Undergraduate RS student
Goal: First contact with SAR; wants intuition plus one full worked example.
Path: Home → Primer → Flood Case showcase → inquiry form for questions.
/ ISPRS CATCON 9 · JUDGES’ PACKET
Remote Sensing Teaching and Practice Platform for Flood Analysis and Student Interaction
A web-based teaching environment built around a real Sentinel-1 SAR flood case — every figure is a real model output, every step is reproducible, and students close the loop with guided inquiry.
Teach remote sensing the way it is actually practiced.
Traditional remote sensing courses show finished maps and finished equations. This platform teaches the workflow — the messy middle where raw satellite archives become decisions. We pick one real disaster event, one operational SAR pipeline, and one clearly-explained segmentation model, and we walk learners through it from L1C GRD to interpreted flood map without hiding the steps.
The purpose is three-fold: (1) give undergraduates an intuitive first contact with SAR; (2) give graduate students a reproducible end-to-end pipeline they can rerun and modify; (3) give instructors a self-contained case they can drop into a one-hour lecture.
Teaching a workflow, tested on a scene the model has not seen.
The flood-segmentation model at the centre of this site was trained and validated on KuroSiwo (Bountos et al., 2023). The case shown throughout — a Sentinel-1A acquisition of the 2025-11-26 Banda Aceh flood — post-dates the dataset and is not part of it. What learners watch is generalisation under a single preprocessing hyperparameter, in a radar regime that runs 5–8× brighter than the training mean.
We frame this as a qualitative, out-of-distribution evaluation, not a test-set accuracy claim. The table below condenses the facts that make it one; the Claims & Limits block that follows spells out what the demonstration does and does not support.
The model was trained on KuroSiwo. The scene shown throughout this site was acquired afterwards and is out of distribution. What you see below is generalisation under a single preprocessing knob — a teachable observation, not a test-set accuracy claim.
- Training data
- KuroSiwo (Bountos et al., 2023) — a global, multi-temporal flood dataset of 33 B m² of labelled water.
- Evaluation scene
- Banda Aceh · 2025-11-26 — acquired two years after the KuroSiwo cutoff and not in the dataset.
- Distribution shift
- VH backscatter runs 5 – 8× brighter than the KuroSiwo training mean (tropical vegetation, saturated paddies).
- Adaptation
- A single inference-time hyperparameter (input clamp). No retraining, no fine-tuning, same weights.
- Outcome
- With clamp = 0.30 the model recovers a sensible flood map — flood / water / land ≈ 4.2 / 5.1 / 90.7 %.
- Failure modes shown
- clamp = ∞ → noisy; clamp = 0.15 → 71 % of VH pixels saturate and the flood disappears.
What we claim
- The 2025-11 Banda Aceh scene is correctly out-of-distribution for this checkpoint.
- A single inference-time hyperparameter (input clamp) recovers a sensible flood map — no retraining.
- The interactive panels show real, reproducible model behaviour on an unseen scene.
What we don’t claim
- That
clamp = 0.30generalises to every future OOD scene — it happens to be the right knob here. - That this pipeline outperforms any state-of-the-art flood-mapping benchmark.
- That qualitative OOD evaluation replaces quantitative held-out metrics for operational use.
How you could falsify us
- Open the Atlas, pick any of the 49 tiles, press Run in LIVE mode — and see where the recovery breaks.
- Re-run the pipeline with
reproduce-no-saron a different Sentinel-1 scene.
Six functions, one teaching loop.
SAR physics primer
Ninety-second explainer pairing VV/VH/RGB imagery with backscatter intuition for first-time learners.
End-to-end flood case
Banda Aceh 2025-11-26 case from raw Sentinel-1 GRD through Cross-Scale Mamba (CS-Mamba) prediction to side-by-side comparison.
Walkable 6-stage pipeline
Scrollable `/pipeline` page with real commands, artefacts, TL;DR and check-yourself prompts per stage.
