Primary document · v1 · last updated from docs/METHODOLOGY.md on render

Basal Informatics — Methodology

One-page brief for ag lenders and loan-review committees. Intended audience: Farm Credit System associations, regional agricultural banks, and the portfolio-management / collateral-review teams that commission ecological condition assessments on parcels they lend against.

Who submits the data

Landowners with active agricultural loans. Submission is either required at loan renewal or incentivized through rate terms — the same documentation pattern that already covers soil tests, yield records, and irrigation logs. The landowner exports their own trail-cam SD cards and uploads a single .zip through a Basal web form; the resulting Nature Exposure Report is shared with their lender for collateral review.

Basal does not collect data from the landowner without their participation. The lender commissions the report; the landowner supplies the raw photos; Basal produces the methodology-backed output.

What we measure

Pipeline outputs — index + density + tier, in that order of primacy:

1. Detection frequency (events per camera-day): the raw relative abundance index, computed directly from independent-event counts. This is the pre-REM, minimum-assumption quantity. Two parcels surveyed with identical camera deployments can be compared on this number alone without invoking any movement-distance assumption.
2. Density estimate (animals/km²): the REM-scaled per-area population derived from the detection rate using per-species daily travel distance. Reported with a bootstrap 95% confidence interval.
3. Tier (Low / Moderate / Elevated / Severe, feral hog only v1): the binary-decision-grade classification per Mayer & Brisbin 2009 hog-density bins. Drives the loan-review committee's go / condition / no-go / survey-required decision path.

Alongside these we also publish an Exposure Score (0–100, piecewise-linear anchored on the tier cutoffs) for visual legibility on a single gauge, and a modeled annual crop-damage projection as supplementary context — but those are derivations, not the primary outputs.

Damage dollars are NOT a pipeline output. The relationship between hog density and parcel-scale crop loss is poorly characterized in the literature (producer-survey recall bias, state-level extrapolations that break down at parcel scale). We produce the density + rate + tier with methodology-defensible confidence intervals; the dollar-scaling is attached separately under a supplementary_projection object in the JSON API, with its own disclaimer. A loan committee with an internal damage model should consume the pipeline outputs directly and ignore our dollar block; committees without an internal model can use our dollar figure as a convenience projection, clearly labeled.

Estimator: Random Encounter Model (REM)

Rowcliffe, Field, Turvey & Carbone 2008 (Journal of Applied Ecology):

D = (y/t) · π / (v · r · (2 + θ))

Why REM, not capture-recapture or N-mixture? REM does not require individual identification, which is unreliable at population scale for species without natural marks (feral hog, deer at distance, raccoon). It also does not require closed-population assumptions, which are violated by transient species at parcel scale.

Per-species movement parameters

Hard-coded in config/settings.py (SPECIES_MOVEMENT):

For species without a published v (e.g. raccoon, opossum), we report the raw detection rate (events per camera-day) as an unscaled index and explicitly omit the density estimate. The recommendation flag reads "insufficient data" with a method note explaining why.

Confidence intervals

Bootstrap 95% via 1000 iterations: - Camera bootstrap: resample cameras with replacement (the design's primary stochastic source per Rowcliffe 2012). - Movement-distance perturbation: per-iteration v_sample ∼ N(v, sd), truncated to [0.5·v, 1.5·v]. Truncation prevents the upper CI tail from inflating ~10× under physically implausible v < 0.5·v_published values; the published sd captures inter-individual / inter-region variation, not within-survey uncertainty.

Bias correction

Camera placement is non-random in operational deployments. Feeders, trails, water sources, and food plots inflate per-camera detection rates by 1.4-9.7× depending on species and context (Kolowski & Forrester 2017). Without correction, REM density inherits this inflation directly. The pipeline applies two complementary corrections on the per-camera detection rate before it enters REM, and reports both alongside the raw rate.

Method 1 — Literature-prior ratio adjustment (PRIMARY)

For each camera at a non-random context, deflate the observed per-camera rate by a per-species, per-context inflation factor sourced from the literature, then average across cameras:

adjusted_rate = mean_i ( rate_i / inflation_factor[species, context_i] )

Inflation factors for feral hog (config/settings.py → bias.placement_ipw.DEFAULT_INFLATION_FACTORS):

This is the only sound correction when no random-placement cameras exist in the deployment, which is typical for hunter-style camera arrays. Trade-off: sensitive to inflation-factor accuracy. Default factors are mid-range estimates; project-specific calibration tightens them.

Method 2 — Hájek IPW with empirical propensities (DIAGNOSTIC)

The textbook IPW estimator (Hájek 1971; Cassel-Särndal-Wretman 1976) reweights the sample to a target placement-context distribution:

weighted_rate = Σ_i ( w_i · rate_i ) / Σ_i w_i
where w_i = q(context_i) / p(context_i)

p is the empirical propensity (proportion of cameras at each context); q is the target marginal (default: uniform across contexts present). Empirical IPW alone cannot correct bias when the entire sample is biased — there is no unbiased anchor. It is reported as a sanity-check companion to the literature-prior method, not fed into REM.

