Histoire  |  Géographie  |  Économie  |  Société

 

Mudit Mehta, 2008 | Lausanne, VD

 

Rivers have been the cradle of human civilization but are prone to devastating floods due to meteorological phenomena and unchecked human activities. Governments spend millions on sophisticated flood prediction models requiring expensive scientific infrastructure, limiting their deployment to a few rivers worldwide. But riverbank width can serve as an excellent alternative indicator for flood risk. However, finding a reliable and cost-effective solution to monitor riverbank widths along the entire length of rivers remains challenging.

This project aims to leverage technology by using computer vision AI to analyze satellite images of rivers, monitor their bank widths, and identify flood risk zones by comparing them with ideal bank widths. This early risk indication helps governments take timely preventive actions, such as clearing bank zones, increasing water carrying capacity, or creating flood detention areas.

Introduction

Fluvial and pluvial floods cause significant damage to communities, infrastructure, and the environment, resulting in an annual economic loss of approximately $8 billion globally. Governments invest heavily in sophisticated flood prediction models that rely on expensive sensors, gauges, and monitoring towers. However, due to high costs, only a small percentage of rivers are equipped with such infrastructure. Riverbank width can serve as an excellent alternative indicator for flood risks. Various studies provide guidelines for ideal bank widths based on the river’s width, but continuous monitoring of bank widths along extensive river lengths is impractical and costly.

This project aims to develop and test a reliable, cost-effective, and scalable solution using a computer vision AI model. The model will analyze satellite images of rivers, measure their widths, assess their bank widths, and identify flood risk zones.

Methods

The development of the AI model involved three steps:
1. Source selection: Various layers of ArcGIS and Google Maps were explored for well-defined bank lines, clear contrast, and minimal noise. Finally the ‹Default› layer of Google Maps was chosen.
2. Image processing: Computer vision techniques were used to find river edges and bank lines by converting images to grayscale and applying Canny Edge detection, including Gaussian Blur filter, Sobel filter, and Edge Hysteresis.
3. Risk identification: The AI model created ideal bank lines based on river width and compared them with existing bank lines to classify river sections as Risk zones or Safe zones.

Results

Computer vision AI model was created in Python and Aare river section (Google Maps @47.2137874, 7.5550917, 1034m), was tested on the model. Model successfully
o Measured the river’s width in meters (87.02 m)
o Created ideal bank lines using guidelines from Federal Office for Water and Geology’s study
o Compared existing bank lines with ideal minimum bank lines
o Classified the river section based on their risk profile
Model was accurate with tolerance of +/- 0.5 m or +/- 0.6% of actual length.

Discussion

The AI model successfully met project objectives but had the following limitations:
• Image quality dependency: Lower resolution images or cloud presence increased noise and worsened edge detection.
• Sharp river bends: Sections with multiple channels or steep curves created multiple contours, requiring concatenation for seamless edges. Improved filters or better coding logic could enhance contour creation.
Further development: As next step, build an AI application around this model that
o Links all river images together to create a continuous view of the entire river.
o Estimates time to disaster based on rate of change of riverbank width.
o Incorporates self-learning function to improve predictive capabilities using historical data.
o Develops prescriptive functionality to suggest suitable interventions.

Conclusions

The AI model developed to monitor riverbank widths and highlight flood risk zones was tested on the Aare River, a Swiss tributary of the Rhine River, at coordinates (47.212903, 7.552900). It successfully accomplished the project objectives of measuring river width, analyzing existing bank lines, creating ideal bank lines, comparing both, and highlighting risk zones with a tolerance of +/- 0.6% accuracy. The model is scalable and can reliably and efficiently provide early flood risk warnings, aiding governments in timely preventive interventions.

 

 

Appréciation de l’expert

Luc Rochat

Le candidat, sensible aux risques que représentent les inondations, notamment pour les populations des pays les moins avancés, propose une solution innovante pour une gestion du risque basée sur l’imagerie satellitaire. Il a méticuleusement développé un code qui, à partir d’une image, permet de définir dans quels secteurs d’un cours d’eau les berges devraient être modifiées pour prévenir au mieux les risques d’inondations. Son modèle a été testé favorablement sur un tronçon de l’Aar. Il cherche maintenant à améliorer son approche pour pouvoir l’étendre à n’importe quel cours d’eau.

Mention:

très bien

 

 

 

International School of Lausanne, Le Mont-sur Lausanne
Enseignante: Smiley Noelle