Deep Learning-Based Aerosol and Ocean Data Retrieval from Satellite Polarimeter Measurements

This paper presents a deep learning (DL)-based approach for retrieving aerosol and ocean optical parameters from polarimeter measurements. The traditional data retrieval method, i.e., Microphysical Aerosol Properties from Polarimetry (MAPP) algorithm, involves vector radiative transfer (VRT) calculations, which are a time-consuming and computationally intensive process. To address this limitation, we propose replacing the VRT calculations in MAPP with DL models to accelerate the data retrieval process. The core idea is to train DL models to replicate VRT calculations used in MAPP. To achieve this, we collected 2 million input-output pairs from the VRT calculations in MAPP to construct a comprehensive training dataset. Three types of DL models, including feedforward neural networks (FNNs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs), were developed to learn the underlying VRT calculation pattern through supervised learning. The performance of these models was evaluated using a test dataset, with the RNN model achieving the highest prediction accuracy. Experimental results indicate that the proposed DL-based approach can improve the data retrieval efficiency of MAPP algorithm while maintaining high accuracy in the data retrieval process.