Colours-of-the-Wind (COLD): A novel framework for image-based air quality prediction using deep learning and spatial-temporal analytics
Year: 2025
Type: dissertation
Abstract: Urban air pollution is a critical global issue that negatively affects public health and the environment. Traditional air quality monitoring relies on physical monitoring stations, which are limited in geographic coverage and expensive to maintain. This creates significant gaps in air quality data, especially in remote areas, water bodies, or underserved urban zones.This thesis presents the "Colours-of-the-Wind" (COLD) framework, which aims to address the limitations of traditional air quality monitoring systems. COLD integrates meteorological camera images with structured data from air quality monitoring stations and historical records to provide accurate, real-time air quality predictions. Unlike current image classification systems, which can only classify the content of the image, this framework enhances monitoring capabilities in regions lacking physical infrastructure by employing machine learning techniques. The research makes several significant contributions:•Innovative Data Use: COLD utilises meteorological camera images as a new data source for air quality prediction, demonstrating the conversion of unstructured images into structured data.•Integration of Data Sources: The framework combines visual and numerical data, allowing for air quality estimation in areas without physical sensors, extending monitoring to remote regions.And three Hybrid Prediction Algorithms:•CORD: A time-series prediction model that combines CNNs and recurrent neural networks (RNNs) for temporal forecasting.•SAID: Spatial analysis algorithms integrating CNNs and artificial neural networks (ANNs) for regional air quality estimation.•STUN: A spatial-temporal prediction model incorporating CNNs, RNNs, and ANNs for comprehensive forecasting.The COLD framework was validated using over 80,000 images from 14 meteorological cameras across Hong Kong and pollutant concentration data from the Environmental Protection Department (EPD). The framework achieved an 85-97% accuracy in classifying air quality conditions and provided 6-hour air quality health index (AQHI) forecasts with over 70% accuracy.Future improvements will focus on incorporating infrared and thermal imaging technologies, significantly enhancing performance in low-light scenarios. Additionally, we aim to develop an automated image classification system that correlates established pollution conditions with the gathered images, thereby reducing manual input and improving efficiency.
Author: Chan, Kenneth
Language: English
Subfield: Environmental Engineering
Field: Environmental Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7204361286