Abstract:Particulate pollution threatens the ecological environment, air quality, and public health. Therefore, it has become an increasing concern for the public and governments in recent decades. In this study, a full coverage PM2.5 (aerodynamic diameter of less than 2.5 $\mu \text{m}$ ) estimation strategy is proposed based on spatiotemporal machine learning approaches, including the convolutional neural network with long short-term memory (CNN-LSTM) and random forest (RF). The RF estimates PM2.5 by considering the features of a single pixel, while the introduction of the CNN-LSTM (size of 7 $\times7\,\,\times4$ ) assists in exploiting the spatiotemporal correlation of surrounding pixel features. Compared with linear models and empirical spatiotemporal weight methods, our CNN-LSTM+RF avoids the uncertainty and complexity owing to actual measurements of the surrounding sites. In addition, full coverage is achieved using both satellite data and reanalysis data. Results showed that the root mean square error (RMSE) and coefficient of determination ($R^{2}$ ) of the CNN-LSTM+RF were 12.790 $\mu \text{g}/\text{m}^{3}$ and 0.910, respectively, in sample-based cross-validation (CV). From the perspective of the season, the best performance of the CNN-LSTM+RF was found in autumn ($R^{2}$ of 0.915), and the lowest was in summer ($R^{2}$ of 0.848). In the meantime, for the different regions of China, the CNN-LSTM+RF also showed stable performance. The proposed method can generate high-precision continuous PM2.5 distribution maps that provide beneficial support for improving environmental and public health, and provide a reference for using deeper networks.

Particle Pollution Estimation from Images Using Convolutional Neural Network and Weather Features

Estimation of Atmospheric PM2.5 Based on Photos and Deep Learning.

Image-Based PM2.5 Estimation and Its Application on Depth Estimation

Image-based air quality analysis using deep convolutional neural network

Visually Estimating and Forecasting PM2.5 Levels Using Hybrid Architecture Deep Neural Network

Ensemble Of Deep Neural Networks For Estimating Particulate Matter From Images

A Shallow ResNet with Layer Enhancement for Image-Based Particle Pollution Estimation.

Estimate Air Quality Based On Mobile Crowd Sensing And Big Data

Particle Pollution Estimation Based on Image Analysis.

PM2.5 Estimation Based on Image Analysis.

Applying Traffic Camera and Deep Learning-Based Image Analysis to Predict PM2.5 Concentrations

A Deep CNN-LSTM Model for Particulate Matter (PM2.5) Forecasting in Smart Cities

Densely Connected Convolutional Networks with Attention Long Short-Term Memory for Estimating PM2.5 Values from Images

A Spatial Neighborhood Deep Neural Network Model for PM2.5 Estimation Across China

PM2.5 concentration estimation using convolutional neural network and gradient boosting machine

Image-based Air Pollution Estimation Using Hybrid Convolutional Neural Network

Full Coverage Estimation of the PM Concentration Across China Based on an Adaptive Spatiotemporal Approach

Feature enhancement and fusion for image-based particle matter estimation with f-mse loss

Enhancing real-time PM 2.5 forecasts: A hybrid approach of WRF-CMAQ model and CNN algorithm

On Estimating Air Pollution From Photos Using Convolutional Neural Network

Real-time image-based air quality estimation by deep learning neural networks