Data Sources & Methodology
Hunt IQ combines public geospatial data (LiDAR, satellite imagery, water, parcels, weather) with published GPS collar research to produce stand recommendations. This page documents every dataset we load, every scoring module we run, and every research source those modules rely on. Every DOI on this page points to a verifiable paper or public dataset.
The registry is updated whenever a new feature, layer, or algorithm lands in the codebase. If you find a number in the app that doesn't trace to a source on this page, please file a bug.
Citations
26
Verified
21
Likely
5
Scoring modules
10
Registry version 1.0 · last updated 2026-04-06
What Hunt IQ is, and what this page is for.
Hunt IQ is a map-first tool that recommends stand locations by combining real public geospatial data (terrain, land cover, water, public-land boundaries, weather forecasts) with real published research on white-tailed deer movement, habitat selection, and hunting-pressure response. The goal is to replace “I think this spot looks good” intuition with evidence the user can verify.
This page is the public methodology document. It exists because a biologist or detail-oriented client should be able to read our sources without having to browse the source code. When a new feature lands in Hunt IQ, the corresponding section here gets updated in the same pull request.
Three principles shape everything on this page:
Three neutral categories describe what a data source IS. All three are legitimate; the badge just tells you which category each number comes from.
Observations of the physical world. USGS 3DEP LiDAR elevation, NLCD land cover pixels, NHD water polygons, PAD-US parcel boundaries, Open-Meteo HRRR weather forecasts, iNaturalist citizen-science sightings, user-reported deer sightings, and GPS collar fixes from published telemetry studies. If the underlying data is a measurement of something that exists in the physical world, it gets this badge.
Published peer-reviewed coefficients and models. The RSF beta coefficients from Darlington 2022, the HMM activity budgets from MSU P3927, the Auburn pressure decay curves, the Oklahoma hunt-day effect from Marantz 2016 — all real published science with real DOIs. The model APPLIES those numbers to local covariates. This is rigorous science; it just isn't a live measurement of your specific hunt area.
Demo data generated in-house for testing. The collar_sample.geojson file is the only synthetic data we ship, and it's explicitly opt-in via the --synthetic flag on load_collar_data.sh. Clearly labeled in every UI surface that displays it so you can never confuse it with an observation. Useful for running the pipeline without requiring network access or external datasets.
10 public datasets loaded or ingested by the Hunt IQ pipeline. Every one is a real observation of the physical world — satellite imagery, LiDAR, GPS telemetry, or field surveys — with a verified DOI or stable URL.
Open-Meteo / NOAA HRRR · 2026
Findings used in Hunt IQ
Referenced from
U.S. Geological Survey · 2024
Findings used in Hunt IQ
Referenced from
U.S. Geological Survey · 2024
Findings used in Hunt IQ
Referenced from
U.S. Geological Survey Gap Analysis Project · 2024
Findings used in Hunt IQ
Referenced from
VanAcker, M., DeNicola, V., DeNicola, A., Aucoin, S.G., Simon, R., Toal, K., Diuk-Wasser, M., Cagnacci, F. · 2023
Findings used in Hunt IQ
Referenced from
Note
DOI verified live against Dryad landing page on 2026-04-06. 89.54 MB. This is the dataset that replaces synthetic collar_sample.geojson as the default load target per STORY-055. ALL DATA MUST BE REAL — this is real NYC DIMP GPS telemetry, not synthesized.
Kauffman, M., Lowrey, B., Berg, J., Bergen, S., Brimeyer, D., Burke, P., Cufaude, T., Cain, J.W., III, Cole, J., Courtemanch, A., Cowardin, M., Cunningham, J., DeVivo, M., et al. · 2022
Findings used in Hunt IQ
Referenced from
Note
Verified live on 2026-04-06 — the parent ScienceBase collection (63598d30d34ebe442503eb3f) hosts UMRv3_Complete_dataset.zip and inspection confirmed all 20 expected white-tailed deer files (4 layers × 5 shapefile components: shp/shx/dbf/prj/xml). White-tailed deer rarely migrate long distances, so this Washington Selkirk subset is the only species-matched corridor dataset available; it is appropriate as a validation reference for our computed corridors but is geographically limited to the WA Selkirk population. STORY-059 / EPIC-010.
