A web-based teaching platform built on one real flood case — Sentinel-1A SAR, 26 November 2025, Banda Aceh. Every figure is a live model output; every step from raw backscatter to flood mask is reproducible in the browser.
A teaching interface organized around workflow, not content piles. Every case walks a single path — from raw data, through method, to interpretation and discussion. Built for classroom demonstration, student practice, and CATCON competition review alike.
Before we look at model predictions, spend ninety seconds with the underlying physics. Radar does not see colour or brightness the way a camera does — it measures how much of its own microwave pulse bounces straight back.
Figure 1aVertical transmit, vertical receive. Most sensitive to smooth surfaces. Water = black, urban = bright.
Figure 1bVertical transmit, horizontal receive. Sensitive to volume scattering in vegetation.
Figure 1cA common false-colour composite. Green highlights vegetation volume scattering.
Three Sentinel-1A acquisitions of the same Aceh coast. Move left to right to follow the event: dry baseline, approaching weather, and co-event inundation.
Figure 2a Baseline. Dry paddy fields, full river discharge, normal coastal outline.
Figure 2b Before landfall. Early moisture accumulation in low-lying areas visible as darker patches.
Figure 2c Peak inundation captured during ascending pass. Large dark zones mark new standing water.
All three scenes are processed with identical radiometric terrain correction (SNAP). What changes is the physical world on the ground, not the sensor.
The model below was trained on KuroSiwo, a global flood dataset published in 2023. The Sentinel-1A acquisition you are about to inspect was recorded on 26 November 2025 — after the dataset’s cutoff. By signal, it is further out still: the tropical vegetation and saturated paddies of Sumatra push VH backscatter 5–8× brighter than the training mean. A radar world the model was never shown.
What we don’t hide. With the stock KuroSiwo
preprocessing, 71 % of this scene’s VH pixels
saturate against the default ceiling and the flood all but disappears.
With a single inference-time knob — clamp = 0.30
— the same weights recover it. Every figure that follows is that
recovery in action; every failure mode is one click away. The lesson
isn’t that the model works. It’s how, and
where, it would break.
Compare the CS-Mamba flood prediction under four input-scaling strategies, using the same Sentinel-1A scene over the Aceh coast. Each view below is a real pre-computed model output — click through the tabs and switch the configuration to see what changes.
Each figure below is a three-panel comparison: prediction A, prediction B, and the pixel-wise difference. Red marks pixels classified as flood by A but not B. Blue marks the reverse. The further two configurations are apart, the noisier the difference panel.
The 84.1% agreement number hides a story: most disagreement is blue (new flood that the training-time clamp missed entirely).
94.6% agreement. A conservative clamp captures the main flood footprint but shrinks the extent.
90.5% agreement. An aggressive clamp starts dropping real floods while over-estimating permanent water.
Input clamping bounds the SAR backscatter before normalization. A tight clamp (0.15) truncates bright returns and under-detects floods. A loose clamp (0.5) preserves bright reflectors but washes out the water signature. The sweet spot at 0.3 balances information loss against signal-to-noise — and more than doubles the recovered flood area compared to the training-time setting.
/ BEHIND THE CURTAIN See the full pipeline that produced these maps 6 stages · SNAP → tiles → stats → CS-Mamba → validation · including a scrubbable training trajectoryInstead of reading how it works, run it. Sample a random KuroSiwo tile and see the 7 bands. Drag the clamp slider and watch truncation live. Press RUN and a 224² patch rides a Cloudflare Tunnel to a WSL + RTX 5070 inference server. Click any cell of the agreement matrix and the real disagreement figure fades in.
Each KuroSiwo-format tile directory packs 6 GeoTIFFs: VV + VH at three acquisition times — pre-event 1 (21 Oct, baseline), pre-event 2 (2 Nov, approach) and co-event (26 Nov, main flood scene). Press Sample to load a random tile from the …-tile Banda Aceh test split. Each click round-trips to the WSL box in ~1 s.
Three Sentinel-1 VV composites — the same scene, ~11 days apart. Each tile the model sees is a 224 × 224 px crop of these, stacking all 6 bands (VV + VH at each of the three dates). About 911 such tiles make up the Banda Aceh test split. When the GPU endpoint is online, the Sample button above will sample one at random and show its 6 bands.
Every bar is a VH backscatter bucket. Everything to the right of your clamp value gets clipped to the ceiling — identical to the model. Find the clamp that keeps the flood tail visible without drowning in speckle. This one runs entirely in your browser.
Same model weights, same unseen 2025 tile — only the preprocessing clamp changes. Flip to LIVE to run on the GPU right now, or stay on CACHED to compare against the figure in the Notebook. Use NEXT TILE to move to a different patch of Banda Aceh. Nothing in the training set came from this scene.
prediction_report.json · full sceneNo ground truth exists for Banda Aceh on 2025-11-26, so we triangulate: measure the pixel-for-pixel agreement between every pair of configurations. Low numbers are not wrong — they're the teaching signal. Clicking a cell pulls up the real disagreement map.
The platform is built by a compact group combining remote sensing research, teaching practice, and web engineering — optimizing for reproducibility and classroom fit rather than scale.
IRIDeS, Tohoku University
Research interests in photogrammetry, remote sensing, computer vision and machine learning. Works on aerial/satellite image processing, DEM/DSM generation, stereo matching and change detection for urban monitoring and disaster science.
IRIDeS, Tohoku University
Ph.D. candidate at the International Research Institute of Disaster Science. Focus on AI-driven disaster analysis, flood mapping and depth estimation, integrating multi-source remote sensing with physical modelling.
OSCARS Lab, The University of Tokyo
Ph.D. candidate at the Graduate School of Information Science and Technology. Research covers adversarial robustness, deepfake detection, privacy-preserving ML and other AI-safety topics.
Center for Spatial Information Science, The University of Tokyo
Professor Emeritus of The University of Tokyo and Vice President of Reitaku University. Pioneering researcher in spatial information science, GIS, and geospatial data analysis with extensive contributions to human activity mapping and disaster applications.
IRIDeS, Tohoku University
Director of the International Research Institute of Disaster Science (IRIDeS). Professor specializing in disaster geo-informatics, remote sensing approaches for disaster impact identification, and tsunami damage detection.
Not just a contact form — this page is part of the platform's pedagogical loop. Students can ask questions, teachers can collect needs, partners can submit collaboration intent.
For concrete questions and methodological doubts students encounter in practice.
To request sample datasets, practice materials, or case support.
For course collaboration, case co-authoring, or thematic expansion discussions.
Submissions are sent to the backend /inquiries endpoint and stored
in the D1 database.