Research published in Newswise – engineering Introduces innovative high-precision aerosol algorithms for geostationary weather satellites. Titled “Deep Learning and Transfer Learning Hybrid Aerosol Search Algorithms for FY4-AGRI: Development and Validation in Asia,” this research article introduces a hybrid approach that addresses the challenges that bring flexibility in traditional physical algorithms, and a hybrid approach that addresses the limited number of challenges of terrestrial sun hologmeter sites required for habit learning.
The proposed method uses an aerosol optical depth (AOD) search algorithm that utilizes deep learning and transfer learning. The algorithm integrates the core ideas of dark targets and deep blue algorithms to facilitate the feature selection of machine learning algorithms. This process involves two main steps: First, we employ a 10 min advanced Himawari Imager (AHI) AOD as a target for building deep neural networks (DNNs) using residual networks. Then fine-tune the DNN parameters using AOD data from the ground station equipped with an 89 solar photometer.
Independent verification demonstrates the high accuracy of the algorithm when acquiring Advanced Earth Coupled Radiation Imager (AGRI) AOD, achieving a measurement coefficient of 0.70 and an average bias error of 0.03, with 70.7% of the data achieving 70.7% within the expected error range. In particular, during extreme aerosol events, searches properly capture the temporal evolution of these phenomena.
This study demonstrates the great potential of combining physical approaches in geoscience analysis with deep learning. Furthermore, the proposed algorithms show versatility and provide applicability to other multispectral sensors.
The main contributions of this study are as follows:
1. Hybrid deep learning and transfer learning aerosol search algorithms have been proposed.
2. The algorithm leads to better Agri AOD recovery compared to previous studies.
3. AgriAod provides more detailed information about aerosol events than the daily multi-angle implementation of Atmospheric Correction (MAIAC) AOD.
In conclusion, this study introduces a pioneering hybrid aerosol search algorithm tailored to geostationary weather satellites. By integrating deep learning and transfer learning techniques, the algorithm overcomes the limitations of traditional physical algorithms and the scarcity of ground-based data sources. Furthermore, the applicability of algorithms to other multispectral sensors paves the way for a wider application in geoscience analysis.
Paper “FY4-AGRI's Deep Learning and Transfer Learning Hybrid Aerosol Search Algorithm: Development and Verification in Asia.” Duan, Xiangao Xia. Full Open Access Paper: https://doi.org/10.1016/j.eng.2023.09.023.
For more information about engineeringplease follow x (https://twitter.com/engineeringjrnl) & Facebook (https://www.facebook.com/engineeringjrnl).
