Mandi, Himachal Pradesh
Retrospective candidate-ranking study
Start with an existing inventory. Then ask whether terrain and imagery can bring more labelled slides to the front of a limited review queue.
Explore the review budgetArchive evidence. Candidate priorities.
No early-warning or district-wide detection claim.
The review universe is fixed. The core cohort contains 2,120 candidates, including 1,181 labelled slides. Recall here concerns those labelled slides inside this cohort.
Tanvi Chauhan / Prof. K. V. Uday’s group supplied the dataset and upstream rule-based inventory. Tejesh’s study ranks candidates already generated by that inventory.
The review-budget explorer
Choose how much of the cohort to review. Compare the number of labelled slides each saved model brings into that first portion of the queue.
“Pre” is a nominal composite, not a verified pre-event observation. See timing limits.
Why this default? Pre + terrain has the best-supported primary contrast against Terrain. It is not a universal winner: at 20%, Pre alone retrieves 330 labelled slides versus 327 for Pre + terrain.
What changes at 100%? Every model reaches the same 1,181 labelled slides. Ordering changes what appears early, not the contents of the cohort.
Each model reviews exactly 424 candidates. This static evidence remains readable when the explorer is unavailable.
Download 80 aggregate summariesCSV contains only model/budget counts and percentages. No candidate records.
| Model | Slides | AP |
|---|---|---|
| Terrain | 307 | 0.6946 |
| Pre alone | 330 | 0.7277 |
| Pre + terrain Default | 327 | 0.7336 |
| Pre + Post + terrain | 323 | 0.7360 |
AP = average precision across ranking thresholds. AP is not accuracy. The default’s 20% precision is 77.1%; its candidate recall is 27.7%.
Reading the evidence
A spatial evaluation of saved candidate rankings. The original scanner and this study do different jobs.
The upstream workflow used Sentinel-2 imagery, cloud/snow/water masking, temporal vegetation averages, vegetation-change conditions and slope screening, followed by manual imagery checks. The original scanner was rule-based, not machine learning.
No comparable AP was supplied for that scanner. Terrain AP 0.6946 belongs to this study’s terrain-only comparison model; it is not the original detector’s score.
The core comparison uses five spatial group folds with 5 km groups and strict >1 km train/test separation. The review explorer uses saved held-out predictions for the 2,120-candidate cohort.
The suite contains 52 arms/configurations, 240 saved fold models and 105,942 held-out rows. No external research data or pretrained weights were used.
“Pre” names a nominal composite from to . Event timing has not been verified. Pre and Post feature labels do not establish a historically valid before/after test.
The boundary of the result
This is exploratory, archive-only retrospective ranking. Treat the results as a way to examine candidate priorities.
No district-wide detection or recall, superiority over the original rule system, early warning, rainfall result, 2025 transfer or continual temporal learning result. There is no historically untouched holdout.
Margin acquisition helps two initializations and loses one. It does not establish an uncertainty detector or expert time saving.
Dated events and scenes, reliable IDs for unvalidated cases, rainfall, independent regional negatives and 2025 data.
307 missing-chip exclusions belong to the core-to-image cohort difference. They are not the approximately 300 reportedly unvalidated cases, whose case IDs are unavailable.
This page publishes 80 saved model/budget summaries and approved aggregate evidence. Original datasets, candidate locations, imagery and model files are not published.