Abstract:Reanalysis datasets combining numerical physics models and limited observations to generate a synthesised estimate of variables in an Earth system, are prone to biases against ground truth. Biases identified with the NASA Modern-Era Retrospective Analysis for Research and Applications, Version 2 (MERRA-2) aerosol optical depth (AOD) dataset, against the Aerosol Robotic Network (AERONET) ground measurements in previous studies, motivated the development of a deep learning based AOD prediction model globally. This study combines a convolutional neural network (CNN) with MERRA-2, tested against all AERONET sites. The new hybrid CNN-based model provides better estimates validated versus AERONET ground truth, than only using MERRA-2 reanalysis.

What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the accuracy of aerosol optical depth (AOD) prediction on a global scale. Specifically, the paper aims to improve the estimation of AOD, especially in extreme events, by combining deep - learning methods (especially convolutional neural networks, CNN) and NASA's MERRA - 2 reanalysis data set. ### Problem Background 1. **Existing Challenges**: - The MERRA - 2 data set has biases when compared with the true values provided by ground - based observatories (such as AERONET), especially in cases of high AOD values, where MERRA - 2 often underestimates AOD. - The spatial resolution of satellite data is low, resulting in large errors when estimating AOD far from grid points. - Aerosol data are difficult to obtain in some regions, limiting the prediction accuracy of reanalysis models that rely on satellite data in these regions. 2. **Research Motivation**: - Previous studies have found significant biases in AOD predictions by MERRA - 2 at certain locations, especially in parts of Asia and Africa. - Existing machine - learning methods mainly rely on simple multi - layer perceptrons (MLP) and do not fully utilize the spatial feature extraction capabilities of convolutional neural networks (CNN). ### Solution The paper proposes a CNN - based hybrid model that uses MERRA - 2 reanalysis data as input and is trained to improve AOD prediction. Specific practices include: 1. **Model Structure**: - The CNN model consists of two main parts: feature learning and classification prediction. - The feature learning part is composed of convolutional layers, pooling layers, and normalization layers, which are used to extract spatial features. - The classification prediction part consists of multiple fully - connected layers (Dense layers) and Dropout layers to prevent over - fitting and improve generalization ability. 2. **Training and Evaluation**: - Use the real AOD data from AERONET sites for training and validation. - The training data is divided into two parts: one part is used for predicting normal AOD values, and the other part is used for predicting extreme AOD values. - Use the mean squared error (MSE) as the loss function and the Adam optimizer for back - propagation. ### Main Results 1. **Global Comparison**: - Compared with using only MERRA - 2 data, the CNN model significantly improves the AOD prediction accuracy on a global scale. - For extreme AOD events, the performance of the CNN model is particularly outstanding, with the RMSE decreasing from 1.098 to 0.775 and the MAE decreasing from 0.986 to 0.436. 2. **Regional Performance**: - Although the performance in Southeast Asia (such as Indonesia) is still not as good as in other regions, compared with MERRA - 2, the error of the CNN model has been reduced. - At some specific sites (such as Hornsund, Norway), despite individual anomalies, overall, the CNN model still shows better predictive ability. ### Conclusion This paper successfully improves the prediction accuracy of aerosol optical depth on a global scale, especially in extreme events, by introducing deep - learning techniques, especially convolutional neural networks. This improvement is of great significance for climate models and environmental monitoring.

Deep learning for Aerosol Forecasting

Deep learning the atmospheric boundary layer height

Deep Learning with Pretrained Framework Unleashes the Power of Satellite-Based Global Fine-Mode Aerosol Retrieval

Deep-AIR: A Hybrid CNN-LSTM Framework forFine-Grained Air Pollution Forecast

Wide and Deep Learning Model for Satellite-Based Real-Time Aerosol Retrievals in China

Improving Prediction of Marine Low Clouds with Cloud Droplet Number Concentration and a Deep Learning Method

Deep Feature Gaussian Processes for Single-Scene Aerosol Optical Depth Reconstruction

A Machine Learning Approach to Retrieving Aerosol Optical Depth Using Solar Radiation Measurements

Machine Learning Approach for Spatially and Temporally Resolved PM2.5 Exposures in the Continental United States

Reconstructing aerosol optical depth using spatiotemporal Long Short-Term Memory convolutional autoencoder

Evaluation of hybrid deep learning approaches for air pollution forecasting

Aerosol optical depth retrieval using scaled digital number (DN) values of multi-spectral satellite and a generating adversarial model based on deep learning application

Recurrent U-net: Deep learning to predict daily summertime ozone in the United States

Deep Air Quality Forecasting Using Hybrid Deep Learning Framework

Forecasting Smog Clouds With Deep Learning

Himawari-8/AHI Aerosol Optical Depth Detection Based on Machine Learning Algorithm

Transfer Learning with ResNet3D-101 for Global Prediction of High Aerosol Concentrations

Deep Learning of Radiative Atmospheric Transfer with an Autoencoder

Deep Inferential Spatial-Temporal Network for Forecasting Air Pollution Concentrations

Retrieval of Atmospheric Aerosol Optical Depth From AVHRR Over Land With Global Coverage Using Machine Learning Method

Hybrid optimized deep recurrent neural network for atmospheric and oceanic parameters prediction by feature fusion and data augmentation model