Movebank Data Repository contributors (related publication: see DOI) · 2022
Findings used in Hunt IQ
Referenced from
Note
Discovered via the Movebank Data Repository DSpace REST search at datarepository.movebank.org/server/api/discover/search/objects?query=Odocoileus. The standard Movebank `direct-read` API requires authentication for anonymous queries (returns 401); the data repository's REST API is fully open. Verified live on 2026-04-06 — bitstream content URLs return 200 and the CSV header parses cleanly with the standard Movebank schema. STORY-057 / EPIC-010.
Multi-Resolution Land Characteristics (MRLC) Consortium · 2021
Findings used in Hunt IQ
Referenced from
Jacobsen, T., Wiskirchen, K., Ditchkoff, S. · 2020
Findings used in Hunt IQ
Referenced from
Note
DOI and file metadata verified live against the Dryad landing page on 2026-04-06. The 13.10 MB total matches the EPIC-009 'Alabama EHRM GPS Data — 13 MB' line item. Dryad's API requires authentication for downloads as of 2026 (https://datadryad.org/openapi.yml: 'Anonymous users... are not allowed to download data files'); the importer supports both OAuth2 client_credentials and a manually-staged cache file. CSV header schema and 191,867 GPS fixes verified by direct parse of the cached file. ALL DATA MUST BE REAL — this dataset is real Auburn University Deer Lab / Ditchkoff GPS telemetry from Alabama, not synthesized.
Hanberry, B.B., Hanberry, P. · 2019
Findings used in Hunt IQ
Referenced from
Note
Replaces the vaporware USDA density layer from EPIC-009 with a real, verified, publicly-available dataset. STORY-058 / EPIC-010.
1 external service APIs queried at runtime or in the pipeline.
iNaturalist Network · 2026
Findings used in Hunt IQ
Referenced from
10 modules contribute to every recommendation. Each module documents its inputs, outputs, and the citations that back its coefficients.
Scores DEM-derived terrain features (slope, aspect, TPI, elevation position, canopy fit) per candidate.
src/lib/scoring/terrain.ts
Classic heuristic terrain scoring. Given a candidate location extracted from the LiDAR pipeline (saddle, funnel, ridge, bench, field edge, or bedding area), computes a 0–100 score based on how well the local terrain matches the weapon-specific canopy range and general deer movement heuristics (preference for moderate slopes, south/southeast aspects, concave TPI for saddles, etc.).
Inputs
Output
0–100 terrain score with per-factor breakdown for the Score Breakdown UI.
Note: This is the heuristic layer that RSF scoring replaces when real GPS-collar data coefficients are available.
Replaces heuristic terrain scoring with published Resource Selection Function beta coefficients applied to local covariates.
src/lib/scoring/rsf-scoring.ts
The RSF surface is pre-computed per geography by pipeline/compute_rsf_surface.py using published seasonal beta coefficients from Darlington et al. 2022 (Scientific Reports) applied to DEM slope/aspect, NLCD land cover, NHD distance-to-water, and computed distance-to-edge. At request time, rsf-scoring.ts reads the candidate's pre-enriched RSF probability, rescales it to 0–100, adds feature-type bonuses (funnel +25, saddle +20, etc.), and applies weapon-specific canopy modifiers. When RSF data is unavailable for a geography (pipeline not yet run), it falls back gracefully to the heuristic terrain score.
Inputs
Output
Blended terrain score, feature bonus, weapon modifier, and a scoring_mode flag (rsf | heuristic | both).
Hour-by-hour prediction of deer behavior states (BEDDING / FORAGING / TRAVELING) from real GPS collar activity budgets.
src/lib/scoring/behavior-states.ts
Implements a 3-state hidden-Markov-inspired model using real GPS collar activity budgets loaded from src/lib/scoring/config/behavior-matrices.json. The matrices encode 24-hour activity percentages from MSU Deer Lab 15-minute GPS fix data (Strickland et al. 2023) plus the counterintuitive midday-buck-peak finding from the Penn State Deer-Forest Study. Photoperiod-based season detection auto-adjusts the rut phase by latitude and date. Moon phase is explicitly excluded per Webb et al. 2010's seven-year, 32-deer finding of zero measurable effect.
