An unsupervised framework for identifying urban pluvial flood event prototypes in Shenzhen using a convolutional autoencoder

Urban pluvial flood responses exhibit substantial temporal and spatial heterogeneity, making event-scale classifications challenging when directly relying on rainfall and water-depth observations. In this study, an unsupervised framework combining a convolutional autoencoder (AE) with K-means clustering was developed to identify prototypes of urban pluvial flood events from paired rainfall–depth time series in Shenzhen, China. The learned representations were evaluated against raw sequences and hand-crafted hydrological features using multiple clustering validity and stability metrics. Sensitivity analysis revealed that an eight-dimensional latent space provided a parsimonious representation, combining near-plateau reconstruction performance with the highest cross-seed clustering stability. At the reference station, the AE representation produced the most favourable overall clustering structure and resolved three hydrologically interpretable response patterns: persistent severe events, mild rainfall-tracking events, and high-rainfall events with weaker rainfall–depth correspondence. A cross-station comparison showed a partial rather than one-to-one recurrence of prototype semantics, further distinguishing the methodological transferability from the direct transferability of prototype labels. Analysis of rainfall–station pairings also presented that the overall clustering quality was broadly preserved, although individual event assignments remained moderately sensitive to rainfall representations. These findings demonstrate that AE-based latent feature learning can improve the event-level representation of urban rainfall–flood responses while preserving hydrologically meaningful structures.

Authors

Publication Details

Journal
Figshare
Published
2026-11-17
DOI
https://doi.org/10.6084/m9.figshare.35039831
Primary Topic
Hydrological Forecasting Using AI
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

An unsupervised framework for identifying urban pluvial flood event prototypes in Shenzhen using a convolutional autoencoder

Upmanu Lall, Zehui Zhou, Shaonan Zhu, Dianchen Sun et al.
3 links
Figshare
Hydrological Forecasting Using AI
article

An unsupervised framework for identifying urban pluvial flood event prototypes in Shenzhen using a convolutional autoencoder

Upmanu Lall, Zehui Zhou, Shaonan Zhu, Dianchen Sun, Huimin Wang, Chi Zhan, Jing Huang, Weiqian Wang
article en

Abstract

Urban pluvial flood responses exhibit substantial temporal and spatial heterogeneity, making event-scale classifications challenging when directly relying on rainfall and water-depth observations. In this study, an unsupervised framework combining a convolutional autoencoder (AE) with K-means clustering was developed to identify prototypes of urban pluvial flood events from paired rainfall–depth time series in Shenzhen, China. The learned representations were evaluated against raw sequences and hand-crafted hydrological features using multiple clustering validity and stability metrics. Sensitivity analysis revealed that an eight-dimensional latent space provided a parsimonious representation, combining near-plateau reconstruction performance with the highest cross-seed clustering stability. At the reference station, the AE representation produced the most favourable overall clustering structure and resolved three hydrologically interpretable response patterns: persistent severe events, mild rainfall-tracking events, and high-rainfall events with weaker rainfall–depth correspondence. A cross-station comparison showed a partial rather than one-to-one recurrence of prototype semantics, further distinguishing the methodological transferability from the direct transferability of prototype labels. Analysis of rainfall–station pairings also presented that the overall clustering quality was broadly preserved, although individual event assignments remained moderately sensitive to rainfall representations. These findings demonstrate that AE-based latent feature learning can improve the event-level representation of urban rainfall–flood responses while preserving hydrologically meaningful structures.

Figshare
Sustainable cities and communities, Climate action
Hydrological Forecasting Using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon