Deep Multi-Sensor Domain Adaptation on Active and Passive Satellite Remote Sensing Data

Abstract:

Studies have shown machine learning (ML) algorithms such as Random Forests could outperform the physical-based algorithms in remote sensing applications. However, these ML algorithms are not well-suited to learn from heterogeneous sources such as multiple active and passive sensors. For example, Random Forest can be either developed for Cloud-Aerosol Lidar and Infrared Pathfinder Satellite Observations (CALIPSO) or Visible Infrared Imaging Radiometer Suite (VIIRS) sensor data, but it cannot jointly learn from both these sensors since there is a mismatch of features (variables) among the sensors. On the other hand, domain adaptation techniques have been developed to handle data from multiple sources or domains. But most existing domain adaptation approaches assume that the source and target domains are homogeneous i.e., they have the same feature space, and the difference between domains primarily arises due to the data distribution drifting. Nevertheless, many real world applications often deal with data from heterogeneous domains that come from completely different feature spaces. For example, in our remote sensing application, the source domain, namely CALIPSO, contains data of 25 attributes collected by the active spaceborne Lidar sensor; and the target domain, namely VIIRS, contains another group of data of 20 attributes collected by passive spectroradiometer sensor. CALIPSO has better representation capability and sensitivity to aerosol types and cloud phase, while VIIRS has wide swaths and better spatial coverage but has inherent weakness in differentiating atmospheric objects on different vertical levels. To address this mismatch of features across the domains (sensors), we propose a novel deep learning based heterogeneous domain adaptation framework called Deep Multi-Sensor Domain Adaptation (DMSDA) to 1) learn the domain invariant representations from source CALIPSO and target VIIRS domains by transferring the knowledge across these domains, and 2) better classify the different cloud phase types in the source and target domains. Our experiments on a collocated CALIPSO and VIIRS sensor dataset showed that DMSDA can achieve 69% classification accuracy in predicting the cloud phase types that is at least 23% improvement and outperformed other ML approaches employed in comparison.