Explorable-explanation notebook
`/paper` presents the model as five interactive widgets — tile explorer, clamp playground, architecture walkthrough, 37-epoch training scrubber, model race. Every teaching point is a control the learner operates, not prose.
Student inquiry loop
Feedback form wired to Cloudflare Workers + D1 so classroom questions and data requests reach the teaching team.
Reproducible artefacts
Every image on the site is a real output; pipeline reruns with one command on the provided dataset.
Name, affiliation, country, email — as required by CATCON.
Primary contact: Wei Yuan (wei.yuan@tohoku.ac.jp). All five members are confirmed and available for Congress correspondence.
RoleNameAffiliationCountryEmail
Principal Investigator · Team Lead Wei Yuan International Research Institute of Disaster Science (IRIDeS), Tohoku University Japan wei.yuan@tohoku.ac.jp
Ph.D. Candidate · Model & Data Engineering Zhongyuan Yang International Research Institute of Disaster Science (IRIDeS), Tohoku University Japan yang.zhongyuan.t2@dc.tohoku.ac.jp
Ph.D. Candidate · AI Methodology Weihang Ran OSCARS Lab, Graduate School of Information Science and Technology, The University of Tokyo Japan ran-weihang@g.ecc.u-tokyo.ac.jp
Professor Emeritus · Funding Support and Advice Ryosuke Shibasaki Center for Spatial Information Science, The University of Tokyo Japan rshibasa@reitaku-u.ac.jp
Director · Funding Support and Advice Shunichi Koshimura International Research Institute of Disaster Science (IRIDeS), Tohoku University Japan koshimura@tohoku.ac.jp
Browser is the whole client. The rest is optional.
For judges and learners (viewing)
- Browser
- Chrome 110+, Firefox 110+, Safari 16+, Edge 110+
- Screen
- Works from 360 px (mobile) to 4K desktop
- Network
- Any broadband; total page weight under ~10 MB
- JavaScript
- Required (ES2022)
- Account
- None — fully public, no sign-in
For reproducers (rebuilding)
- Frontend
- Nuxt 3 · Vue 3 · TailwindCSS · @nuxt/image
- Backend
- Cloudflare Workers · D1 SQLite
- Modelling
- Python 3.10 · PyTorch 2.x · SNAP 9 · rasterio
- Build
- Node 18+, pnpm or npm
- Deployment
- Cloudflare Pages (static) + Workers (API)
Why this is not just another course website.
- 01 The segmentation model itself is original research: Cross-Scale Mamba (CS-Mamba), our SAR-specific extension of RSMamba (Chen et al., 2024), submitted to the ISPRS 2026 Congress — the qualification on which this CATCON entry stands.
- 02 Built around a single real disaster case rather than a catalogue of toy datasets.
- 03 Parameter sensitivity is taught as a first-class lesson: the same model, four input-clamp configs, four visibly different flood maps.
- 04 Every figure on the site is a real model artefact — no stock images, no staged screenshots.
- 05 The teaching loop closes with an inquiry form backed by Cloudflare Workers + D1, not a mailto link.
- 06 Pipeline page exposes real shell commands and file layouts so students can rebuild the case at home.
Three personas, three entry paths through the platform.
Graduate / MSc student
Goal: Wants to reproduce the pipeline and understand parameter sensitivity.
Path: Pipeline page → Training replay → clone toolkit → run `reproduce-no-sar`.
Instructor preparing a lecture
Goal: Needs a self-contained case they can project and discuss in one session.
Path: Flood Case showcase → teaching-flow prompts → discussion questions.
Six screenshots and a one-page PDF, for judges who skim.
Six frame grabs from the live deploy plus a one-page PDF summary — all SHIPPED. The live site is itself the walkthrough.
Screenshots (3–6)
SHIPPED
One-page PDF summary
SHIPPEDcatcon9-summary.pdf Source:
competition/isprs-catcon-2026/pdf-summary-source.md