Inputs
Output
Per-hour movement probability and dominant behavior state for the timing chart.
Note: Paterson 2023 is cited for HMM methodology only — that paper studied pronghorn and mule deer, not white-tailed deer. WTD-specific numbers come from Strickland 2023 and Webb 2010.
Photoperiod-based rut phase detection with daily distance and home range predictions.
src/lib/scoring/rut-tracker.ts
Walks forward and backward from the given date to detect phase boundaries using photoperiod (day length), then looks up phase-specific daily movement distance (yards), home range expansion (acres), and hunting strategy from the MSU P3927 publication. The 1.7× rut-movement claim in EPIC-009 was reconciled against this source: the actual published numbers are 4,800 yd/day pre-rut → 7,500+ yd/day peak rut (1.56× daily total), with NET DISPLACEMENT growing 1.7× (see STORY-065 for details).
Inputs
Output
Rut phase, days into phase, phase progress, buck daily distance, home range acres, hunting strategy text, day length.
Multi-factor hunting pressure model combining season-long decay, stand-specific learning, and the Oklahoma hunt-day effect.
src/lib/scoring/pressure.ts
Composite 0–1 pressure score that multiplies the behavior score. Four factors: (1) temporal decay via exp(−0.025 × daysSinceOpener) calibrated to Penn State's ~25% of opener by late November finding; (2) stand-specific sigmoid avoidance via 1 / (1 + exp(2.0 × (effectiveVisits − 2.5))) calibrated to Auburn's 4+ visit abandonment finding (Sullivan et al. 2018); (3) calendar weekend/weekday modifier; (4) land-type multiplier (0.67 public / 1.0 private). STORY-063 added the explicit Oklahoma hunt-day effect from Marantz et al. 2016: weekends during hunting season get an additional multiplier (−17% daytime / −31% nighttime movement) applied via applyHuntDayEffect. STORY-040 also exposes getCoverPreferenceShift() for the Penn State 240% cover-use escalation used by the terrain path.
Inputs
Output
Composite pressure score 0–1, per-factor breakdown, human-readable label, days-to-recovery estimate, cover-shift multiplier.
Rate-of-change weather modifier (temperature delta, pressure trend, cold fronts) applied to the base weather score.
src/lib/scoring/weather-movement.ts
GPS collar research consistently shows that RATE OF CHANGE is a stronger movement predictor than absolute values. A 15°F+ temperature drop with rising barometric pressure (a cold front) produces the single strongest movement response in the literature. This module computes a composite multiplier (0.3–2.0) from temp delta, pressure trend, wind impact, and precipitation impact, then applies it to the raw weather score from the Open-Meteo HRRR forecast. Moon phase is explicitly excluded.
Inputs
Output
Weather-movement modifier 0.3–2.0, per-factor breakdown, cold-front-detected flag.
Predicts deer arrival time at food sources with pressure-adjusted shift from early to late season.
src/lib/scoring/food-plot-timing.ts
For candidates near food sources (NLCD land cover 71/81/82 or field_edge feature), predicts the estimated arrival time relative to sunset. Opening-week arrivals average 30–60 min before sunset; by gun-season week 2 arrivals shift to at-or-after dark. Includes a rut-disruption flag for bucks during peak rut (when food-plot patterns break down in favor of doe-seeking). Pressure level is derived from the pressure model's spatial factor.
Inputs
Output
Estimated arrival HH:MM, minutes relative to sunset, legal-hours flag, recommended setup time, rut-disrupted flag, explanation text.
A* least-cost-path corridors between predicted bedding and feeding areas with funnel detection.
src/lib/corridor/auto-corridors.ts
Pre-computes a cost surface per geography from DEM slope + NLCD land cover + water barriers (pipeline/compute_cost_surface.py). At recommendation time, detects bedding and feeding clusters from PostGIS terrain features, runs A* with a MinHeap across the cost grid to find least-cost paths between each bedding-feeding pair, and identifies topographic funnels (pinch points) along the paths. Candidates near a corridor or funnel get a terrain score boost of 0–15 points.
