SinkAlert
Research only · not CAP
บทสรุปผู้บริหาร · Thai executive summary

SinkAlert ประเมินความเสี่ยงหลุมยุบเชิงพื้นที่ในกรุงเทพฯ–ปริมณฑลด้วย Random Forest วิจัย (phase_b_rf_v1_56) จากข้อมูลเปิด — ยังไม่ใช่ระบบปฏิบัติการ (operational_use=false) และไม่ออกแจ้งเตือน CAP อัตโนมัติ

แคตตาล็อกเหตุการณ์ที่ยืนยันแล้วถึง 54 เหตุการณ์ BMR / 49 จุดไม่ซ้ำ (Round 11; docs/117) แต่โมเดลบนเว็บยังตรึงที่ชุดเทรนเดิม 12 บวก / 60 ลบ — รีเทรนป้ายหนา + spatial-block CV (v1_89v1_96) ไม่ผ่านเกตคู่ จึงไม่โปรโมต: ซื่อสัตย์กว่าการอ้างตัวเลขสูงจากชุดเล็ก

ต้องการความร่วมมือข้อมูล: MEA / MWA / LiCSBAS (หรือ PSI ระดับเมือง) และพันธมิตร GISTDA (RS / InSAR / ข้อมูลเปิดเชิงลึก) เพื่อปลดล็อกเส้นทางปฏิบัติการ

English executive summary · for GIS / RS reviewers

SinkAlert ships a frozen open-data Random Forest susceptibility layer (phase_b_rf_v1_56) for BMR research screening only — operational_use=false, never auto-CAP. Cite OOF/LOPO metrics, not train accuracy.

Verified news catalog is now 54 BMR events / 49 unique points (<100 m co-located flagged; 2026-07-29 Round 11 · docs/117). The web default remains the 12-positive freeze: denser-label LOPO (v1_95 AP 0.648 / F1 0.591) and spatial-block CV (v1_96 k-means AP ~0.38 / AUC ~0.75) failed dual gates vs v1.56 (AP ~0.945 / F1 ~0.957). Holding the freeze is scientifically honest — more labels + geographic CV expose open-proxy limits, not a free accuracy upgrade.

Collaboration ask: MEA / MWA underground GIS, citywide LiCSBAS / numeric PSI, and a GISTDA data partnership (flood / land-use / InSAR path) to unlock an operational track.

Partner brief · GISTDA GIS / Remote Sensing

Phase B RF susceptibility — honest research status

What the open-data RF layer is, how it is validated, which rasters and vectors it uses, why denser labels did not promote a new web model, and what partner data unlocks next.

operational_use = false cap_auto_issue = false Frozen · phase_b_rf_v1_56 Catalog · 54 events / 49 unique v1_95–v1_96 dual-gate HELD Never auto-CAP

1. What SinkAlert is

Enterprise-oriented Bangkok Metropolitan Region (BMR) sinkhole / multi-hazard ops + citizen reporting platform (BDI Hackathon 2026 finalist). Combines citizen reports (web + LINE OA), a verified news-mined incident catalog, research susceptibility mapping, and a Contabo-hosted dashboard with an MHEWS research API layer.

Audience for this page: GISTDA GIS / remote-sensing specialists evaluating whether the science layer is honest enough to discuss data partnership — not a CAP product pitch.

2. Catalog vs freeze — why we still ship v1.56

54 / 49
Events / unique points
12 / 60
Freeze train set (web)
HELD
v1_89–v1_96 retrains
FieldValue
Web modelphase_b_rf_v1_56 — sklearn RF, 500 trees, max_depth=4, frozen 2026-07-29
Train labels (freeze)12 BMR positives / 60 negatives
Catalog now54 verified BMR events · 49 unique points (<100 m) + 2 national held-out (docs/106 · docs/117)
Latest denser LOPOv1_95 (54 events) AP 0.648 / F1 0.591 · v1_96 unique LOPO AP 0.593 / F1 0.542HELD
Spatial-block CVv1_96 k-means k=8 AP ~0.376 / F1 ~0.436 / AUC ~0.750 (diagnostic; not web default)
Statusoperational_use=false · never auto-CAP · no XGB/MLP demo promotion on this label regime
Scientific honesty: Dual promotion gates require OOF AP and best F1 not to regress vs the freeze. Adding verified positives increases spatial / temporal diversity (2019–2026 MEA shafts, MWA voids, drain scour). Retrains that look “worse” on LOPO are expected when the old 12-point set was easier — we do not swap the web default for a denser-label model that fails gates, and we do not invent operational accuracy KPIs.

