Integration of hard and soft supervised machine learning for flood susceptibility mapping.

Area of Science:

• Environmental Science
• Geospatial Analysis
• Machine Learning

Background:

• Flooding is a major global hazard causing significant casualties and economic losses annually.
• Efficient flood susceptibility mapping (FSM) is crucial for mitigating flood risks and informing management strategies.

Purpose of the Study:

• To evaluate the predictive performance of artificial neural network (ANN) algorithms for flood susceptibility mapping (FSM).
• To compare hard and soft supervised machine learning classifications using different ANN models and activation functions for FSM.

Main Methods:

• Employed three ANN algorithms: multi-layer perceptron (MLP), fuzzy adaptive resonance theory (FART), and self-organizing map (SOM), with various activation functions (sigmoidal, linear, commitment, typicality).
• Integrated these models to predict flood spatial expansion in the Ajichay river basin using 10 flood-influencing factors.
• Validated models using flood inventory data, sensitivity analysis (OFAT, AFAT), and total operating characteristic (TOC) with area under the curve (AUC).

Main Results:

• The MLP with a sigmoidal activation function (MLP-S) achieved the highest success rate (92.1%) and projection rate (90.1%).
• All evaluated flood-influencing factors positively impacted FSM, with altitude being the most influential and curvature the least.
• MLP-S and MLP-L demonstrated strong flood prediction capabilities, highlighting the effectiveness of integrating machine learning with specific activation functions.

Conclusions:

• Artificial neural networks, particularly MLP-S, are highly effective for accurate flood susceptibility mapping.
• The findings support the use of advanced machine learning techniques for improved flood risk management and spatial planning.