Inputs
Output
CorridorResult[] with paths, travel times, funnels detected, and per-candidate corridor proximity boost.
A* walking route from the property edge to each recommended stand, with wind, bedding, and water cost layers.
src/lib/routing/approach-planner.ts
For every top-N recommendation, plans the lowest-cost walking route from a polygon edge entry point to the stand. The ~10m grid composes four cost layers: (1) base distance, (2) downwind scent exposure — a wind cone projected from each other predicted deer position toward the hunter so the route avoids carrying scent into the rest of the herd, (3) bedding-zone avoidance — a gradient around predicted bedding sites so the hunter doesn't bump deer on the approach, and (4) NHD HR water polygons — a hard barrier inside the polygon plus a soft shoreline buffer within ~30m. The auto-detected entry point is biased toward crosswind/downwind approach bearings, or the user can pin an explicit entry point. The static layers (polygon mask, bedding cost, water cost) are built once per request via buildSharedRouteContext and reused across all candidate routes; only the per-candidate scent overlay is recomputed. A* uses a binary min-heap open set, and the path is smoothed via Chaikin's corner-cutting after pathfinding.
Inputs
Output
GeoJSON LineString attached to each recommendation as approach_route, with distance_miles, time_minutes, wind_exposure_score (0–100), and a natural-language start_description.
Note: Water avoidance was added to fix routes that drew across reservoirs and large lakes. The hard-barrier polygon mask plus soft shoreline buffer guarantee no path segment crosses standing water, even at the cost of longer walks. The shared static grid + binary heap optimization in STORY-068 took the per-request planner cost from 25,778 ms to 40 ms.
Validates predicted RSF probabilities against real Dryad GPS collar fixes using the continuous Boyce Index and AUC.
pipeline/validation/validate_rsf_against_dryad.py
Standard wildlife-ecology validation methodology. Loads real used-points from the collar_tracks PostGIS table (with a hard safeguard that REFUSES to validate against synthetic data), generates 10× random available-background points within the study minimum convex polygon, samples the RSF GeoTIFF at both sets of points, and computes the continuous Boyce Index (Hirzel et al. 2006) plus Mann-Whitney U AUC plus per-season calibration bins. Outputs a markdown report and machine-readable JSON artifact the UI can display.
Inputs
Output
Boyce Index, AUC, calibration bin table, research/validation/rsf_validation_<season>.md + .json
Note: This is the story that lets a client ask "how do you know the RSF model actually works?" and get a real number back instead of hand-waving.
Citations
How a single candidate becomes a ranked recommendation.
When a user draws a hunt area and clicks “Optimize”, Hunt IQ runs the following pipeline. Each step either reads real input data or applies a literature-derived model, with every coefficient traceable to a citation on this page.
Query the terrain_features PostGIS table for saddles, ridges, funnels, benches, field edges, and bedding areas within the user's drawn polygon. These come from the USGS 3DEP LiDAR pipeline.
The terrain_features query already excludes any feature whose location intersects an NHD HR water polygon. This is enforced as a schema-level invariant — the is_in_water flag is populated once at seed time by intersecting every terrain feature against the water_bodies table — rather than re-evaluated per request. Flooded candidates never reach the scoring engine. Deer don't bed in open water.
Run terrain.ts heuristic + rsf-scoring.ts RSF blend. RSF takes precedence when pipeline has pre-enriched rsf_score for the geography. Apply feature-type bonus (funnel +25, saddle +20, etc.) and weapon modifier.
Query Open-Meteo HRRR for the hunt date/time and location. Compute base weather score (wind / pressure / precip / temperature). Apply weather-movement.ts cold-front modifier on top.
Run behavior-states.ts HMM model to predict hour-by-hour movement probability for the hunt window. Apply rut-tracker.ts phase-specific multipliers.
Run pressure.ts to compute the composite pressure modifier. Multiply the behavior score by pressure.score, then apply the Oklahoma hunt-day effect (−17% on weekend in-season days) via applyHuntDayEffect.
