Social sensing is an analytical method to study the interaction between human and space through extracting reliable information from massive volunteered information data. During the ongoing COVID-19 pandemic, there are a large number of Internet social sensing data. However, most of them lack geographic attribute. In order to resolve this problem, this paper proposes a convolutional neural network geographic classification model based on keyword extraction and synonym substitution (KE-CNN) which could determine the geographic attribute by extracting the semantic features from text data. Besides, we realizes the non-contact pandemic social sensing and construct the co-word complex network by capturing the spatiotemporal behaviour of a large number of people. Our research found that (1) mining co-word network can obtain most public opinion information of pandemic events, (2) KE-CNN model improves the accuracy by 5%–15% compared with the traditional machine learning method. Through this method, we could effectively establish medical, catering, railway station, education and other types of text feature set, supplement the missing spatial data tags, and achieve a good geographical seamless social sensing.

社会感知是一种通过从大量自愿信息数据中提取可靠信息来研究人与空间之间相互作用的分析方法。在持续的COVID-19大流行期间，有大量的Internet社会感知数据。但是，大多数缺乏地理属性。为了解决这个问题，本文提出了一种基于关键词提取和同义词替换（KE-CNN）的卷积神经网络地理分类模型，该模型可以通过从文本数据中提取语义特征来确定地理属性。此外，我们通过捕获大量人的时空行为，实现了非接触式大流行社会感知，并构建了共词复杂网络。我们的研究发现（1）挖掘共词网络可以获取有关大流行事件的大多数民意信息；（2）KE-CNN模型与传统的机器学习方法相比，其准确性提高了5％–15％。通过这种方法，我们可以有效地建立医疗，餐饮，火车站，教育等多种类型的文本特征集，补充缺失的空间数据标签，实现良好的地理无缝社会感知。