Deep multi-task learning for early warnings of dust events implemented for the Middle East

The study addresses the challenge of forecasting high dust-loading events, which pose significant environmental and public-health risks and are frequent in the Middle East, particularly Israel at the eastern Mediterranean margin of the global dust belt. Physics-based numerical dust models require complex coupling of multiscale atmospheric dynamics and aerosol processes and often show low skill due to uncertainties in dust emission, transport, and deposition. The authors posit that deep learning can leverage large-scale meteorological information to improve dust event forecasts with actionable lead times. They identify three key challenges hindering prior machine-learning applications to dust forecasting: highly correlated meteorological samples, rarity of dust events (class imbalance and low entropy), and the limited applicability of standard data augmentations due to the geo-referenced nature of meteorological fields. The study’s hypothesis is that a meteorology-based deep multi-task learning framework that jointly learns regional and local PM10 behavior can overcome these challenges and yield skillful forecasts 12–72 hours ahead, with robust validation against ground truth.

Prior work includes dust detection and forecasting using machine learning across Asia and the Middle East. Pixel-wise dust detection has been performed using SVM and CNN on satellite data, while PM10 forecasting over short lead times has been attempted using deep learning with station histories and spatial interpolation approaches. In Israel and the broader Middle East, studies have leveraged satellite Aerosol Optical Depth (AOD) with regression models for intra-daily PM10 prediction and CNNs for dust storm direction or dust source identification. However, many approaches focus on nowcasting or few-hour lead times, rely on definitions of dust events of questionable accuracy, or are trained on limited, non-generalizable samples. A comprehensive, meteorology-based deep neural forecast model with ≥24 h lead time validated against ground truth PM10 has been lacking. The paper situates its contribution within this gap, extending beyond shallow baselines (e.g., persistence, XGBoost) by leveraging regional-scale information through multi-task learning.

Data: Half-hourly PM10 measurements from 30 ground air-quality stations in Israel from January 2003 through December 2020 were obtained from the Israeli Ministry of Environmental Protection and aggregated to 3-hourly averages (≈52,543 samples, ~10% missing averaged points). Dust events are defined when local PM10 exceeds 2 standard deviations above a background (summer average), yielding a threshold of 65.2 μg m⁻³. About 13% of 3-hourly samples meet this criterion; most events occur in winter (33.7%) and spring (36.5%), fewer in fall (25.5%), and very few in summer (1.8%). For interpretability analyses, independent events are sequences of dust samples preceded by at least 96 h dust-free, yielding 356 events. Regional atmospheric fields covering 19–43°N, 18–45°E (Mediterranean, Sahara, Arabian Peninsula) were sourced from ERA5 (e.g., geopotential height, winds, vertical velocity, potential vorticity, specific humidity, temperature at multiple pressure levels; SLP; 10 m winds; total column water) and CAMS (dust AOD at 550 nm and regional PM10). Fields were interpolated to 0.5° resolution (49×49 grids), reindexed to 3-hourly to match local PM10, and standardized to anomalies relative to the 18-year mean. Missing spatial data (<0.5% pixels) were imputed by cubic splines. Problem setup: The primary (local) task is to forecast local PM10 level categories at lead times k ∈ {12, 24, …, 72} hours. To address heavy-tailed PM10 distributions and avoid brittle binary thresholds, the local forecast is posed as an ordinal regression over 13 PM10 bins (with the two highest bins above the dust-event threshold). The auxiliary (regional) task reconstructs satellite-based regional PM10 fields from the same encoded meteorological representation. Model architecture: A deep multi-task network shares a spatiotemporal encoder between tasks. Inputs are sequences of meteorological tensors over a 96-hour history sampled every 12 hours (8 frames per sample), producing a compromise between redundancy and temporal coverage. The encoder consists of stacked CNN blocks with batch normalization and ReLU, augmented by transformer-style residual blocks with multi-head attention and feed-forward layers that process spatial position encodings, yielding a 512-dimensional code per time step. These codes feed: - A decoder (stacked CNN + deconvolution layers) to reconstruct the regional PM10 field (autoencoder pathway). - A local classifier: concatenates the N codes, applies dropout (0.5), batch norm, fully connected layers with ReLU, and a final softmax to produce probabilities over 13 PM10 categories at the target lead time. Training: The model jointly optimizes local (ordinal) classification loss and regional reconstruction loss via a weighted multi-task objective, acting as regularization and inducing an inductive bias toward regional-scale dynamics. Mini-batches of 32 samples are used; the encoder outputs per-time-step codes, which are passed in parallel to decoder and classifier. Baselines include a Naive persistence classifier and a shallow extreme gradient boosting (XGBoost) classifier trained on flattened inputs. Evaluation: Metrics include recall and precision for dust-event detection at various lead times, confusion matrices across the 13 PM10 bins, and qualitative assessments of Φreg’s regional reconstructions. Interpretability uses Integrated Gradients to attribute the local PM10 probability to input pixels across variables, space, and time, with a zero baseline corresponding to the 18-year mean. Attributions are averaged over independent events to identify consistent precursors 24–72 h before onset. Hyperparameters and design choices: Code size of 512 optimized performance for the local task while preserving regional reconstruction quality. A 96 h window with 12 h spacing provided best performance, balancing redundancy and temporal context. Alternative auxiliary/self-supervised tasks (e.g., reconstructing inputs, forecasting future regional PM10, sequence sorting) were explored but did not outperform the chosen multi-task regional reconstruction.

