Deep Learning based Cloud Retrieval Techniques Using Multiple Satellite Remote Sensing Data

Abstract

Clouds are a critical component of the EarthÕs climate, impacting EarthÕs energy, hydrological and biological cycles. Satellite-based remote sensing is an important instrument to monitor cloud properties on a regional to global scale. Active and passive satellite sensors have been designed to observe and retrieve cloud properties. For cloud remote sensing, the advantages of active sensors include their capability of resolving the vertical distribution of the cloud layer and better performance during nighttime and polar region in comparison with passive sensors. On the other hand, passive sensors generally have up to three or four order-of-magnitude wider swaths and thereby substantially better spatial coverage. Challenges hinder the attempt of conducting a comprehensive analysis of multiple satellite remote sensing data. Firstly, the two remote sensor datasets collected by active and passive sensors respectively are heterogeneous, e.g., different feature spaces and dimensionalities. Moreover, distribution drift can happen from on-track collocated data to off-track data, due to environment change or different surface types where the data were collected. Lastly, a large number of passive remote sensing data are unlabeled and carry rich information due to their global coverage. In this dissertation, I first introduce an end to a deep domain adaptation method with heterogeneous domain mapping and correlation alignment (DAMA) to employ both active and passive sensing data in cloud type detection. It can learn domain invariant representation from source and target domains by transferring knowledge across these domains. DAMA-WL extends the DAMA by incorporating noisy/weak label supervision. Secondly, I develop VDAM, a deep domain adaptation model based on variational auto encoder (VAE) and Convolutional Neural Networks (CNN). It exploits the characteristics of a satellite orbiting track and applies 1D-CNN to capture spatial correlations. It also develops a VAE based generative domain adaptation method to learn the latent representations and introduces domain alignment methods on both feature space and label space to develop a powerful domain adaptation technique. Thirdly, I introduced DRLO, a self-supervised representation learning model using pre-training and fine-tuning VAE models to leverage the vast amounts of unlabeled off-track data and incorporate the distinctive aspects of off-track data. The developed model can capture the unique attributes of off-track data through pre-training and fine-tuning strategies. Quantitative evaluations show our deep learning based cloud detection approaches can achieve higher accuracy in predicting cloud types in the challenging passive satellite remote sensing data. I also perform a climatology analysis to demonstrate the effectiveness and plausibility of our methods in cloud property prediction.