Artificial intelligence empowered image-based ambient pollution estimation for smart city

Abstract

Air pollution monitoring plays an important role in a smart city, as accurate ambient pollution measurement is important in policy-making and guiding outdoor activities. Governments rely on accurate measurements to adjust the policy to minimize the total pollution emission, and citizens, especially vulnerable people, can avoid long-time exposure to severe air pollution. However, due to the high cost of constructing and maintaining air quality monitoring stations (AQMSs), there are only limited AQMSs installed in a city. Sparsely distributed AQMSs cannot provide fine-grained pollution information, and the citizens who live far from the AQMSs cannot have accurate and timely pollution information about their surrounding environment. Instead of building new high-cost AQMSs everywhere, using stationary cameras-taken images as an alternative proxy of the pollution levels gradually becomes a hot topic. Compared to the AQMSs, the widely distributed stationary cameras can provide timely and high-density images almost everywhere in the city, and no extra implementation is needed. This thesis demonstrates a comprehensive study of image-based pollution estimation from three different aspects, including sequential information extraction, cross-camera information sharing, and limited information inferring, and provides a comprehensive path from theoretical modeling to real-world application in a smart city.

First, we propose a deep learning model to estimate ambient pollution by extracting temporal-domain features from sequential images directly. The temporal-domain features extracted from the past consecutive images contain rich information and can be used to improve the model's capability for pollution estimation. The numerical results show that the model can help to flatten the estimation fluctuations and reduce the estimation errors.

Second, we propose a personalized method to enable information sharing between different cameras. Since each camera can only provide limited samples and sceneries, the locally trained estimation models can easily overfit. Hence, we propose to merge all images together to train a single global model first, so a more general knowledge of mapping images to pollution estimation can be learned. Then the global model is fine-tuned on each camera separately to adapt the local features.

Third, we propose a contrastive approach to infer the ambient pollution even when the labeled images are limited. The proposed model estimates the pollution differences between images, and a probabilistic pollution estimation can be achieved by comparing each image from the limited set with multiple images taken by other cameras. Furthermore, to enhance the model's capability in extracting pollution-related features, the scenery images are synthesized and used to eliminate the scene-related features from each image.

In summary, this thesis explores the possibility of image-based pollution estimation from a single image to sequential images, from a single camera to multiple cameras, and from sufficient labeled images to limited labeled images with the help of deep learning techniques, which demonstrate an effective and practical method for the real-world ambient pollution estimation based on the stationary camera-taken images.