- Forecast skill: At 24 h lead time, the model detects approximately 76% of dust events overall, with even higher recall for winter–spring events (~83%). Reported precision at 24 h is about 67%. Performance degrades with longer lead times; at the longest lead times, recall gains are traded for lower precision, eventually converging to around 51% for both recall and precision. - Baselines: The shallow XGBoost classifier performs only marginally better than persistence at ≥24 h lead times, underscoring the difficulty of learning from high-dimensional meteorological inputs with small-sample regimes. The proposed deep multi-task model maintains an increasing advantage over baselines as lead time grows, indicating successful generalization of large-scale spatial information. - Confusion structure: The 24 h confusion matrix is concentrated near the diagonal, indicating that most misclassifications are of small magnitude in PM10 level. - Misclassification analysis: Events that are misclassified tend to have weaker regional PM10 build-up and are driven by local sources (e.g., Negev or Sinai deserts) with short-range transport and rapid evolution, making them difficult to forecast beyond ~24 h using large-scale precursors. Well-classified events show clear regional PM10 enhancements over northern coastal Africa and the Arabian deserts advancing towards Israel. - Interpretability: Integrated Gradients reveals that 24–72 h prior to events, the model focuses on lower-tropospheric winds, AOD signals, and synoptic patterns over North Africa propagating eastward. Variables of highest importance at 24 h are lower-tropospheric winds (u, v), AOD, and SLP; temperature and specific humidity contribute less. Importance generally increases toward lower-tropospheric levels for wind, humidity, and geopotential height; potential vorticity shows mid-tropospheric peaks, suggesting roles in destabilization and momentum transfer. Signals from the Libyan deserts and eastern Mediterranean u10 distinguish large-scale versus local events. The model also associates higher dust probability with eastward-advancing high SLP over the Mediterranean, potentially reflecting complex seasonal or covariate relationships.

The findings demonstrate that multi-task deep learning leveraging regional PM10 reconstruction as an auxiliary task can extract robust, regionally coherent meteorological features that improve local dust-event forecasts at actionable lead times (≥24 h). By sharing an encoder between regional and local tasks, the model mitigates data scarcity, high temporal correlation, and the limited applicability of augmentation, improving generalization versus shallow or persistence baselines. The interpretability analysis aligns with known meteorology: precursors include lower-tropospheric cyclonic activity, strong near-surface winds over Saharan source regions, low humidity, relatively high temperatures, and elevated AOD, propagating eastward toward Israel. The model captures distinctions between large-scale Saharan intrusions and locally forced events, with the former more predictable at longer lead times. The unexpected positive association of advancing high SLP with dust probability suggests interactions with other variables or seasonal patterns; further research is needed to disentangle these relationships. Overall, the approach addresses the research question by showing that regional-scale meteorological signals contain predictive skill for local dust hazards, especially in winter–spring, and that multi-task learning enhances this skill by coupling local forecasts to regional aerosol dynamics.

This work introduces a meteorology-based deep multi-task learning framework that jointly reconstructs regional PM10 fields and classifies local PM10 levels to forecast dust events in Israel 12–72 hours in advance. The model achieves strong 24 h performance (≈76% recall overall; ≈83% in winter–spring; ≈67% precision), outperforming persistence and a strong shallow baseline. Interpretability analyses highlight physically consistent precursors—lower-tropospheric winds and AOD over North Africa—and elucidate why locally forced, short-range events remain challenging. Contributions include: (i) a validated, regionally informed deep-learning pipeline for dust-event early warning; (ii) an ordinal-regression framing that avoids brittle binary thresholds; and (iii) a transparent attribution analysis linking forecasts to meteorological drivers. Future work could: extend to broader regions and stations; integrate additional aerosol/speciation data; explore probabilistic forecasting and calibrated uncertainty; refine event definitions and multi-scale targets; and investigate auxiliary tasks or physics-informed constraints to further improve long-lead predictions and local-event skill.

- Data limitations: Dust events are relatively rare, leading to class imbalance and low sample entropy; meteorological fields sampled at 3-hourly resolution are highly temporally correlated, reducing effective sample size. CAMS/remote-sensing data availability (e.g., AOD) and differing temporal resolutions necessitate reindexing and may introduce inconsistencies. - Augmentation constraints: Standard image/video augmentations (e.g., flips, rotations) are not geophysically valid for geo-referenced fields, limiting data augmentation options. - Generalization to local events: Short-range, locally sourced events with weak or late-emerging regional precursors are often misclassified beyond ~24 h lead time, indicating limited skill for mesoscale/local dynamics not captured by regional inputs. - Event definition: Reliance on a PM10 threshold (2σ above summer background) influences labels; alternative thresholds or definitions can alter performance metrics. Ordinal regression mitigates but does not eliminate this sensitivity. - Modeling trade-offs: Larger code sizes improve regional reconstruction but risk overfitting the local task; chosen hyperparameters reflect a balance rather than an optimum for all objectives. Some architectural text indicates exploration of auxiliary/self-supervised tasks without consistent gains.