Evaluation of dust detection using multiple machine learning algorithms against physics-based approach on Visible Infrared Imaging Radiometer Suite (VIIRS) data

Abstract:

Dust has significant impacts on human activities and Earth environment. It affects air quality and visibility, influence weather system, modulating biogeochemical cycles, and perturbs radiative balance and thus further impacts Earth climate. Remote sensing of dust that provide wide spatiotemporal coverage is a major tool that used to study dust and its impacts on Earth. However, identifying dust posts challenges to the passive remote sensing communities. That is because traditional way of identifying dust, which utilizes dust spectrum signals, misclassifies very thin or thick dust to other kind of aerosols or clouds.

In this study, we use 5 different machine-learning (ML) and deep-learning (DL) based algorithm to identify dust (Logistic Regression, K Nearest Neighbor, Random Forest, Feed Forward Neural Network, and Convolutional Neural Network). We build our training data from the collocated Cloud‐Aerosol Lidar with Orthogonal Polarization (CALIOP) and Visible Infrared Imaging Radiometer Suite (VIIRS) products. Each satellite holds its advantage: CALIOP can detect dust well using its depolarization capability, but has a narrow swath. The dust classification from CALIOP can be used as a benchmark for identifying dust. VIIRS has wider footprint, which is more suitable to capture dust events, but is a passive sensor that is not equipped with the best tool to detect dust. We trained 5 ML/DL algorithms to predict dust under clear sky condition globally, then validated their performances on a subset of the collocated dataset that is not used as training data. Our analyses shows promising results of using FFNN to identify dust over both land and ocean. Further we compared our results to NOAA’s Aerosol Detection Product (ADP) on VIIRS, which is a popular product that classifies dust/smoke/ash using traditional physical-based algorithm. The comparisons are both on/off CALIOP track. Case studies were also performed to investigate the different performances between detecting dust originated from North Africa and East Asia.