Machine learning and SHAP-based susceptibility assessment of storm flood in rapidly urbanizing areas: a case study of Shenzhen, China

Figures & data

Figure 1. Geographical location of Shenzhen city.

Table 1. Landscape pattern index.

Scale	Index name	Index abbreviation	Description
Category Level	Number of patches	NP	Characterize the number of patches of the same landscape type.
	Landscape shape index	LSI	Reflect the complexity of the patch shape.
	Patch Cohesion Index	COHESION	Reflect aggregation degree of patches.
	Landscape Division Index	DIVISION	Reflect the fragmentation degree of patches of the same landscape type.
	Largest Patch Index	LPI	Identify the patches with dominant degree.
Landscape Level	Contagion	CONTAG	Describe the agglomeration degree or extension trend of different patch types.
	Shannon’s Diversity Index	SHDI	Reflects the overall complexity of landscape.

Table 2. Feature variables (the serial number in the names of landscape pattern features correspond to the serial number of land use type in Figure 1D).

Feature category	Quantity	Feature name
Terrain topography	5	Elevation、Slope、Surface roughness、Topographic fluctuation、Flow accumulation
Land use	5	Impervious surfaces、Green spaces、Bare land、Cultivate land、Water body
Landscape pattern	27	NP_1、NP_2、NP_3、NP_4、NP_5、LSI_1、LSI_2、LSI_3、LSI_4、LSI_5、COHESION_1、COHESION_2、COHESION_3、COHESION_4、COHESION_5、DIVISION_1、DIVISION_2、DIVISION_3、DIVISION_4、DIVISION_5、LPI_1、LPI_2、LPI_3、LPI_4、LPI_5、CONTAG、SHDI

Table 3. Experimental protocol.

Experimental protocol	Combination of features	Number of features
1	Terrain topography	5
2	Terrain topography + Land Use	10
3	Terrain topography + Landscape Pattern	32
4	Terrain topography + Land Use + Landscape Pattern	37
5	Optimal combination of features	—

Figure 2. Technical framework for urban waterlogging susceptibility assessment.

Figure 3. SHAP-based feature importance distribution.

Figure 4. Comparison of the AUC values of each classifier with different feature combinations.

Table 4. Feature combinations for each classifier to obtain the maximum AUC.

Classifier	Feature combination	AUC
ANN	COHESION_1 + Flow accumulation + NP_2 + Topographic fluctuation + CONTAG + LSI_1	0.9150
NB	COHESION_1 + Elevation + NP_2 + Topographic fluctuation + Slope	0.9042
RF	COHESION_1 + Impervious surfaces + Elevation + Flow accumulation + CONTAG + LSI_1+ Slope	0.9425
SVM	COHESION_1 + Elevation + Flow accumulation + NP_2 + Topographic fluctuation + LSI_1	0.9184
XGBoost	COHESION_1 + Impervious surfaces + Elevation + Flow accumulation + NP_2 + CONTAG + Slope	0.9592

Figure 5. The number of different feature types in the optimal feature combination (A) and the proportion of their Shapely values (B).

Figure 6. AUV values of each classifier in different scenarios (A) and their ROC (B).

Figure 7. XGBoost-based urban flooding susceptibility prediction (a), susceptibility level classification (B), and zoning statistics (C).

Data availability statement

Data will be made available on request.