LLM-Enhanced Pollution Source Attribution: A Multimodal Framework Integrating Vision Transformers with Semantic Feature Extraction

Traditional pollution source attribution relies on expensive chemical speciation and handcrafted statistical models that cannot leverage the rich semantic information embedded in environmental monitoring metadata.We present EnhancedPollutionAttributorWithLLMs, a novel multimodal framework that integrates Large Language Models with advanced neural architectures for automated pollution source attribution.Our approach extracts semantic features from EPA Air Quality System site descriptions and combines them with spatiotemporal environmental data using rigorous geographic holdout validation.Experimental evaluation on 3,899 EPA monitoring records demonstrates that Vision Transformer with LLM-enhanced features achieves exceptional performance (R² = 0.9969, RMSE = 0.3116), significantly outperforming traditional approaches.Ablation studies reveal that LLM-derived semantic features contribute 1.2% improvement in R², while comprehensive interpretability analysis using SHAP provides actionable insights for environmental policy.This work establishes the first reproducible benchmark for LLM-enhanced environmental monitoring and demonstrates the potential for automated, large-scale pollution source attribution.