地理与地理信息科学

环境空气质量监测是保障公众健康与支撑环境治理决策的关键基础，监控图像为获取近地大气颗粒物信息提供了新的数据来源。本文提出一种基于ResNet50-LSTM的时序监控图像空气质量推断方法，在不额外引入气象要素和显式时间编码的前提下，通过监控图像序列实现PM_2.5、PM₁₀和AQI的联合估计。该方法首先利用ResNet50提取图像多尺度空间特征，再引入长短期记忆网络(LSTM)建模污染过程的时间依赖，最后通过回归层输出多指标估计结果。以2021年上海市某固定监控摄像机图像与邻近地面自动监测站观测数据为例，构建8 132组图像与实测配对样本数据集，采用2折交叉验证系统评估模型在全天不同时段下的适应性、不同序列长度设置下的敏感性，以及在公开数据集中的迁移能力。结果表明，与仅使用单帧图像的ResNet50基准模型相比，该方法在3项空气质量指标上的平均决定系数R²提高约0.20,误差明显减小，在清晨和夜间等弱光时段仍能保持较高精度，体现出较强的时间适应性和对图像质量变化的鲁棒性；在公开数据集上，R²高达0.92、RMSE约为6.10μg/m³,优于多种代表性方法。研究表明，利用监控图像序列联合挖掘空间与时间特征用于空气质量推断可行且有效，可为依托现有视频监控基础设施构建高频、自动化、低成本的城市空气质量监测体系提供技术支撑。

Ambient air quality monitoring is fundamental for safeguarding public health and supporting evidence-based environmental governance.With the large-scale deployment of urban video surveillance systems, surveillance images have become a new data source for retrieving near-surface particulate matter information.This study proposes an air quality inference method based on a ResNet 50-LSTM framework, which jointly estimates PM_2.5,PM₁₀,and the air quality index(AQI) from image sequences without explicitly incorporating meteorological variables or time-encoding features.The method first uses ResNet50 to extract multi-scale spatial features from images, then employs a long short-term memory(LSTM) network to model the temporal dependence of pollution processes, and finally outputs multi-indicator estimates through a regression layer.Using hourly images from a fixed surveillance camera in Shanghai in 2021 and observations from a nearby ground automatic monitoring station, a dataset of 8 132 paired images and in situ measurements are constructed.A 2-fold cross-validation scheme is adopted to systematically evaluate the model′s adaptability across different times of day, its sensitivity to sequence length settings, and its transferability to a public dataset.The results show that, compared with a ResNet50 baseline model using only single-frame images, the proposed method increases the average coefficient of determination(R²) for the three air quality indicators by about 0.20,substantially reduces overall errors, and maintains high accuracy even during low-illumination periods such as early morning and night-time, demonstrating strong temporal adaptability and robustness to variations in image quality.On the public dataset, the model achieves a maximum R² value of 0.92 and a root mean squared error(RMSE) of about 6.10 μg/m³,outperforming several representative methods.These findings confirm that jointly mining spatial and temporal features from surveillance image sequences is a feasible and effective approach for air quality inference.This study provides technical support for constructing high-frequency, automated, and low-cost urban air quality monitoring systems based on existing video surveillance infrastructure.