If a cost grid exists for the geography, run auto-corridors.ts to compute A* least-cost paths between predicted bedding and feeding clusters. Boost terrain score for candidates near a corridor or funnel.
For candidates near food sources, run food-plot-timing.ts to predict arrival time relative to sunset, factoring in pressure level and rut phase.
Query the sightings PostGIS table for real user-reported sightings within 500m. Apply a recency-weighted boost (14-day exponential half-life) to the behavior score.
composite = 0.40·terrain + 0.20·weather + 0.25·behavior + 0.15·access. Rank candidates by composite score and return the top N.
For each top-N recommendation, run approach-planner.ts to find the lowest-cost walking path from a polygon edge entry point to the stand. The cost grid composes downwind scent exposure (a wind cone projected toward other predicted deer positions), bedding-zone avoidance, NHD HR water polygons (hard barrier inside polygons + ~30m soft shoreline buffer), and base travel distance. Returns a smoothed GeoJSON LineString attached to the recommendation as approach_route with distance, time, and a wind-exposure score.
Build the movement_intelligence.citations[] array from the union of CITATIONS_USED exports across every module that contributed. This is what powers the clickable CitationTip popovers in the UI.
How we know the models actually work (and where we don't yet know).
Any RSF-style habitat model is worth less than the validation backing it up. Applying published coefficients to local covariates is arithmetic; verifying that the resulting surface tracks real deer use is science.
Hunt IQ ships a validation harness at pipeline/validation/validate_rsf_against_dryad.py that samples the predicted RSF probability at real GPS fixes from the Dryad NYC dataset (VanAcker et al. 2023) versus random available-background points, and computes two standard validation metrics:
Spearman rank correlation between habitat-suitability class and P(used)/P(available) ratio. Values above 0.5 are good, above 0.8 excellent, below 0 are a red flag (inverse relationship).
Hirzel et al. 2006 — Ecological Modelling, DOI 10.1016/j.ecolmodel.2006.05.017
Probability that a random used point has a higher RSF value than a random available point. Values above 0.7 are good, above 0.8 excellent, below 0.6 indicate the model barely beats random.
Standard ROC analysis
The harness refuses to run against synthetic study sources as a hard safeguard against fake validation numbers. Actual execution requires the Dryad NYC cache to be pre-staged (Dryad auth policy blocks anonymous downloads) and the RSF GeoTIFF to be computed for the matching region. When those are in place, the results will be published at research/validation/rsf_validation_<season>.md and surfaced on the RSF layer legend.
Honest science requires honest limitations.
Hunt IQ does not ingest live collar data from an ongoing study. The GPS datasets we load (Dryad NYC, Dryad Alabama, Movebank Florida) are historical and static. When collar partnerships materialize in the future, the ingestion pipeline is ready (pipeline/collar_data/) but it's not currently running against a live feed.
The primary RSF source (Darlington et al. 2022) studied female white-tailed deer in Alberta, Canada during winter. Hunt IQ applies those betas across five seasons (early / pre-rut / rut / post-rut / late). This is a reasonable extrapolation but it's not a guarantee that the coefficients generalize perfectly. STORY-060's validation harness is the check.
The specific 25% / 240% / 10→21% / midday buck peak numbers appear in Penn State Deer-Forest Study extension articles and blog posts on deer.psu.edu, not in peer-reviewed journal publications. The study itself is real and long-running, but these specific numbers are extension content. The registry entry for penn-state-deer-forest-study is marked LIKELY rather than VERIFIED to reflect this.
This HMM methodology paper is cited for its modeling approach, not for WTD-specific findings. The WTD-specific empirical grounding comes from Strickland et al. 2023 (MSU P3927) and Webb et al. 2010. The species caveat is documented on the citation entry.
The Kauffman et al. 2022 UMRv3 dataset is the only published white-tailed deer migration corridor dataset on USGS ScienceBase. It's geographically limited to 43 GPS-collared deer in Northeast Washington. Hunt IQ users outside that extent will see no published reference corridors — which is a data-availability reality, not a bug.
The RSF coefficients currently apply equally to buck and doe hunting. STORY-064 is planned to add buck/doe differentiation using real published sex-specific betas where available.
