Multi-Temporal Anomaly Detection for SAR Earth Observations
Principal Investigator (PI): Hook Hua, NASA's Jet Propulsion Laboratory
Synthetic aperture radar (SAR)-based geodetic imaging has revolutionized Earth science research in disciplines such as solid earth, ecosystems, and cryosphere. Yet the ability to effectively utilize SAR data for research, long-term monitoring of spatial areas of interests (AOIs), and rapid hazard response has been limited due to processing complexity, data volume sizes, and latencies in the end-to-end process. For example, barriers in urgent response include the lack of automated data triggers from forecasts, the need for specialized processing parameters that currently rely on expert intervention, and the manual delivery of actionable science data products to the decision support communities. Our team's AIST-2011 and AIST-2014 efforts towards an Advanced Rapid Imaging and Analysis (ARIA) data system has demonstrated the capability to automate high-volume SAR image analysis in a cloud computing environment, but a limiting factor has been the continued need for expert analysis for detection of features in the L2 data products as well as transients in the L3 time series data products.
Decision support products, which often are most useful if they are generated rapidly and with simplified information (e.g., "damaged or not damaged" or "flooded or not flooded"), often require change detection-based approaches utilizing the before and after event scenes to be processed - often with threshold values set on the underlying SAR measurement values of either amplitude or coherence. These steps requiring a human-in-the-loop have become a bottleneck for rapid and reliable exploitation of geodetic data for both long-term monitoring and event response with SAR data. If done correctly, SAR data can also be used to monitor for events such as volcano inflation or precursory signals to landslide events. Automating these time domain-based feature detection have even larger barriers due to processing of large temporal co-registered data stacks, processing complexity, as well as human expertise needed to assess the time domain signals. Machine learning approaches for earth science data have typically been applied on these types of single scene feature detection. In our AIST-2014, we prototyped automated classification of phase unwrapping features in processed L2 inteferograms from the Sentinel-1A/ B data streams. For change detection techniques however, most methods have focused on paired "before/after" observations. Now with more availability of low-latency and global multi-temporal remote sensing data, opportunities exists to exploit detection of more time dependent features of highly temporal earth science observations. Unlike the current state of the art pair-wise change detection techniques, multi-temporal spatial prediction techniques that leverage long-term historical observations yield more accurate and more interpretable predictions.
We will: (1) on-ramp onto the cloud computing platform, an automated and large-scale pre- processing of multi-temporal data stacks, and (2) develop automated large-scale machine learning analysis of multi-temporal transient detection and precursory signal analysis of multi-temporal stack datasets. We will demonstrate real science value by exploiting NASA's Sentinel-1A/B archive in preparation for the upcoming NASA-ISRO SAR (NISAR) mission. All input data will be acquired by discovering and accessing NASA's Earth Science Data and Information System (ESDIS) Distributed Active Archive Center (DAAC) data holdings via queries to the EOSDIS Common Metadata Repository (CMR). Our demonstration will pre-process co-registered stack of single looks complex data in the Amazon Web Services cloud. This is both computationally intensive as well as disk I/O intensive and therefore fits the need for cloud optimized pre-processing. The generic time-dependent anomaly detection approach will exercise example use cases such as automated landslide detection, automated volcanic uplift early detection, and/or automated detection of pre-event time-series patterns vs co-event flood detection.
Last Updated: Jun 13, 2019 at 12:04 PM EDT