promoting factors. Peak Rainfall (SHAP = −0.36) and Early Rainfall (SHAP = −0.13) are inhibiting factors. Elevation (SHAP = +0.21) and NDVI (SHAP = +0.17) are minor promoting factors for occurrence.

4. Discussion

4.1. Comparison of Model Performance and Applicability

This paper systematically evaluated the performance of three models—RF, XGBoost, and CNN in flood susceptibility mapping of the Poyang Lake basin. The results demonstrated that the overall classification performance of the ensemble learning models (RF, AUC = 0.9536, XGBoost, AUC = 0.9528) was significantly superior to that of the CNN model (AUC = 0.9355) (Figure 7). This finding contrasts with conclusions drawn by other researchers [60], such as those based on mountainous landslide analysis [61]. This discrepancy arises from the intrinsic relationship among model principles, data structure, and geographical complexity.

The significant advantages of the RF and XGBoost models in this study arise from their intrinsic algorithms, which align well with the requirements of flood susceptibility modeling. Flood systems are influenced by the complex interactions of multiple factors, such as topography, meteorology, and hydrological conditions, exhibiting strong nonlinear characteristics. Models based on trees naturally capture the interactive effects between variables through recursive splitting, eliminating the need for complex feature engineering. Additionally, XGBoost reduces model complexity through regularization constraints [50], while RF decreases variance by randomizing feature subsets [62], making both models highly robust to scale differences and feature collinearity (e.g., Late Rainfall and Total Rainfall) [63,64]. These findings are consistent with the results of this study, as both models demonstrate high heterogeneity and rich detail in spatial predictions (Figure 8a,b), accurately capturing the impact of topographic features on a small scale (e.g., depressions, embankments) in Poyang Lake on flood risk, thereby exhibiting superior generalization capabilities.

In contrast, CNN models perform slightly worse and require further analysis in conjunction with model features and the study context. While CNNs excel at handling spatial features, their generalization ability may be inferior to that of models based on trees when dealing with tabular or flood data not based on images, as they are susceptible to limitations in the scope of training data [65,66]. Khattab et al. note that models based on trees (such as XGBoost and RF) outperform CNNs on complex datasets [66], which aligns with the findings of this study. In the specific context of this research, three factors may explain this outcome: first, due to sample size limitations, CNN models typically require large datasets to train deep networks capable of learning generalized features, and the 6440 samples in this study may be insufficient to fully capture the complex patterns driven by floods, consistent with the conclusions of Vemula et al. [67]. Second, the feature processing method may have influenced the results. This study employed recursive RFE to select the optimal feature subset, which significantly improved the performance of models based on trees but may have disrupted the original local autocorrelation structure of the geospatial data, thereby limiting the spatial feature extraction capability of CNN convolutional kernels [60]. Third, the complexity of the study area poses a challenge. As a large connected lake basin, Poyang Lake’s patterns driven by floods mechanisms are influenced by multiple factors such as lake–river backwater effects and rainfall. Additionally, human activities over the long term have altered the land surface, resulting in blurred and irregular local spatial patterns [25,68], which further reduce the efficiency of CNN convolution operations in extracting spatial features.

To further validate the predictive capabilities of the three machine learning models, this study statistically analyzed the distribution of flood locations across the hazard levels generated by each model (

Table 3. Flood Point Statistics Table.

Model	Very Low (Count, %)	Low (Count, %)	Medium (Count, %)	High (Count, %)	Very High (Count, %)	Total Count
RF	0 (0.00%)	4 (0.11%)	24 (0.65%)	163 (4.38%)	3529 (94.87%)	3720
XGB	1 (0.03%)	10 (0.27%)	47 (1.26%)	223 (5.99%)	3439 (92.45%)	3720
CNN	1 (0.03%)	32 (0.86%)	177 (4.76%)	513 (13.79%)	2997 (80.56%)	3720

). As shown in the table, RF correctly classified 94.87% of historical flood samples into the extremely high-risk zone, outperforming XGBoost (92.45%) and CNN (80.56%). Additionally, RF exhibited a misclassification rate of only 0.11% in the extremely low and low susceptibility categories, significantly lower than XGBoost (0.30%) and CNN (0.89%). These statistical results align with the ROC test outcomes, effectively confirming that the RF model demonstrates the highest discriminative accuracy and stability.

