應用長短期記憶模式與多時期雷達影像於崩塌判釋

Global warming has been considered a critical factor to intensify the frequencyand magnitude of landslide hazards when they are generally triggered by heavyrainfall events. Nowadays, Synthetic Aperture Radar (SAR) sensors onboardEOS facilitate the detection of landslides over large geographical areas at a lowercost. The most common and simple approach to detect affected areas caused bylandslides is change detection which compares a pair of optical scenes acquiredpre- and post-event. Nonetheless, for SAR data, errors in the classification areexpected since not all changes in SAR signals are landslide-induced, such asdifferent satellite observation parameters, atmospheric conditions, terrain relief,and systematic changes of land cover. Therefore, by conducting a time seriesanalysis of historical data, this study assumes that it is possible to boost changedetection and improve its mapping accuracy. In this study, first, to reduce theatmospheric effect, the spatial and temporal coherence derived by SAR timeseries, after the atmospheric phase screen (APS) will be used to extract thedetection area. Second, a proposed algorithm that combines InSAR SensitivityIndex and multi-looking images will be used to minimize the layover and shadowareas caused by hilly topography in the SAR images. Finally, to incorporate thespatial and temporal signatures for the change detection, a modified deeplearning architecture, Convolutional Long Short-Term Memory (ConvLMST) isapplied for simulating the forecast image using SAR time series. All the above willbe incorporated into the proposed framework and this study expects to improvethe mapping accuracy and capability of landslide detection by applying this novelchange detection framework.