TickSurv/RiskMap-SouthernQC-Pilot

Risk Map Inference Data — Southern Quebec

Geospatial input data for running tick habitat establishment risk predictions over seven administrative regions in Southern Quebec, covering Estrie (05), Montréal (06), Chaudière-Appalaches (12), Laval (13), Lanaudière (14), Montérégie (16), and Centre-du-Québec (17) — approximately 60,200 km².

Purpose

This dataset contains everything needed to run sliding-window inference with the TickSurv dual-stream production model and produce a continuous probability risk map at 250 m resolution.

Dataset Structure

├── tiles/
│   ├── apr/          # 130 Sentinel-2 composites — April 2024
│   │   ├── T0000_APR_2024.tif
│   │   └── ...
│   ├── jul/          # 130 Sentinel-2 composites — July 2024
│   │   ├── T0000_JUL_2024.tif
│   │   └── ...
│   ├── oct/          # 130 Sentinel-2 composites — October 2024
│   │   ├── T0000_OCT_2024.tif
│   │   └── ...
│   └── tile_manifest.csv
├── tabular/          # 16 environmental predictor rasters
│   ├── dd0c.tif
│   ├── bio5.tif
│   ├── ...
│   └── aspect.tif
├── boundaries/
│   └── study_area.geojson
└── tabular_scaler.json

Sentinel-2 Tiles (390 files, ~47 GB)

Seasonal median composites exported from Google Earth Engine for the year 2024.

Property	Value
Seasons	April (spring), July (summer), October (fall)
Tiles	130 non-overlapping tiles covering the study area
Tile size	27.4 km × 27.4 km (25 km core + 1.2 km margin)
Resolution	10 m/pixel
CRS	EPSG:32618 (UTM zone 18N)
Bands	7 — B2, B3, B4, B8, B11, B12, cloudprob_median
Data type	Float32
Values	Raw Sentinel-2 DN (0–10,000); divide by 10,000 for reflectance
Compositing	Pixel-wise median of cloud-masked images (cloud prob < 30%)

Band Order

Index	Band	Description	Wavelength
0	B2	Blue	490 nm
1	B3	Green	560 nm
2	B4	Red	665 nm
3	B8	NIR	842 nm
4	B11	SWIR-1	1610 nm
5	B12	SWIR-2	2190 nm
6	cloudprob_median	Cloud probability	—

Tabular Predictor Rasters (16 files, ~16 MB)

Environmental features covering the full study area, exported at 250 m resolution.

Feature	Source	Native Resolution
dd0c	DAYMET v4 — degree days below 0°C	1 km
bio5	DAYMET v4 — max temp of warmest month	1 km
bio6	DAYMET v4 — min temp of coldest month	1 km
bio12	DAYMET v4 — annual precipitation	1 km
bio13	DAYMET v4 — precip of wettest month	1 km
bio15	DAYMET v4 — precipitation seasonality	1 km
ph	SoilGrids v2 — soil pH (×10)	250 m
silt	SoilGrids v2 — silt fraction (g/kg)	250 m
clay	SoilGrids v2 — clay fraction (g/kg)	250 m
soc	SoilGrids v2 — soil organic carbon	250 m
bulk_density	SoilGrids v2 — bulk density	250 m
percent_broadleaf	ESA WorldCover 2021 — broadleaf %	10 m → 250 m
percent_urban	ESA WorldCover 2021 — urban %	10 m → 250 m
elevation	SRTM GL1 — elevation (m)	30 m
slope	SRTM GL1 — slope (degrees)	30 m
aspect	SRTM GL1 — aspect (degrees)	30 m

All rasters are Float32, EPSG:32618, with Deflate compression.

Study Area Boundary

GeoJSON polygon of the study area (union of seven Quebec administrative regions). Sourced from Données Québec administrative boundaries at 1:100,000 scale (CC-BY 4.0).

• Regions: Estrie (05), Montréal (06), Chaudière-Appalaches (12), Laval (13), Lanaudière (14), Montérégie (16), Centre-du-Québec (17)
• Area: ~60,200 km²
• CRS: EPSG:4326 (WGS84)
• Bounding box: 44.99°N–47.76°N, 74.89°W–69.63°W

Tile Manifest

tiles/tile_manifest.csv maps each export file to its tile ID, season, year, CRS, resolution, and geographic bounding box (in WGS84). 390 rows (130 tiles × 3 seasons).

Tabular Scaler

tabular_scaler.json contains pre-computed mean and standard deviation for each of the 16 predictor columns, fitted on the training set (n = 1,361). These statistics are used to z-score normalize tabular inputs at inference time.

Running Inference

# Clone the inference code
git clone https://github.com/<org>/lyme-disease-mapping.git
cd lyme-disease-mapping

# Download this dataset
huggingface-cli download TickSurv/RiskMap-SouthernQC-Pilot \
  --repo-type dataset --local-dir risk-map/data

# Run inference (GPU recommended)
python risk-map/scripts/run_inference.py --device cuda --batch-size 128

The script automatically downloads the production checkpoint from TickSurv/Ticksurv-Resnet-SSL-MOCO.

Outputs

File	Description
risk_map_2024.tif	Probability of tick establishment (0–1), 250 m resolution
attention_map_2024.tif	3-band seasonal attention weights (Apr, Jul, Oct)

Model

The inference pipeline uses the dual-stream TickSurv model:

• Image stream: ResNet-50 (MoCo v2 SSL pre-trained, frozen) with temporal attention pooling over 3 seasons
• Tabular stream: 2-layer MLP on 16 environmental predictors
• Fusion: Late concatenation → classification head → sigmoid probability
• Training: 5-fold spatial cross-validation, AUC = 0.821 ± 0.033
• Checkpoint: TickSurv/Ticksurv-Resnet-SSL-MOCO (r50_ssl_moco_defaults_production.ckpt)

Pilot Scope & Limitations

• Regional coverage — covers 7 of 17 Quebec administrative regions. Results do not generalize beyond the study area without additional validation.
• Single year — composites are from 2024; interannual variation is not captured.
• Cloud masking — seasonal medians may contain residual cloud/shadow artefacts, especially in April (snow/cloud confusion).
• Tabular resolution — climate variables are natively 1 km, bilinearly resampled to 250 m. Fine-scale variation in climate features is smoothed.
• Water bodies — tiles covering large water bodies (St. Lawrence) will have NaN regions; these are masked out by the study area boundary.

Citation

If you use this dataset, please cite:

@dataset{ticksurv_riskmap_southernqc_2026,
  title   = {Risk Map Inference Data — Southern Quebec},
  author  = {TickSurv},
  year    = {2026},
  url     = {https://huggingface.co/datasets/TickSurv/RiskMap-SouthernQC-Pilot},
  license = {CC-BY-4.0}
}