FewShot-AIRNET: Temporal 2D-Variation Modeling for Air Pollution Forecasting

Urbanization and industrialization have led to a rapid increase in atmospheric air pollution, highlighting the urgent need for intelligent systems to monitor and predict pollution levels. Current state-of-the-art (SOTA) models rely on one dimension (1D) air pollutant data for forecasting, but the limited availability of such data poses significant challenges in extracting meaningful patterns. To overcome this limitation, we introduce FewShot-AIRNET, a novel approach that extends air pollution prediction into two dimension (2D) space by transforming 1D sequences into 2D tensors, specifically designed for few-shot learning settings. The model employs TimeBlock as a general backbone, which effectively captures multi-periodicity and complex patterns within the transformed 2D tensors, enhancing prediction accuracy. Our proposed FewShot-AIRNET significantly outperforms four SOTA models in air pollutant forecasting, especially in scenarios with limited data. Specifically, FewShot-AIRNET achieves SOTA improvements, with an average of 12.10% on AotiZhongxiz, 5.08% on Tiantan, and 15.31% on Wanliu in terms of mean squared error (MSE) using 5% training data. By leveraging small datasets and focusing on temporal 2D-variation modeling, FewShot-AIRNET presents a more robust solution for long-term air pollution forecasting, addressing the critical need for accurate predictions in data-scarce environments. The code for this repository is available: https://github.com/Prasanjit-Dey/FewShot_AIRNET.