Modelling flood susceptibility based on deep learning coupling with ensemble learning models.

Area of Science:

• Environmental Science
• Geographic Information Science
• Artificial Intelligence in Environmental Modeling

Background:

• Flood susceptibility modeling is crucial for mitigating disaster losses.
• Data-driven methods, including ensemble and deep learning, are state-of-the-art.
• The combined effect of deep learning and ensemble learning in flood modeling remains unexplored.

Purpose of the Study:

• To propose and evaluate three novel deep learning (DL) coupled with ensemble learning models for flood susceptibility.
• To investigate the performance of DL combined with Filtered Classifier (FC), Rotation Forest (RF), and Random Subspace (RSS).
• To assess the effectiveness of these hybrid models in producing accurate flood susceptibility maps.

Main Methods:

• A case study in Dingnan County, China, was used to generate flood and non-flood data.
• Frequency ratio analysis identified ten significant flood-influencing factors.
• Three hybrid models (FC-DL, RF-DL, RSS-DL) were developed and compared against a standalone DL model.

Main Results:

• All developed models demonstrated good performance (AUC > 0.8) on validation data.
• The FC-DL model achieved the highest AUC (0.996) for training data, outperforming RF-DL, RSS-DL, and DL.
• Hybrid models showed more reliable and excellent performance compared to the standalone deep learning model.

Conclusions:

• Deep learning coupled with ensemble learning models offers superior performance for flood susceptibility modeling.
• The proposed hybrid approach provides a reliable tool for land-use planning and flood risk management.
• This methodology is adaptable and applicable to flood susceptibility mapping in diverse geographical regions worldwide.