In summary, ensemble models based on tree structures, such as RF and XGBoost, demonstrate superior adaptability, robustness, and predictive accuracy when addressing flood susceptibility in complex geographical systems like Poyang Lake. These findings provide empirical evidence to guide model selection in this region.

4.2. Heterogeneity of Flood Driven Mechanisms

Global and local explanations based on SHAP jointly reveal that the patterns of mechanisms driven by floods of Poyang Lake exhibit significant characteristics dependent on scale. In the global analysis, elevation, late rainfall, and peak rainfall are the dominant factors (Figure 9), indicating that regional topography and key meteorological elements collectively govern the overall pattern prone to floods, consistent with the findings of Luo et al. [69]. However, local sample analysis shows that under specific scenarios, the terrain wetness index (TWI = 6.14) replaces elevation as the primary risk driver (SHAP = +1.05), while total and peak rainfall act as risk suppressors (SHAP values of −0.58 and −0.36, respectively) (Figure 11).

This discrepancy highlights the specificity of local hydrological processes. High TWI values indicate that the sample is located in a topographic depression or near a confluence area [70]. Although the total and peak rainfall amounts are high, they fall within a range where the SHAP contribution is negative or highly variable (Figure 10a,b). This suggests that while such rainfall has the potential to cause disasters, its ultimate effects are modulated or even reversed by stronger local processes. We infer that this modulation is related to the unique mechanism of transfer of backwater, flooding, and risk of Poyang Lake. Spatial localization verification (Figure 12) confirms that the sample point is indeed situated in an agricultural area that is low lying (Elevation 20–50 m) at the edge of a flood retention zone. This may be explained by the following formation mechanism: when the entire region experiences heavy rainfall (total rainfall > 660 mm), the combined effects of water inflow from the Five Rivers and Yangtze River backwater cause the main river channel water level to rise or activate the flood retention area. This process may reduce flooding risk at this point by altering the spatial priority of flooding or triggering drainage scheduling, resulting in a negative contribution from the rainfall factor. The positive contribution of distance to river (SHAP = +0.53) further supports the view that this location is susceptible to regulation by the main river system [71].

To quantitatively validate the interaction between topography and rainfall, this study further analyzed the global dependency relationships of TWI (Figure 13). The results indicate that TWI exhibits a significant nonlinear relationship with flood risk: when TWI is less than 7.5, SHAP values are predominantly positive, consistent with the traditional understanding that convergence increases risk. However, when TWI > 7.5, SHAP values generally become negative, contradicting findings from natural watershed studies [72]. This discrepancy reflects the unique characteristics of the Poyang Lake embankment area. In some regions with extremely high TWI values, flat depressions have been diked and controlled by pump stations, and the artificial water network has disrupted and restructured the natural convergence pathways. Consequently, flood dynamics in these regions are no longer primarily governed by upstream water inflow but are more influenced by engineering operations and external water level backwater control, leading to an inversion of their statistical patterns [71]. The flood sample in this study has a TWI of 6.14 (SHAP > 0) (Figure 11), placing it in a transitional zone still significantly influenced by natural runoff processes, although its risk patterns have been altered by other regulatory factors.

Combining global and local explanatory analyses underscores the limitations of relying solely on approaches at a single scale. The global model explanation captures the general rule that “increased rainfall typically increases risk”, but its contribution ranking is easily influenced by the distribution of training data and feature collinearity, which can smooth out or obscure key local processes. In contrast, local explanations precisely capture specific responses within particular geographical contexts (such as depressions) and system states (such as backwater saturation and scheduling impacts). This suggests that the complexity of the Poyang Lake flood system arises from the interplay among static geographical features, dynamic hydrological processes, and human activities. Therefore, future flood risk assessments should integrate multiple-scale perspectives and fully consider regional characteristics, such as river–lake interactions, to achieve a more comprehensive understanding of flood formation mechanisms.

4.3. Interaction Effects of Temporal Rainfall Factors

SHAP single-factor dependency analysis revealed an anomalous phenomenon: when late rainfall exceeded 160 mm, its contribution to flood susceptibility became negative, with the SHAP value remaining negative beyond 200 mm (Figure 10c). This suggests that under extreme late rainfall conditions, this factor exerts a weak inhibitory effect on flood probability, contradicting conventional hydrological understanding. To further explore the mechanism behind this phenomenon, this study conducted a two-factor interaction analysis involving Late Rainfall with total rainfall and early rainfall (Figure 14).

The interaction analysis reveals that when late rainfall exceeds 200 mm and the SHAP value is negative, most samples correspond to lower early rainfall (<110 mm) and higher total rainfall (>660 mm). As shown in Figure 10a, the SHAP value fluctuates between positive and negative contributions within the 100–160 mm range for early rainfall, indicating that its influence exhibits significant threshold effects and scenario dependence. Lower early rainfall may result from insufficient soil saturation during the preceding period, while extremely high total rainfall (>660 mm) suggests that the basin is in a state of overall saturation. This phenomenon can be explained by the unique lake–river relationship and backwater effect of Poyang Lake: the late rainfall period defined in this study (8–14 July) completely overlaps with the critical period when Poyang Lake experienced intense backwater effects and even backflow from the Yangtze River in 2020 [73]. Under the influence of elevated water levels in the Yangtze River mainstem, Poyang Lake’s flood discharge was severely obstructed, and the risk of inundation in the lake basin reached saturation. At this point, although the later rainfall generated additional runoff, its marginal contribution to the already extremely high overall risk significantly decreased. Furthermore, this extreme scenario may trigger spatial restructuring of the inundation area and operational responses, thereby relatively reducing inundation risk in certain local regions. It is important to emphasize that the negative contribution indicated by the SHAP value does not imply that flooding will not occur. Rather, it signifies that the incremental contribution of this factor to the model’s predicted probability is negative under specific system conditions (such as backwater saturation). If rainfall during the later period is lower, the model’s predicted flood probability may increase.

The results above indicate that in basins such as Poyang Lake, which are governed by complex river–lake interactions, patterns driven by flood mechanisms exhibit significant scenario dependency, and the impact of late rainfall is strongly influenced by external hydrological pressures. Therefore, future flood warning practices should not consider rainfall factors in isolation but should instead establish a dynamic, multiple-factor collaborative assessment framework. This paper recommends incorporating real-time indicators that reflect the intensity of backwater effects—such as the water level difference between the Hukou and Xingzi stations and backflow flow rates—into the risk assessment system, while dynamically adjusting the weights assigned to different rainfall factors based on hydrological conditions. For example, once backwater or backflow events are confirmed, the weighting of later rainfall indicators should be appropriately reduced, with greater emphasis placed on dynamic and static factors such as terrain depressions, antecedent soil moisture, and engineering operations that more accurately indicate local vulnerability. This study enhances our understanding of the mechanisms underlying flood formation in lakeside basins and provides a foundation for improving the explanatory power and warning accuracy of risk models under complex hydrological conditions.

4.4. Uncertainty Analysis and Limitations

The robustness of any predictive modeling study is inherently constrained by its uncertainties and limitations, which must be carefully examined to properly contextualize the findings. Guided by this principle, we present a structured analysis of potential sources of error and their propagation through our analytical framework, distinguishing between inherent data uncertainties and methodological constraints.

The foundation of our model depends on input data that inherently contain irreducible uncertainties. First, despite the rigorous manual interpretation protocol used to construct the flood inventory, errors primarily arise from difficulties in distinguishing floodwaters from stable water bodies and wet soils in SAR imagery. These labeling errors systematically propagate through the model’s learning process, potentially biasing it toward misclassifying the spectral or spatial signatures of ambiguous pixels and consequently reducing overall confidence in prediction probabilities. Second, the vertical error in the ASTER GDEM (~±15 m), propagates nonlinearly into key topographic derivatives [74]. This is especially critical in low-slope floodplain areas, where it significantly affects the TWI. Consequently, the model’s reliance on elevation and TWI introduces higher inherent uncertainty into susceptibility predictions for pixels in very gentle terrain. This issue is further exacerbated by the 30 m resolution resampling, which smooths micro-topographic features essential for accurately identifying flood accumulation.

