Development Of A Deep Neural Network Based Method For Quality Control Of Modis Maiac Aerosol Data For Aerosol Modeling Applications

• Main Author: Aldridge, Joel Braxton
• Format: Text
• Language: unknown
• Published: UND Scholarly Commons 2021
• Subjects: AERONET; AOD; Machine Learning; MAIAC; Neural Network; Quality Control; Aerosol Robotic Network
• Online Access: https://commons.und.edu/theses/4150; https://commons.und.edu/cgi/viewcontent.cgi?article=5151&context=theses

Quality-assured satellite aerosol data have been shown to improve aerosol analysis and forecasts in Chemical Transport Models. However, biases present in the satellite-based aerosol data can also introduce non-negligible uncertainties in the downstream aerosol forecasts and impact model forecast accuracy. Therefore, in this study we evaluated uncertainties in Moderate Imaging Spectroradiometer (MODIS) Multi-Angle Implementation of Atmospheric Correction (MAIAC) aerosol products and developed a deep neural network (DNN) based method for quality control of Terra and Aqua MODIS MAIAC Aerosol Optical Depth (AOD) data using the version 3 level 2 AErosol RObotic NETwork (AERONET) data as the ground truth. This method is done using 14 years of Aqua MODIS (2002-2016) and 16 years of Terra MODIS (2000-2016) MAIAC data which are collocated with the AERONET observations. The resulting trained network, which is tested on one year of Aqua/Terra data, can detect and significantly reduce noisy retrieval in MAIAC AOD data resulting in an approximate 31%/27% reduction in Root-Mean-Square-Error in Aqua/Terra MODIS MAIAC AOD with an associated 14%/16% data loss. A sensitivity study performed in this effort suggests that reducing the number of output categories and hidden layers can significantly improve performance of the deep neural network in this case. This study suggests that DNN can be used as an effective method for quality control of satellite based AOD data for potential modeling applications.