This is a deliberate design choice, not a limitation. MSU Deer Lab and Webb et al. 2010 (32 deer, 7-year GPS collar study) both found zero measurable effect of moon phase on deer movement rates. Including moon phase would be pseudoscience.
15 published papers and extension publications. Grouped by kind.
Paterson, J.T., Johnston, A.N., Ortega, A.C., Wallace, C., Kauffman, M. · 2023
Findings used in Hunt IQ
Referenced from
Note
Species caveat: this paper did NOT study white-tailed deer. It is cited as the HMM ungulate methodology reference. The white-tailed-deer-specific empirical grounding comes from strickland-2023-buck-movement and webb-2010-fine-scale-movements.
Darlington, S., Ladle, A., Burton, A.C., Volpe, J.P., Fisher, J.T. · 2022
Findings used in Hunt IQ
Referenced from
Note
Author correction (2026-04-06): prior EPIC-009 drafts guessed 'Kjær' — the real lead authors are Darlington, Ladle, Burton, Volpe, Fisher, verified via the DOI landing page. Sample is 38 female WTD in Alberta (not mixed-sex), winter-focused.
Kays, R., Davidson, S.C., Berger, M., Bohrer, G., Fiedler, W., Flack, A., et al. · 2022
Findings used in Hunt IQ
Referenced from
Long, E.S., Diefenbach, D.R., Lutz, C.L., Rosenberry, C.S., Wallingford, B.D. · 2021
Findings used in Hunt IQ
Referenced from
Sullivan, J.D., Ditchkoff, S.S., Collier, B.A., Ruth, C.R., Raglin, J.B. · 2018
Findings used in Hunt IQ
Referenced from
Marantz, S.A., Long, J.A., Webb, S.L., Gee, K.L., Little, A.R., Demarais, S. · 2016
Findings used in Hunt IQ
Referenced from
Michelot, T., Langrock, R., Patterson, T.A. · 2016
Findings used in Hunt IQ
Referenced from
Lutz, C.L., Diefenbach, D.R., Rosenberry, C.S. · 2015
Findings used in Hunt IQ
Referenced from
Little, A.R., Demarais, S., Gee, K.L., Webb, S.L., Riffell, S.K., Gaskamp, J.A., Belant, J.L. · 2014
Findings used in Hunt IQ
Referenced from
Webb, S.L., Gee, K.L., Strickland, B.K., Demarais, S., DeYoung, R.W. · 2010
Findings used in Hunt IQ
Referenced from
Beier, P., Majka, D.R., Spencer, W.D. · 2008
Findings used in Hunt IQ
Referenced from
Chetkiewicz, C.B., St. Clair, C.C., Boyce, M.S. · 2006
Findings used in Hunt IQ
Referenced from
Hirzel, A.H., Le Lay, G., Helfer, V., Randin, C., Guisan, A. · 2006
Findings used in Hunt IQ
Referenced from
Strickland, B.K., Demarais, S., Cain, R., Chance, A., Henderson, C., Street, G., Resop, L., McKinley, W. · 2023
Findings used in Hunt IQ
Referenced from
Diefenbach, D.R., McDill, M., Pennsylvania Cooperative Fish and Wildlife Research Unit · 2013-ongoing
Findings used in Hunt IQ
Referenced from
Note
CAVEAT (2026-04-06 research agent audit): the specific numeric claims listed above (25%, 240%, 10→21%, 12-1 PM midday peak) could NOT be traced to peer-reviewed Deer-Forest Study publications. They appear in extension articles, blog posts, and podcast transcripts on deer.psu.edu, not in journal papers. They are still grounded in real GPS collar observations from a real long-term study, but any scientific client should be told these specific percentages come from extension content rather than peer-reviewed literature. Peer-reviewed Diefenbach-lab outputs include: Lutz et al. 2015 (Journal of Mammalogy) on female dispersal, and Long et al. 2021 (Ecology and Evolution, DOI 10.1002/ece3.7227) on dispersal survival. See lutz-2015-dispersal and long-2021-dispersal-survival citations for peer-reviewed alternatives.