Flood Susceptibility ML

Leakage-aware classical machine learning for flood susceptibility mapping under a strict spatial holdout.

Scope

This repository packages a conference-oriented flood susceptibility mapping study using only classical machine learning in the main workflow:

• Decision Tree
• Random Forest
• Extra Trees
• AdaBoost
• Gradient Boosting
• HistGradientBoosting
• XGBoost
• LightGBM

Neural-network experiments are excluded from the main study and treated as archived side experiments only.

Scientific Focus

The repository is organized around five methodological principles:

1. Leakage-aware spatial evaluation
2. Reproducible preprocessing for mixed geospatial predictors
3. Classical ML comparison under the same spatial split
4. Explainability for selected models
5. Publication-ready outputs and documentation

Dataset

Raw source:

• [data/raw/Flood_data.csv](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/data/raw/Flood_data.csv)

Key dataset facts:

• Original rows: 15,000
• Rows with missing target removed: 12
• Final labeled rows: 14,988
• Positive rate: 27.42%
• No duplicate rows
• No duplicate coordinate pairs

Preprocessing

The final classical pipeline uses:

• removal of rows with missing target
• coordinate reservation for spatial splitting and export only
• one-hot encoding of LULC
• cyclic transformation of Aspect into Aspect_sin and Aspect_cos
• missing-indicator variables
• median imputation of continuous predictors
• reproducible feature assembly without coordinate leakage

Spatial Evaluation

Main split design:

• strategy: spatial
• block size: 5000
• spatial folds: 10
• validation fold: 0
• test fold: 1

This strict spatial holdout is used to reduce leakage from nearby samples and better reflect real flood susceptibility mapping use cases.

Current Main Findings

• Random Forest achieved the highest test ROC-AUC: 0.6270
• XGBoost achieved the highest test average precision: 0.4024
• XGBoost achieved the highest test F1: 0.4516
• Tree/boosting models are the strongest family under spatial evaluation

Recommended primary model:

• XGBoost

Recommended benchmark baseline:

• Random Forest

Repository Layout

ml_flood_susceptibility_mapping/
|- README.md
|- requirements.txt
|- environment.yml
|- .gitignore
|- LICENSE
|- CITATION.cff
|- data/
|- notebooks/
|- src/
|- outputs/
|- docs/
|- scripts/
`- archive/

Quick Start

Create the environment and run the full classical package:

python scripts/run_all.py

This produces:

• processed splits in data/processed/
• metrics and tables in outputs/metrics/ and outputs/tables/
• publication figures in outputs/figures/
• explainability outputs in outputs/shap/
• susceptibility maps in outputs/maps/
• paper draft files in docs/

Key Output Files

• [table_test_summary.csv](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/outputs/tables/table_test_summary.csv)
• [model_comparison.png](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/outputs/figures/model_comparison.png)
• [conference_paper.md](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/docs/conference_paper.md)
• [conference_paper.docx](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/docs/conference_paper.docx)

Archived Material

Neural-network experiments are not part of the main paper package. See:

• [README.md](/h:/Flood Susceptibility/ml_flood_susceptibility_mapping/archive/excluded_experiments/README.md)