Dust storm detection for ground-based stations with imbalanced machine learning

Abstract

Dust storms, common meteorological hazard in arid and semi-arid regions, have significant environmental and societal impacts. Rapid and accurate detecting dust storms is critical for early warning systems. Over the past few decades, dust storm detection primarily relied on satellite remote sensing techniques using multi-channel imagery, but these methods have limitations in temporal resolution. With the recent expansion of China's observation network, the dense distribution of ground-based sensors offers a promising data source for real-time dust storm detection. This study proposes a machine learning approach to detect dust storms using ground-based sensor networks. By combining undersampling strategies and ensemble algorithms, this method improves model's performance in detecting dust storms. Compared with the state-of-the-art models, this approach improves the Recall rates for different dust storm levels by 24.32% and the G-Mean by 18.58%, achieving superior dust storm detection performance. This approach can offer the near-real-time, hourly updated dust storm detection products.