3. Validation metrics (OOF on freeze — cite these)

~0.945
OOF Average Precision
~0.957
Best F1 (OOF)
~0.968
LOPO AUC (mean)

Leave-one-positive-out (LOPO): for each of the 12 freeze positives, train on the remaining positives + all 60 negatives; score the held-out positive vs negatives. At best F1 (~t=0.30): typically 11 TP / 1 FN / 0 FP.

Contrast: denser LOPO unique AP 0.593 / F1 0.542 and spatial k-means AP ~0.38 / AUC ~0.75 (docs/117) — geographic honesty check, not a web promotion candidate.

Honesty: Research OOF on a small freeze set — not citywide operational accuracy, not CAP readiness, not in-sample train scores.

4. Hard false negative — Sai Noi / MEA cable-pit

Sole hard FN at best F1 on the freeze: nbi-sainoi-2025-07-19 — Bang Kruai–Sai Noi Rd MEA cable manhole collapse (outer Nonthaburi), OOF p≈0.08. Open OSM power/manhole proxies do not replace MEA underground inventory. Recovering this FN without flooding FPs needs partner utility GIS.

5. Data stack in v1.56 (open features)

Copernicus DEM GLO-30

Elevation, slope, relief, TWI proxy, curvature (+ pos_copdem_curv)

BMA drains

District pipe count / length / mean dim / coverage aggregates

PSI (sparse)

Zenodo validation stations (~16 pts) IDW velocity / subsidence + local std — not citywide InSAR

Groundwater

DGR / Zenodo well density and static levels

OSM + transit

Roads, waterways, rail, tunnels, MRT construction / transit distances; dense-road log count

Google Open Buildings

Log building density 1–2 km + slope/relief × GOB interactions

Climate / flood / soil

Open-Meteo rain, amphoe flood index, clay IDW; HII × rain; WorldPop × PSI

Open enrichment path

Geofabrik / GHSL / GSW / HydroRIVERS / peri-urban trials through v1.88 — dual-gate failed; further open churn alone is not productive

Full feature list: web/pages/data/phase-b-rf-metadata.json · freeze: docs/100-phase-b-rf-v156-FREEZE.md

6. Collaboration ask — unlock operational path

Why partner data matters: Catalog already has 54 events / 49 unique points with spatial-block CV baselined — enough to expose open-proxy limits. Dual-gate holds show the next lift is features with underground / deformation truth, not more open churn. With MEA/MWA/LiCSBAS + GISTDA layers, a new RF (or successor) can be gated honestly toward operational_use — still never auto-CAP without written sign-off.

Shareable one-pager: Partner GIS ask → · Honesty boundary: What we do NOT claim · Duty desk: Duty-officer brief

7. Enterprise pilot use vs operational

8. Live demos (Contabo)

Prefer dashboard.cfoth.ai. sinkalert.com is often Cloudflare-stale vs Contabo — do not treat afolio as source of truth unless a fresh CF deploy is confirmed.

Repo docs: docs/100-phase-b-rf-v156-FREEZE.md · docs/106-label-harvest-bmr-2026-07-29.md · docs/007-phase-b-scientific-rf-spec.md · docs/000-AGENT-MAIN-REFERENCE.md

Evidence annex · endpoints we cite

No invented evidence — these are the public endpoints (and what they mean) for the “research-only / never auto-CAP” claim.

9. Adjacent track (not this mission)

Separately, SinkAlert has an OIC InsurTech Award 2026 prep page for Challenge #3 (disaster-related risk awareness for insurers). That is an adjacent InsurTech track — same research RF honesty rules, different audience and mission from this GISTDA GIS/RS brief. Do not mix CAP / insurance claims into the GISTDA data partnership ask.

OIC InsurTech pitch → · research RF only · not operational rating