Retrieval of Aerosol Optical Thickness from Satellite Images through Machine Learning

摘要

Aerosol measurements are routinely obtained through ground observation networks of sun-photometers such as the AERONET. Such implementation gives accurate and periodic observations of aerosol properties at each ground-based station. Satellite observations can provide wide areal coverage of aerosol optical thickness (AOT). However, the retrieval of AOT from satellite data usually involves execution of complex radiative transfer codes and is complicated by the influence of different ground covers. In this study, the retreival of AOT at 550 nm (AOT550) from satellite images using machine learning is presented. Simulated satellite signals were generated using the Py6S. a python interface to the 6S (Second Simulation of a Satellite Signal in the Solar Spectrum) Radiative Transfer Model (RTM). The simulated data were based on the different permutations of varying geometrical conditions, atmospheric models, spectral conditions and ground reflectances. A feedforward neural network (NN) with backpropagation algorithm was implemented. The dataset was used for training, validation and testing of the NN model. The trained NN model was then applied to retrieve the value of AOT550 from the satellite images. The plot of actual against retrieved AOT550 had achieved a coefficient of determination (R~2) of 0.716. The machine learning method has shown a promising capability for the spatial and temporal monitoring of aerosols distribution, especially during episodes of biomass burning.