Exploring sentiment dynamics and their driving factors in megacity residents’ environmental complaints through deep learning and multimodal data

As urbanization continues to accelerate, ecological challenges in cities have intensified, resulting in a growing number of environmental complaints from residents. Effectively exploring the potential public emotions behind complaints is helpful for improving the urban environmental governance capacity. However, most existing studies emphasize the drivers of environmental complaints, while giving limited attention to the mechanisms underlying residents' negative sentiment (RNS). In addition, the influence of the built environment on RNS remains insufficiently examined. Taking Guangzhou as a case study, this research applies the BERT model to conduct sentiment analysis on environmental complaint text data. Furthermore, a Light Gradient Boosting Machine-SHapley Additive exPlanation (LGB-SHAP) model is employed to characterize the nonlinear associations between RNS and its potential drivers. Results indicate that RNS is predominantly concentrated in the central built-up areas of Guangzhou, with stronger expressions observed during nighttime. Spatial overlap is evident between high-density complaint zones and RNS hotspots, highlighting critical areas for enhanced environmental surveillance. The plot ratio emerges as the strongest determinant of RNS. Moreover, the plot ratio often interacts with other factors, exerting either amplifying or mitigating effects on RNS within different threshold ranges. The influence of driving factors also varies across different land use types, where plot ratio and openness exert dominant impacts. This study integrates multimodal data to detect the emotional dynamics of residents’ environmental complaints and elucidates the driving mechanisms of RNS in relation to the built environment and socioeconomic factors, thereby providing a reference for more targeted and responsive urban environmental governance strategies.