Quantifying noise complaint intensity via multi-LLM and revealing vertical greenery mitigation thresholds.

Urban noise pollution has become a critical environmental issue affecting residents' quality of life. Existing research struggles to capture heterogeneity in noise sensitivity across different populations and cannot precisely quantify the mitigation effects of green space on noise. First, this study proposes an LLM-based method to quantify noise complaint intensity. Using four large language models, we constructed a benchmark dataset of noise complaint intensity. By replacing direct scoring tasks (where LLMs underperform) with comparison tasks (where LLMs excel), the fine-tuned RoBERTa model (a variant of BERT) achieved 61.48% accuracy when transferred to new datasets. This approach enables a fine-grained understanding of complaint data and captures large-scale distributions of resident complaint intensity. Second, green space exposure measurement integrated the NDVI from Sentinel-2 imagery and the 3D Green View Index (including tree and grass pixel ratios) from street view images. Generalized Additive Models results revealed significant vertical structural differences and non-linear threshold effects. The overhead greenspace coverage (NDVI) and grass coverage (low-level green space) showed highly similar trends with strong positive correlation: at low exposure levels, low-lying vegetation not only failed to mitigate noise but actually intensified complaint severity; whereas high-level green space (tall vertical greenery) represented by trees demonstrated strong noise suppression effects, though marginal benefits diminished at higher coverage (>0.12). In contrast, road density showed a reversal pattern of "initial suppression then amplification" (threshold ≈ 0.01). These findings indicate that constructing tall vertical green spaces exceeding effective thresholds is a crucial strategy for protecting vulnerable populations in urban planning.