RTS Segmentation Model v2

Semantic segmentation of Retrogressive Thaw Slumps (RTS) in Arctic satellite imagery for pan-arctic mapping.

Overview

This project trains a deep learning model to detect RTS from PlanetScope basemap imagery (up to 74N) and deploys it for pan-arctic inference to produce an RTS survey map.

Status (2026-07): model v2 (3-seed UNet++/EfficientNet-B5 ensemble, RGB+NDVI, thr 0.65) is deployed — the pan-Arctic South run (2025 Q3 imagery, ≈50–76°N, 41.57M tiles) is complete and its products are shipped (see Deployed products). Model v2.1 (DINOv3-L MAE self-supervised pretraining) is in progress on branch v2.1-pretraining.

This README is the map of the repo — every canonical document is linked below, and the Source of truth table says where each kind of fact lives. Specs are the source of truth: always read the relevant doc before implementing (see CLAUDE.md §Rule 1).

Data

• Training: 2024 PlanetScope Quarterly Basemap (RGB 3m)
• Inference: 2025 PlanetScope Quarterly Basemap
• Labels: Refined from ARTS dataset on 2024 imagery (~2–3k positive, ~20–25k negative tiles)
• Auxiliary (optional): Sentinel-2 NDVI/NIR, ArcticDEM derivatives

Deployed products

The 2025 Q3 pan-Arctic South run produced a tiered product family in gs://rts-mapping-v2-usw1/inference/2025q3_south/products/ — catalog SSoT: post-inference/south_products.md.

• High-confidence map: south_rts_high_confidence.gpkg — 19,068 polygons / 529.7 km², QC-calibrated rts_class from a measured precision grid (279 human ratings).
• Full candidate inventory (MMU≈0): 60,167 polygons / 688.2 km² across confidence tiers; original delivered south_rts.gpkg (10,984 / 238.1 km² at thr 0.65) kept for provenance.
• Rasters: probability mosaic (1,633 COG shards + WMTS z10 re-tile), 95 m browse likelihood surface, 10 km / 0.5° hotspot density grids (threshold-free expected RTS area 1,037.4 km²).
• Open in ArcGIS Pro: post-inference/arcgis_south_products.md.

Document map

Project & process

Document	Purpose
CLAUDE.md	How to work in this repo: rules, structure, technical constraints, code style, the 3-doc update ritual
docs/experiment_ledger.md	Experiments SSoT — every run, the locked recipe, per-family findings, dropped ideas (scores auto-harvested)
current_working_status.md	Project diary — rolling progress (just-completed · now · future); links to the ledger for numbers
docs/report.html	Generated analytical + visual report (build with scripts/build_report.py)

Data

Document	Purpose
data/data.md	Data pipeline spec — sources, labels, splits, normalization, disk layout (§9 = EXTRA bands)
data/data_format.md	Format standards for all data (CRS, tile size, label values, dtypes)
data/datacheck.md	Data-validation checks at each lifecycle stage

Training

Document	Purpose
training/training.md	Model, loss, metrics, training loop, train–inference consistency contract
training/experiments.md	The phased experimentation plan (sequential elimination + multi-seed lock)
docs/baseline_unetpp_effb5.md	Living experiment record for the UNet++/EfficientNet-B5 baseline
docs/optimization_roadmap.md	Cross-aspect optimization roadmap (training/inference/infra) + experiment-fairness & validity audit

Inference & post-inference

Document	Purpose
inference/inference.md	Deployment workflow — tiling, overlap aggregation, merging, vectorization
post-inference/post-inference.md	Post-processing, QC, evaluation, threshold tuning (spec complete; multi-scale fusion deferred)
post-inference/south_products.md	Product-catalog SSoT — every shipped South product: provenance, decode, tier table, caveats
post-inference/arcgis_south_products.md	How to download and open the South products in ArcGIS Pro
deliverables/README.md	ADC/PDG handover doc — submission manifest, WMTS tiling convention, methods, attribute dictionary

Computing

Document	Purpose
computing/infrastructure.md	Infra SSoT — GCP projects, buckets, VM inventory, regions, compute budget, data storage map
computing/vm_instruction.md	Daily VM/SSH how-to — start/stop, config, Python env, file transfer
computing/docker_training.md	Docker build/run how-to — image, mounts, GCS auth
computing/artifact_inventory.md	Artifact → bucket/path map — where every durable artifact lives, SSoT vs backup

Domain

Document	Purpose
domain/inference_domain.md	Inference domain and circumpolar subregions (H. Rodenhizer)
domain/training_data_distribution.md	Geographic/ecological distribution of the training data (H. Rodenhizer)

Tests

Document	Purpose
tests/tests.md	Test-suite living doc — per-test inventory, strictness, coverage gaps

Source of truth

This repo follows a single-source-of-truth standard. Where each kind of fact lives:

Concern	Source of truth
Config values — hyperparameters, paths, thresholds	configs/*.yaml (configs/base_v2_fast.yaml is the canonical base for current experiments; configs/baseline.yaml records the original Phase-0/1 baseline + shared infra keys)
MLflow tracking URI	configs/baseline.yaml:mlflow.tracking_uri
Core constants — CRS, tile size, label values, seed	CLAUDE.md §Technical Constraints
Data disk layout & EXTRA bands	data/data.md (§9 for bands)
Status, roadmap, project decisions	current_working_status.md
Experiments, scores, recipe, findings	docs/experiment_ledger.md
Test inventory	tests/tests.md
Infra facts — projects, buckets, VMs, regions, budget	computing/infrastructure.md
Shipped South products — files, numbers, caveats	post-inference/south_products.md
Artifact locations — what lives in which bucket	computing/artifact_inventory.md

Todos

1. training in multi-scale DONE (ledger family M, 2026-07-02): 0.5× re-stage + joint dual-scale training, 3 seeds — gates 1+2 pass, gate 3 (fusion recall) fail; inference multiscale path implemented, deploy stays scales:[1.0].
2. pan-Arctic South inference + products DONE (2026-07): full run + tiered QC-calibrated products shipped (see Deployed products).
3. v2.1 — DINOv3-L MAE self-supervised pretraining on the 295k-tile South corpus (in progress, branch v2.1-pretraining); v3 hard-negative mining seeded from qc_false_hard_negatives.gpkg (152 QC-verified FPs).
4. explore GEE satellite embedding as input feature
5. 2025 micro set to test temporal domain shift