Diagnostics

For each survey period the pipeline reports Kish's effective sample size, ESS = (Σw)² / Σ(w²) (Kish 1965), and the maximum-weight ratio max(w) / mean(w). Caveats fire automatically when:

• ESS drops below n / 2 (significant statistical-power loss from weighting).
• max-weight ratio exceeds 5× (one camera dominates the estimate; Cole & Hernán 2008 recommend stabilization or trimming).
• No random-placement cameras in the deployment (literature-prior factors carry the entire correction; cannot be cross-validated).

Each camera carries a placement_context value supplied by the landowner at setup. Both the raw and bias-adjusted rates appear in the dashboard and JSON API. REM density is derived from the adjusted rate; the raw rate is reported alongside for transparency.

Recommendation logic

Per species, per survey period:

Thresholds are tunable in config/settings.py: MIN_CAMERA_DAYS_FOR_DENSITY, MIN_DETECTIONS_FOR_DENSITY, DENSITY_CI_RATIO_THRESHOLD.

What the loan-review committee gets

For each species on each parcel for each survey period, both the lender dashboard (/lender/<slug>/parcel/<id>) and the JSON API (/lender/api/<slug>/parcel/<id>/exposure) return a structured record suitable for import into the lender's internal portfolio- management system:

The record shape explicitly separates pipeline outputs from supplementary modeled projections, so downstream importers cannot accidentally treat the damage dollar figure as a pipeline output:

{
  "species_key": "feral_hog",

  "pipeline": {
    "tier":                           "Elevated",
    "score_0_100":                    50.6,
    "density_animals_per_km2":        5.13,
    "density_ci_low":                 1.29,
    "density_ci_high":                16.64,
    "detection_rate_per_camera_day":           0.397,
    "detection_rate_adjusted_per_camera_day":  0.099,
    "recommendation":                          "recommend_supplementary_survey",
    "caveats": [
      "Cameras at non-random placements (feeder, trail) violate REM's
       movement-independence assumption. Inverse propensity weighting
       (Kolowski & Forrester 2017) corrects for residual bias but does
       not eliminate it."
    ],
    "method_notes": [
      "Daily travel distance: v = 6.0 km/day (sd 2.5). Source: Kay et al. 2017."
    ]
  },

  "supplementary_projection": {
    "label":                     "MODELED PROJECTION",
    "source":                    "Anderson et al. 2016; APHIS Wildlife Services annual Program Data Reports",
    "annual_damage_usd":         25561,
    "annual_damage_ci_low_usd":   6421,
    "annual_damage_ci_high_usd": 82980,
    "crop_modifier":             1.30,
    "per_hog_annual_usd":        405,
    "disclaimer": "Not a pipeline output. Derived from third-party loss
                   data (Anderson et al. 2016 per-hog damage figures
                   × parcel area × crop modifier). Intended as context
                   for loan-review committees that have not yet built
                   their own damage model; a committee with an internal
                   model should consume the pipeline outputs above
                   instead."
  }
}

The parallel under stats also includes n_cameras, total_camera_days, and total_detections at the property-level summary.

Audit trail is retained at camera-day granularity: individual detection timestamps, SpeciesNet inference confidence per photo, and the raw ZIP the landowner submitted. Available for any downstream review the lender or its auditor runs.

Pricing

• Per parcel-verification: $1,500 per report. One-time per survey window.
• Portfolio unlimited: $5,000 / month. Unlimited parcels in the lender's portfolio.

The alternative is a ~$40,000 independent field-biologist survey that's point-in-time and goes stale inside six months. Basal is ~25× cheaper at the per-parcel tier and continuously refreshable rather than stale.

What we do not claim

• We do not estimate damage dollars.
• We do not estimate density for species without published movement parameters.
• We do not infer presence outside the camera's detection cone.
• We do not extrapolate beyond the surveyed property without explicit habitat-similarity tooling (separate product line).

References

1. Rowcliffe JM, Field J, Turvey ST, Carbone C. 2008. Estimating animal density using camera traps without the need for individual recognition. Journal of Applied Ecology 45: 1228–1236.
2. Rowcliffe JM, Carbone C, Jansen PA, Kays R, Kranstauber B. 2011. Quantifying the sensitivity of camera traps using an adapted distance sampling approach. Methods in Ecology and Evolution 2: 464–476.
3. Kolowski JM, Forrester TD. 2017. Camera trap placement and the potential for bias due to trails and other features. PLOS ONE 12: e0186679.
4. Kay SL et al. 2017. Quantifying drivers of wild pig movement across multiple spatial and temporal scales. Movement Ecology 5: 14.
5. McClure ML et al. 2015. Modeling and mapping the probability of occurrence of invasive wild pigs across the contiguous United States. PLOS ONE 10: e0133771.
6. Webb SL, Hewitt DG, Hellickson MW. 2010. Survival and cause-specific mortality of mature male white-tailed deer. Journal of Wildlife Management 74: 1416–1421.
7. Andelt WF. 1985. Behavioral ecology of coyotes in south Texas. Wildlife Monographs 94: 3–45.