DustNet: skillful neural network predictions of Saharan dust

TL;DR

The results show DustNet has a potential for fast and accurate AOD forecasting which could transform the understanding of dust impacts on weather patterns and create predictions in 2 seconds on a desktop computer.

Abstract

Suspended in the atmosphere are millions of tonnes of mineral dust which interacts with weather and climate. Accurate representation of mineral dust in weather models is vital, yet remains challenging. Large scale weather models use high power supercomputers and take hours to complete the forecast. Such computational burden allows them to only include monthly climatological means of mineral dust as input states inhibiting their forecasting accuracy. Here, we introduce DustNet a simple, accurate and super fast forecasting model for 24-hours ahead predictions of aerosol optical depth AOD. DustNet trains in less than 8 minutes and creates predictions in 2 seconds on a desktop computer. Created by DustNet predictions outperform the state-of-the-art physics-based model on coarse 1 x 1 degree resolution at 95% of grid locations when compared to ground truth satellite data. Our results show DustNet has a potential for fast and accurate AOD forecasting which could transform our understanding of dust impacts on weather patterns.

Figures (13)

• Figure 1: Schematic representation of the DustNet model. Each of 6,205 inputs is first padded with a border of zeros using ZeroPadding2D (light blue arrow) to increase dimensionality and allow the convolution windows to detect the borders. The features are then extracted by 2D convolution window (pink arrows) which decreases dimensionality while increasing the number of trainable parameters. Then deconvolution is applied (yellow arrow) by including a 2D transpose network, which increases the size of the input (dark blue arrows) while maintaining connectivity between the layers. The output is then cropped back to match the initial input size (green arrow) and represents a 24-hr ahead prediction.
• Figure 2: Metrics indicating model performance. Results for 24-hour predictions of AOD values (2020-2022) compared with the ground truth data from MODIS. The RMSE for DustNet (A) and CAMS (B), where the brighter the colour the smaller the error. Note, that the maximum error for DustNet is 0.62 AOD (medium green shades), while the maximum RMSE for CAMS reaches above 1.2 AOD (dark blue). In C) the difference in RMSE between CAMS and DustNet where all yellow to deep brown shades indicate the advantage of DustNet, while the blue shades indicate the advantage of CAMS. White grid cells indicate locations where both of the models performed equally when compared to the ground truth data. Note the lack of deeper blue shades and the dominance of yellow and brown grid cells where DustNet outperformed CAMS. D) and E) show the ACC for DustNet and CAMS respectively, where values above 0.6 (bright to white) indicate a valuable forecasting capability, while lower values (green to dark blue) indicate little to no predictive value. The ACC values in darkest blue indicate a misleading forecast.
• Figure 3: Annual and quarterly mean AOD for 2020-2022. Mean AOD values calculated from 24-hr predictions. The left column represents AOD values from MODIS observations, while model predictions from DustNet are in the middle and from CAMS in the right column. Row A) compares the 3-year annual mean AOD between the observations and models, where DustNet skillfully captures the locations of the main dust events and the higher AOD around Nigeria and the Gulf of Guinea. In row B) the 3-year mean AOD for Q1: January - March, where the influence of the Harmattan wind has a visible effect on the mean AOD with a south-westward transport of mineral dust from the main generation site of the Bodélé Depression (dark blue). The effect of this transport is clearly picked up by our model. An increased AOD from biomass burning is also captured below 5° N. In row C) these same means are shown but for Q2: April - June where again the DustNet predictions skillfully capture the change in wind direction and westward aerosol transport in comparison to MODIS. Row D) shows that both models, CAMS and DustNet skillfully detected the northward shift of mean AOD transport during Q3: July-September. Here, CAMS forecasts tend to overestimate the AOD along the 20° latitude, but represent biomass burning related AOD around equator more realistically than DustNet, whose smoother contours seem to overestimate the AOD below 10° N. In row E) the seasonal decrease in aerosol activity for Q4: October - December is skillfully captured by both models when compared to observations from MODIS. Here, DustNet captures the position of the Bodélé Depression more accurately than CAMS and shows the lack of aerosol generation from the eastern locations. Note here the change in the colour-bar range.
• Figure 4: Mean AOD predictions for each day of the year (2020-2022) at chosen locations. Shown are daily means (2020-2022) of AOD predictions from DustNet (golden line) and CAMS (light-sea-green line) as compared to MODIS (black line) and climatological mean (dotted line). At all four locations predictions from DustNet are closer to MODIS values than CAMS forecasts. An increase in AOD can be seen in the first 90 days of the year in A) the Bodélé Depression, with lower but still elevated values towards B) Kano and C) Gulf of Guinea. These elevated AOD values during quarter 1 are not observed in D) Nouadhibou, which is consistent with the south-western direction of the Harmattan wind. DustNet also predicts daily and seasonal AOD variability at each site more skillfully than CAMS, whose forecasts tend to stay closer to or below the climatological mean. Both models struggle to fully capture the highest AOD peaks recorded by MODIS at the westmost location - Nouadhibou, however the DustNet model replicates these peaks better than CAMS. The background image, showing the position of the chosen locations (top), shows the December view of Blue Marble available from NASA https://visibleearth.nasa.gov/collection/1484/blue-marble?page=4.
• Figure 5: Fig. S1: Daily spatial mean AOD (2020-2022) regressed between model predictions and MODIS data.S1. Linear regression with corresponding y equation, Pearson's r$^2$ and p values were calculated for daily spatial mean AOD over the Sahara for 2020 - 2022. Shown in A) AOD prediction results from DustNet correspond to MODIS data well with high r$^2$ = 0.91, and only a slight tendency to overestimate higher AOD. In B) the mean AOD forecasts from CAMS are shown to correspond with MODIS data well, r$^2$ = 0.71 though, with more frequent tendency to underestimate both low and high AOD values. Results from both predictions are highly significant with p$<0.0001$.