Vision-Language Models (VLMs) in GeoAI Systems: Enhancing Brownfield Change Detection through Semantic Reasoning

Abstract

This paper presents a hybrid GeoAI methodology for semantic change detection in satellite imagery by integrating Vision–Language Models (VLMs) into an unsupervised clustering-based pipeline. Building on earlier work using K–Means clustering to validate pre-selected brownfield regions from SPOT (2021–2023) and aerial imagery, we address the persistent issue of semantic ambiguity, particularly in high-uncertainty zones. We introduce a zero-shot, reasoning-based verification layer that evaluates whether visual differences across time are structurally meaningful. This approach improves interpretability, traceability, and diagnostic robustness. Evaluation across 1,000 human-labeled samples and a temporally unseen test set of size 200 (SPOT23–SPOT24) demonstrates notable improvement in ambiguous zones in reasoning quality and error transparency, especially where prior methods faltered. Our framework maintains the speed and scalability of clustering while injecting semantic precision through natural language decision paths. Designed with the UN’s SDG 11 (Sustainable Cities and Communities) in mind particularly for brownfield redevelopment this work contributes to scalable, interpretable, and operationally viable GeoAI systems.