Geospatial machine learning · Academic work 03

Flood Susceptibility Mapping in Fangshan District

A LeNet-style convolutional neural network combines verified vehicle insurance claims with fine-grained terrain and environmental data to map where flooding is most likely to occur.

Study area: Fangshan District, Beijing
Model: LeNet-style CNN
Authors: Jiankun Xu · Jia Li
Presented: July 2026

Flood susceptibility map of Fangshan District showing five susceptibility levels and four highlighted high-susceptibility areas, A1 through A4. — Model-produced flood susceptibility map. Four high-susceptibility clusters are highlighted as A1-A4.

Abstract

Using Fangshan District as the study area, this project converts verified vehicle insurance claims from the July 2023 “23·7” rainfall event into ground-truth flood locations. A LeNet-style CNN combines these observations with eleven topographic, hydrological, land-surface, and soil factors to produce a fine-grained susceptibility map. The model achieves an AUC of 0.876 and identifies four principal high-susceptibility clusters, offering a data-driven approach to rapid spatial risk screening.

Method

Spatial context, not isolated pixels

LeNet-style CNN architecture with two convolution and pooling stages followed by a fully connected layer and binary output. — Architecture of the proposed LeNet-style CNN, extracted from the full paper.

Research design
Ground truth	880 verified flood locations
Inputs	11 factors, including a 5 m DEM
Sample	23 × 23 × 11 neighborhood tensor
Model	Two convolution-pooling stages
Training	Adam · 20 epochs · balanced sampling

Results

A few results worth remembering

Primary model
ROC-AUC	0.876
Precision	0.863
F1-score	0.765
Recall	0.686

Spatial pattern

Risk concentrates in four clusters.

A1-A4 capture river-road corridors, mountain-front runoff convergence, and dense eastern urban areas; most northwestern mountain terrain remains lower susceptibility.

Ablation insight

A higher AUC did not mean a better map.

Removing DEM and slope raised AUC to 0.920 but reduced precision to 0.735 and over-smoothed the eastern plain, showing why spatial realism matters alongside aggregate metrics.

Robustness: under a 1:3 flood-to-non-flood ratio, recall remained 0.686 and AUC reached 0.891. Insurance claims nevertheless remain spatially clustered around roads and vehicle-dense areas.

Full paper

Flood Susceptibility Mapping in Fangshan District Based on Convolutional Neural Networks

The complete paper includes the study-area description, conditioning factors, model architecture, validation, ablation analysis, figures, and references.

Conference presentation · July 2026

Presented at the First Symposium on Geocomputation and System Simulation and the 2026 Annual Meeting of the Geocomputation and System Simulation Working Group of the Geographical Society of China.