AirFusion: sensor data fusion for air quality monitoring

Since traditional air quality monitoring methods often rely on geographically sparse and costly air quality monitoring stations, image-based air quality methodologies are recently offering a compelling alternative that utilizes images from sources like satellites, traffic cameras, and even smartphones to monitor pollution levels by using estimation models, image-processing techniques, and deep learning models. In this paper, we introduce a novel, multimodal dataset designed to address the limitations of existing resources, which are often restricted in size, geographical coverage, and fixed-scene imagery, impeding the generalization of existing deep learning prediction models. Our AirFusion dataset comprises 9,411 images paired with synchronized meteorological and geospatial readings, collected by a portable commercial air quality sensor, from 179 diverse locations. Moreover, we introduce AirFusionNet, which leverages transfer learning of the pre-trained ResNet50 and uses an attention mechanism to extract features from both image and sensory data to predict five key air quality parameters: PM 1, PM 2.5, PM 10, temperature, and humidity. Our analysis of AirFusionNet establishes baseline results on this challenging dataset. Our model achieves a Root Square Mean Error (RMSE) of 10.44, 11.56, 13.18, 2.98, and 9.02 for PM1, PM2.5, PM10, temperature, and humidity, respectively, on the filtered day dataset and achieves an RMSE of 9.08, 9.75, 10.69, 2.63, and 8.49 for the same parameters, respectively, on the filtered combined day-night dataset.