Explainable Multimodal Approach for Infrastructure-Driven Socioeconomic Assessment Using Satellite Imagery and Open-Source LLMs

Abstract ―​

Accurate and interpretable assessment of socioeconomic conditions is vital for equitable policy-making, yet traditional methods face limitations in scalability and resolution. This paper presents a novel explainable multimodal framework that synergizes high-resolution satellite imagery with large language models (LLMs) to infer infrastructure- driven development levels. Unlike opaque deep learning models, our approach extracts visual features (e.g., roof materials, road density) and leverages LLMs to generate human-understandable insights, bridging the gap between geospatial data and actionable policy recommendations. We propose a hybrid architecture that aligns visual and textual modalities, enabling transparent analysis of underdeveloped regions. Experiments on diverse urban datasets demonstrate superior performance in identifying informal settlements and infrastructure gaps. By integrating explainability with remote sensing, this work advances responsible AI for social impact, offering governments and NGOs a scalable tool for targeted interventions.