DustNet (v1): skilful neural network predictions of dust aerosols over the Saharan desert

Suspended in the atmosphere are millions of tonnes of mineral dust that interact with weather and climate. Accurate representation of mineral dust in weather models is vital, yet it remains challenging. Large-scale weather models use supercomputers and take hours to complete forecasts. Such computational burdens allow them to include only monthly climatological means of mineral dust as input states, inhibiting their forecasting accuracy. Here, we introduce DustNet, a simple, accurate, and fast forecasting model for predictions 24 h in advance of aerosol optical depth (AOD). DustNet is a custom-built 2D convolutional neural network (CNN) equipped with transposed convolution layers. The model is trained on selected ERA5 meteorology and past MODIS AOD observational data as inputs. Our design of DustNet ensures that the model trains in less than 8 min and creates predictions in 2.1 s on a desktop computer, without the need to utilise any graphics processing units (GPUs). Predictions created by DustNet outperform the state-of-the-art physics-based model at coarse 1°×1° resolution at 95 % of grid locations when compared to ground truth satellite data. The test results show that the daily mean AOD over the entire Saharan desert area is highly correlated with MODIS observational data, with Pearson's r2=0.91. Our results demonstrate DustNet's potential for fast and accurate AOD forecasting, which can easily be utilised by researchers without access to supercomputers or GPUs.