PMNet: Improving Aerosol Predictions using Deep Neural Nets for Limited Ground Stations

Observational datasets which accurately measure aerosol optical depth (AOD) on the ground are not readily available at all locations, instead reanalysis data is used as a substitute for ground measurements. Predictions from reanalysis however are not very accurate and at significantly lower resolutions than measurements by sun-photometers. This study examines the use of convolutional neural networks (CNNs) to improve AOD predictions. MERRA-2 reanalysis data with a CNN is used to predict ground truth at AERONET sunphotometer stations measuring AOD. Implementing CNNs to improve AOD predictions yielded improvements on a global scale across all test statistics examined. The improvement was greatest amongst extreme events, classified as an observed AOD greater than 0.7, where the error associated with AOD forecasts from reanalysis data decreased by 30% for the RMSE and 56% for the MAE. However, improvements were not made across all locations. The neural network performed particularly poorly in Australia where the error increased across all performance measures at all locations with the exception of Canberra, Australia. Further work is in progress to examine different neural network architectures such as RNNs alongside CNNs, and additional input features considering more variables than just AOD, to predict aerosols, but the initial results are encouraging.