Large Language Models for Interpreting the Earth: A New Paradigm from Perception to Decision Intelligence

Large-scale Earth observation now produces petabytes of multimodal data, yet transforming these vast streams into causal insights and decision-ready narratives remains challenging. Traditional models excel at pixel-level perception but cannot explain why environmental changes occur or what actions they imply, creating a growing insight gap. Large Language Models (LLMs) help close this gap by enabling semantic reasoning, cross-modal alignment and natural language interaction, which supports concrete applications such as image-text retrieval, open-domain visual question answering, automated geospatial reporting, spatiotemporal interpretation, and interactive decision support. Major obstacles persist, including domain knowledge misalignment, the computational burden of large-scale satellite data, multimodal heterogeneity across optical, synthetic aperture radar, LiDAR, and hyperspectral sensors, limited explainability, and emerging ethical concerns. This review synthesizes current advances and highlights specific methodological pathways such as scalable multimodal foundation models, domain adaptation and knowledge grounding techniques, physics-informed modeling, and trustworthy explainable artificial intelligence frameworks that can mitigate these challenges. These developments outline a pathway for LLMs to evolve from semantic interpreters into reliable, causal-reasoning engines for next-generation geospatial intelligence, which is pivotal for addressing pressing environmental challenges.