High dust events, extreme meteorological phenomena, pose significant environmental and health risks. Their frequency is increasing due to climate change and drought, making accurate forecasting crucial for mitigation. The Middle East, particularly Israel, experiences frequent dust events due to its geographic location and diverse climate. Traditional physics-based numerical models struggle due to the multiscale nature of dust dynamics and the complex coupling between atmospheric particles and wind patterns. These models often suffer from low skill due to inherent uncertainties and model-dependent assumptions. Data-driven machine learning offers a promising alternative. Deep learning, particularly, has shown success in various fields and is increasingly applied in Earth and climate sciences. Previous studies applied machine learning to dust modeling and forecasting, but these models either focused on nowcasting (real-time data) or short-term forecasting, often with limited spatial or temporal range and questionable accuracy in defining dust events. The existing models often lack the capacity for long-term prediction (e.g., 24h or more) and reliable validation using ground truth data. This research addresses these limitations by developing a meteorology-based deep neural model for forecasting dust events in Israel 12–72 hours in advance. The model leverages a multi-task learning approach to overcome challenges posed by highly correlated meteorological data, relatively rare dust events, and meteorological noise and shift-invariance. The multi-task model simultaneously predicts regional PM₁₀ (using satellite data) and local PM₁₀, effectively incorporating valuable information from a correlated task to improve local dust event forecasting.

Numerous studies have investigated high dust events and their impacts. Research highlights the associations between high dust loading and adverse health effects, particularly cardiovascular and respiratory diseases. The Middle East, especially Israel, has been a focus due to its location within the global dust belt and diverse climate influenced by various synoptic systems. While physics-based numerical modeling is the traditional approach for dust forecasting, its limitations in handling multiscale dynamics and complex aerosol-wind interactions have spurred interest in data-driven alternatives. Several studies successfully used machine learning for dust modeling. Some focused on pixel-wise dust detection using techniques like Support Vector Machines (SVM) and Convolutional Neural Networks (CNN). Others concentrated on PM₁₀ forecasting using deep learning, but typically with short lead times and limited spatial coverage. Existing research often employed methods with questionable accuracy for dust event definition or lacked the capacity to generalize results to larger spatial or temporal scales due to limited training data. This study aims to address these limitations by developing a deep learning model capable of accurate longer-term forecasting, validated with ground truth data, and incorporating a multi-task learning framework to improve model performance.

This study utilizes a deep multi-task learning approach for forecasting dust events in Israel. The model employs 18 years (2003-2020) of regional meteorological data to predict local PM₁₀ concentrations (local task) and satellite-based regional PM₁₀ concentrations (regional task). The data includes half-hourly PM₁₀ measurements from 30 ground air quality monitoring stations across Israel, provided by the Israeli Ministry of Environmental Protection. Regional atmospheric data, obtained from the European Centre for Medium-Range Weather Forecasts Reanalysis 5th Generation (ERA5) database, covers the Mediterranean, Sahara Desert, and Arabian Peninsula. These data encompass various meteorological variables at different pressure levels (geopotential height, zonal and meridional winds, vertical velocity, potential vorticity, specific humidity, air temperature, and sea-level pressure), along with Aerosol Optical Depth (AOD) and regional PM₁₀ data. The model architecture consists of two deep neural networks trained simultaneously: Φreg (for regional PM₁₀ prediction) and Φloc (for local PM₁₀ prediction). Φreg uses an autoencoder (U-Net architecture) to learn a representation of the meteorological input, which is then used by Φloc to predict local PM₁₀. The model uses a 96-hour history of 12-hour time intervals as input. A code size of 512 elements was determined as optimal for balancing representation richness and overfitting risk. The choice of multi-task learning aims to overcome challenges related to data scarcity, high temporal correlations, and shift-invariance in meteorological data. The model uses an ordinal regression setup, categorizing PM₁₀ into 13 levels instead of a simple binary classification, enhancing its ability to capture the continuous nature of dust accumulation. Model interpretability is addressed using integrated gradients, to understand the influence of different meteorological variables on the model's predictions. This method measures how changes in input variables affect the model's output, revealing the importance of different features in forecasting dust events. The integrated gradients are computed by integrating gradients along a path from the input sample to a baseline (zero values, representing the 18-year pixel-wise average). The analysis focuses on the spatial and temporal distribution of feature importance to identify key meteorological patterns preceding dust events.

The deep multi-task learning model demonstrates skillful forecasting of dust events in Israel. Specifically: * **High accuracy in dust event detection:** The model achieves 76% recall in detecting dust events 24 hours in advance, with even higher accuracy (83%) for winter and spring events. The precision is 67%. * **Regional-scale dynamics drive forecasting skill:** Analysis shows that most misclassified events are driven by local dynamics occurring on smaller time scales (less than 24 hours), indicating that the regional meteorology is critical for longer-term prediction. Well-classified events are characterized by high PM₁₀ concentrations advancing from North Africa and the Arabian Peninsula towards Israel. Misclassified events showed no significant change in regional PM₁₀ until less than 24 hours before the event. * **Identification of key meteorological features:** Model interpretability analysis reveals the importance of lower-tropospheric winds and AOD in predicting dust events. The model focuses on meteorological signals originating in North Africa 3 days before the event in Israel, indicating that the model effectively captures the long-range transport of dust. Other features, such as sea-level pressure, also play a role, with the model distinguishing between large-scale dust storms originating from Libya and local events near Israel. * **Continuous PM₁₀ prediction:** The model's ordinal regression setup allows for continuous prediction of PM₁₀ levels, enabling a more nuanced evaluation than a simple binary classification.

The findings demonstrate the effectiveness of deep multi-task learning for forecasting dust events. The model's ability to accurately predict dust events 24 hours in advance, especially during winter and spring, has significant implications for mitigation efforts. The model's success in leveraging regional-scale meteorological patterns underscores the importance of considering large-scale dynamics for improved forecasting. The analysis shows that local dynamics account for the model's misclassifications, suggesting a need for incorporating high-resolution local data for more accurate short-term prediction. The interpretation of the model's output highlights lower-tropospheric winds and AOD as key predictive features, consistent with our understanding of the atmospheric processes governing dust events. The model's ability to distinguish between large-scale dust storms and local events further improves its predictive capacity.

This study presents a novel approach to dust event forecasting using deep multi-task learning. The model demonstrates superior performance compared to simpler methods, achieving high accuracy in predicting dust events up to 72 hours in advance. The integration of regional-scale meteorological data and a multi-task learning framework significantly enhanced the model's predictive power. Future research could focus on incorporating higher-resolution local data to improve short-term forecasting accuracy, exploring alternative model architectures, and expanding the model's application to other regions.

The study's reliance on historical data limits its ability to fully capture the variability of dust events. The model's performance could be influenced by changes in climate patterns or dust emission sources. The definition of a dust event, based on PM₁₀ threshold, may not perfectly capture the complex nature of all dust events. The model's regional coverage is primarily focused on the Middle East, which limits its direct generalizability to other geographic regions. Further research is needed to assess the model's robustness in the face of changing climate conditions and varying dust sources.