Using social media for environmental insight: a multi-model deep learning framework approach

Monitoring ecological conditions and detecting environmental issues are critical for safeguarding human health and sustainable development. Previous studies have shown that social media data can complement traditional methods, such as remote sensing, by capturing public sentiment; however, existing text-processing frameworks for social media face challenges in constructing comprehensive keyword lexicons and achieving efficient information extraction. To address these limitations, this study proposes a multi-model Weibo text-processing framework for environmental issues. An iterative expansion strategy was employed to rapidly develop an extensive environmental topic lexicon, and a named entity recognition model based on the BERT – BiGRU – CRF architecture was implemented to extract environmental and geographic keywords with high efficiency. Using this framework, we built a lexicon of 1,019 environmental topic keywords and collected 47,737 posts from Weibo between June 2023 and June 2024, each of which was analyzed for topic classification, geographic location, and sentiment polarity. The experimental results demonstrate that the framework achieves F1 scores of 0.90 for topic classification, 0.80 for sentiment analysis, and over 0.87 for both environmental and location keyword recognition. Significant negative sentiment was detected in Beijing and Chengdu, while the Shandong – Jiangsu region exhibited consistently strong positive sentiment. High-frequency keyword analysis revealed that air pollution (31.54%), environmental protection (14.85%), and ecological destruction (13.09%) were the top concerns. Water pollution case studies further validate the framework’s ability to rapidly localize pollution events. Overall, the proposed framework significantly enhances the timeliness and accuracy of environmental public opinion monitoring, providing novel insights and decision-support capabilities for policymakers.