SkyFuseNet: A Cross-Attention Swin Transformer-Based Model for PM2. 5 Estimation Using Ground-Level Images and Air Quality Readings
Document Type
Conference Proceeding
Publication Date
2026
Abstract
Fine particulate matter (PM2.5) is one of the most harmful air pollutants, and long-term exposure can lead to cardiovascular and respiratory diseases. The estimation of PM2.5 concentration using traditional methods is a big challenge, especially for regions where it is an economic burden to deploy large numbers of monitoring stations and sensors. Current image-based estimation methods mostly rely on CNNs, which only capture local features, without using concurrent sensor data, which is also correlated with the concentration of PM2.5. In this paper, SkyFuseNet is presented as an efficient multimodal, multitask deep learning model with 49.63 million parameters. The model uses a Swin Transformer, a learnable attention-pooling neck, a multi-layer perceptron for the sensor encoder, and a cross-attention fusion module that employs an intermediate fusion strategy. The model uses a dual-head structure, where information from the later stage passes through a dense layer and is then split into regression and classification tasks. The output is then combined using a weighted loss function. The model was evaluated on the India and Nepal Air Pollution Image Dataset, where the model was able to achieve a test MAE of 3. 0 3μg/m3, RMSE of 5. 7 0μg/m3, R2 of 0. 9 9 8 1, MAPE of 3.10%, and AQI classification accuracy of 91.45%.
Recommended Citation
Abdelsalam, M. S., & Sayed, G. I. (2026, July). SkyFuseNet: A Cross-Attention Swin Transformer-Based Model for PM2. 5 Estimation Using Ground-Level Images and Air Quality Readings. In 2026 Intelligent Methods, Systems, and Applications (IMSA) (pp. 489-494). IEEE. https://doi.org/10.1109/IMSA70415.2026.11700341