GeoAI with limited data: Making AI-driven asset detection accessible for small organisations

Traditional geospatial AI models for road asset detection require massive, costly labelled datasets placing advanced automation out of reach for many towns and startups.

This session introduces a novel few-shot learning approach that detects road features from georeferenced panoramic imagery using limited training data, eliminating the need for large-scale manual labelling. Besides introducing the topic, the presentation demonstrates robust performance across multiple countries, camera systems, and diverse road environments. The method delivers accuracy comparable to large-data models while adapting quickly to new scenarios making scalable, high-performance geospatial AI accessible to organisations of any size.