Quantum image segmentation for flood mapping
Insurance
Brainstorming and exploration
At an initial brainstorming meeting with teams from an insurance company, we presented the various quantum technologies and their specific advantages. This productive exchange kick-started a discussion on the potential use cases that quantum computing could facilitate. Together, we identified several areas where this technology could offer real added value, forming a selection of opportunities to be further explored during a month-long in-house research workshop. During this workshop, QbitSoft explored quantum methods, while the company's internal Data teams focused on deploying conventional solutions available on the market.
Use case selection (why is flood mapping a priority?)
Flooding events are among the most frequent and destructive natural disasters, ranking just behind storms in terms of insurance costs. Accurate flood detection is essential for insurers, enabling them to efficiently estimate the number of claims and allocate resources cost-effectively and quickly, ultimately leading to greater customer satisfaction.
The goal: to create business value
Traditional flood detection methods, which rely heavily on supervised learning models, are costly due to the need for manually labeled data and the high cost of computational resources.
Conventional unsupervised detection methods struggle to deliver good performance, often being sensitive to noise and light variations. The aim is to push back the limits of the state-of-the-art in unsupervised segmentation, while minimizing costs and ecological impact. The aim is to develop more robust and efficient approaches, capable of overcoming these challenges without requiring excessive computing resources.
Exploration: Proof of Concept(POC)
Our second quantum module, called Q-SCAN, is a quantum image segmentation module designed to efficiently detect flooded areas from satellite images. We validated our solution using the Sen1Floods11 dataset, which covers 18 flooding events worldwide, and benchmarked against unsupervised methods as well as state-of-the-art supervised machine learning models. Q-SCAN performed exceptionally well, outperforming traditional unsupervised approaches and rivaling state-of-the-art supervised methods. What's more, as an unsupervised approach, Q-SCAN bypasses the problems associated with the acquisition of labeled data, requires no prior training phase and therefore considerably reduces the cost and ecological impact of segmentation processes.
Impact: towards large-scale transformation
Satisfied with these promising results, we have finalized the validation and optimization of our pipeline. Q-SCAN is now a ready-to-deploy solution, bringing a real revolution in claims detection methods. At the same time, our team is working on adapting Q-SCAN for other use cases, notably forest fire detection.