Beyond data uncertainties, algorithmic and methodological choices introduce additional constraints. To quantify the model’s instability, the Random Forest model was run 100 times. The results showed negligible performance variation (AUC = 0.9536 ± 0.0018), confirming that the model’s predictions are highly stable and not sensitive to random seed initialization given our data. However, model selection bias is evident from the superior performance of models based on trees (RF, XGBoost) compared to the CNN. This suggests a potential architectural bias favoring our tabular data structure, indicating that CNNs may systematically underestimate risk in areas where natural spatial patterns are disrupted by anthropogenic features such as embankments—a limitation clearly reflected in the CNN’s oversmoothed susceptibility map (Figure 8c).

Finally, our methodological framework inherently defines the scope of our conclusions. Although the selection of the nine final factors was rigorously optimized using VIF and RFE, this step remains somewhat subjective; alternative factor sets (e.g., including soil type or antecedent moisture) could alter local SHAP importance patterns, even though the global dominance of topography and rainfall remains consistent. Additionally, the use of the Natural Breaks method for zonation means that the areal percentages of each risk level are not directly comparable across models, directing interpretation toward spatial patterns rather than quantitative extents. Most importantly, the exclusion of dynamic, real time hydrological drivers (e.g., Yangtze River water levels) and detailed anthropogenic controls (e.g., pump operations) simplifies the complex dynamics of Poyang Lake. This limitation likely results in an underestimation of risk in areas where backwater effects—insufficiently captured by static topographic indices—are the primary controlling factors.

5. Conclusions

This study uses the 2020 Poyang Lake flood as a case study to develop a high-precision, interpretable flood susceptibility mapping framework. By integrating multiple temporal Sentinel-1 SAR and optical imagery through manual interpretation, a highly reliable flood sample database was created. The performance of three machine learning models—RF, XGBoost, and CNN—was systematically compared. The SHAP framework was employed to thoroughly analyze the contributions of driving factors and their interaction mechanisms. The main conclusions are as follow:

(1)

RF and XGBoost outperformed the CNN model in the application of FSM in this study area. The RF model achieved the highest prediction accuracy (AUC = 0.9536), effectively capturing the spatial heterogeneity of risk driven primarily by microtopography and demonstrating the greatest applicability to this region. This finding provides a critical empirical benchmark for model selection in complex, human-modified river–lake systems, suggesting that models based on tree ensembles are more adept at handling their disrupted spatial patterns than deep learning models designed for coherent image data.

(2)

Globally, topographic factors, particularly elevation, are the primary drivers of flood susceptibility, followed by meteorological factors (late rainfall and peak rainfall). However, SHAP local explanations reveal significant spatial heterogeneity, indicating that the importance ranking and direction of influence of these driving factors dynamically change depending on the geographical context and system state (e.g., backwater conditions). Locally, the TWI can replace elevation as the dominant factor, while high-intensity rainfall factors may even exhibit an inhibitory effect due to the “spatial-risk transfer” mechanism. This underscores a fundamental limitation of relying solely on global explanations and highlights the necessity of multiple-scale, context-specific analysis for accurate local risk assessment and management.

(3)

The flood susceptibility of Poyang Lake arises from complex nonlinear interactions between natural geographical factors and human activities. SHAP two-factor analysis revealed a significant interaction between “topography and rainfall,” confirming that unique hydrological and geomorphological processes, such as lake–river backwater effects, modulate the driving mechanisms. This finding underscores the limitations of traditional statistical methods in analyzing such complex systems. More importantly, it reveals that flood drivers are not static but can invert under extreme system states (e.g., backwater saturation), a phenomenon that traditional models fail to capture.

In summary, this study confirms the superiority and robustness of models based on trees, particularly RF, in applying the FSM approach to complex human–landscape coupled basins such as Poyang Lake. The proposed “high-precision sampling–multiple model comparison–SHAP machine interpretation” framework not only significantly improves mapping accuracy but also greatly enhances the understanding of flood formation mechanisms. This provides a solid scientific foundation and practical methodological tools for flood risk management in the region. Beyond Poyang Lake, our work demonstrates the power of explainable AI to move beyond black-box prediction and uncover the complex, non-linear, and often counterintuitive mechanisms that govern environmental risks in the Anthropocene, ultimately paving the way for more adaptive and intelligent risk management strategies.