BEGIN:VCALENDAR VERSION:2.0 PRODID:-//pretalx//talks.osgeo.org//foss4g-2026//speaker//88JR7P BEGIN:VTIMEZONE TZID:JST BEGIN:STANDARD DTSTART:20000101T000000 RRULE:FREQ=YEARLY;BYMONTH=1 TZNAME:JST TZOFFSETFROM:+0900 TZOFFSETTO:+0900 END:STANDARD END:VTIMEZONE BEGIN:VEVENT UID:pretalx-foss4g-2026-NXBLNJ@talks.osgeo.org DTSTART;TZID=JST:20260903T133000 DTEND;TZID=JST:20260903T140000 DESCRIPTION:Floods are becoming more frequent and severe worldwide due to c limate change. These disasters cause significant human and economic losses \, making the development of rapid and highly accurate flood mapping techn ologies essential for clearly identifying damage conditions and supporting subsequent rescue activities.\n\nRemote sensing with openly available Syn thetic Aperture Radar (SAR) plays an essential role in disaster monitoring . As an active sensor that transmits its own microwave signals\, SAR can a cquire observations regardless of weather or time of day\, enabling reliab le monitoring even in extreme conditions.\n\nHowever\, conventional SAR-ba sed flood detection methods face several challenges. Threshold-based appro aches applied to a single SAR image struggle to distinguish permanent wate r bodies from newly inundated areas. To address this\, methods that compar e pre-flood and flood-time images have been developed. Yet openly availabl e data such as Sentinel-1 have satellite revisit cycles that create tempor al gaps of several days to weeks between compared images. During this inte rval\, seasonal vegetation changes and speckle noise differences are frequ ently misidentified as flood-induced changes\, leading to increased false positives. To address these false positives caused by temporal misalignmen t\, this study proposes a flood detection method that synchronizes the tem poral axes of the compared images.\n\nIn our experiment\, we used the open ly available "UrbanSARFloods" dataset. This dataset was created for global -scale flood detection and covers 18 flood events from five continents\, i ncluding flood and non-flood label data. Each image chip has a spatial res olution of 10 m and a size of 512 × 512 pixels\, and data from a selected 628 area of interest were used in this study.\n\nWe used Sentinel-1 Groun d Range Detected (GRD) data as our primary flood detection source. Sentine l-1 GRD is a multi-looked SAR product projected to ground range using an E arth ellipsoid model. In Interferometric Wide (IW) mode\, it provides roug hly 10-meter resolution with dual-polarization (VV and VH) across a 250 km swath. Sentinel-1 acquires observations regardless of weather or time of day\, with a 6–12 day revisit cycle. We collected SAR time series from G oogle Earth Engine (GEE)\, covering one year prior to each flood event up to 40 images per location. These data include two polarization modes: VV a nd VH. To ensure consistency\, only data acquired from the same orbit (asc ending or descending) as the flood-time observation were used.\n\nGRD data contain inherent speckle noise that requires filtering. To improve data q uality\, we reduced speckle noise using a median filter (3 × 3) and perfo rmed normalization using Min–Max scaling. For datasets with missing time -series images\, we interpolated missing data using three-dimensional spli ne interpolation.\n\nThe revisit gap can cause temporal misalignment betwe en compared images. To obtain an image representing non-flood conditions\, we simulated imagery from past SAR GRD sequences at the timing of the flo od observation. A three-layer ConvLSTM model captured spatio-temporal depe ndencies and generated predicted "non-flood" SAR images. The model used a hybrid Mean Absolute Error (MAE)-Structural Similarity (SSIM) loss functio n to preserve edge details and suppress speckle noise. This function prior itizes structural and statistical similarity over standard MSE.\n\nWe crea ted multiple image chip pairs by using six consecutive time steps as one s et and sliding the window forward by one time step. Each image chip pair c onsists of five continuous input images and one subsequent image as refere nce. The model training was performed by minimizing the error between the predicted image generated from the five input images and the reference ima ge. A dedicated ConvLSTM model was built for each location to learn its un ique surface characteristics. After training\, we input the five most rece nt image chips into the model to generate a predicted image for the next t ime step.\n\nFinally\, we performed flood detection using a Siamese Networ k to compare the simulated images with the observed images acquired during flooding. This CNN-based model uses an encoder–decoder architecture wit h skip connections\, extracting high-dimensional features while preserving spatial boundary information. After computing pixel-wise differences betw een the feature maps\, we applied a sigmoid function to produce a flood pr obability map. The final flooded areas were detected through thresholding (0.5).\n\nQuantitative evaluation using a pre-split test dataset consistin g of 30 samples demonstrated considerable performance improvements. The pr oposed method achieved an F1 score of 0.605 (Precision: 0.556\, Recall: 0. 661)\, representing a 42% improvement over the conventional pre-/post-floo d comparison method\, which achieved only 0.427 (Precision: 0.328\, Recall : 0.610). This improvement was primarily attributed to the substantial red uction in false positives caused by temporal misalignment.\n\nThe precisio n improvement from 0.328 to 0.556 indicates that the proposed temporal syn chronization approach effectively suppresses misdetections caused by seaso nal vegetation changes and speckle noise variations. Meanwhile\, the recal l score of 0.661 demonstrates the method's capability to detect actual flo od events without sacrificing sensitivity.\n\nQualitative evaluation furth er confirmed the effectiveness of the proposed approach. Visual inspection revealed that the method successfully suppresses speckle-like noise patte rns that frequently lead to false positives in conventional methods. The a pproach accurately detected both large-scale inundated areas\, such as flo oded river plains and urban zones\, as well as small\, isolated flooded re gions that are typically difficult to identify. The generated flood probab ility maps showed clear spatial boundaries between flooded and non-flooded areas\, enabling more reliable damage assessment.\n\nWhile we demonstrate d the usefulness of the flood detection method\, several challenges were i dentified. In regions with complex topography\, frequent misdetections occ urred due to terrain-induced radar shadow\, geometric distortions (layover )\, and changes in vegetation's dielectric properties. In urban areas\, fl ood detection often misses actual flooding because double-bounce scatterin g reduces the backscatter intensity differences between images. Future wor k will consider incorporating Digital Elevation Model (DEM) data and coher ence data\, which quantify the phase correlation of microwave signals betw een two temporal images\, to account for physical terrain constraints and further suppress misdetections in urban areas\, enabling more effective fl ood detection.\n\nIn conclusion\, this study proposed a flood detection me thod that eliminates temporal lag between compared datasets and demonstrat ed improved performance in detecting flood-induced changes. By enabling ne ar-real-time flood detection under extreme weather conditions\, this metho d is expected to strengthen societal response capabilities to flood disast ers. DTSTAMP:20260717T225729Z LOCATION:Cosmos2 SUMMARY:Eliminating Temporal Misalignment in SAR Flood Detection with a Con vLSTM-Siamese Approach Using Sentinel-1 Time Series - Narumasa Tsutsumida\ , 中嶋達也 URL:https://talks.osgeo.org/foss4g-2026/talk/NXBLNJ/ END:VEVENT END:VCALENDAR