Machine Learning Based Algorithms for Global Dust Aerosol Detection from Satellite Images: Inter-Comparisons and Evaluation

Identifying dust aerosols from passive satellite images is of great interest for many applications. In this study, we developed five different machine-learning (ML) based algorithms, including Logistic Regression, K Nearest Neighbor, Random Forest (RF), Feed Forward Neural Network (FFNN), and Convolutional (CNN), to identify in the daytime Visible Infrared Imaging Radiometer Suite (VIIRS) under cloud-free conditions on a global scale. order train ML collocated state-of-the-art detection product Cloud-Aerosol Lidar with Orthogonal Polarization (CALIOP) VIIRS observations along CALIOP track. The 16 M-band center wavelength ranging deep blue thermal infrared, together solar-viewing geometries pixel time locations, are used as predictor variables. Four sets training input data constructed combinations validation comparison results indicate that FFNN method all available variables best performing one among methods. It has an averaged accuracy about 81%, 89%, 85% over land, ocean whole globe, respectively, compared CALIOP. When applied off-track pixels, retrieves geographical distributions good agreement on-track well statistics. For further evaluation, our algorithms NOAA’s Aerosol Detection Product (ADP), which classifies dust, smoke, ash using physical-based reveals both similarity differences. Overall, study demonstrates potential methods proves these can be trained track then granule granule.