Geospatial ML · Species Distribution Modeling
Geospatial machine learning that predicts where edible mushrooms are fruiting across Washington State.
Overview
SporeCast forecasts habitat suitability for 20 gourmet mushroom species across Washington, combining species distribution modeling, satellite imagery analysis, weather-aware phenology, and real-time web mapping into a single system.
The hard part is not fitting a classifier. Presence-only biodiversity data is spatially biased toward roads and trailheads, absence data does not exist, and a naive model learns to separate forest from desert rather than good habitat from ordinary habitat. Most of the engineering here is about making an honest model out of dishonest data.
Every data source is a keyless public API or a server-side reduction, so there are no raster downloads, no vendor lock-in, and the whole pipeline redeploys from scratch.
Architecture
01
Pool presence records from four independent biodiversity archives, then de-duplicate and filter by coordinate accuracy and habitat plausibility.
02
Attach 40+ covariates to every observation and prediction cell across climate, terrain, hydrology, land cover, and satellite embeddings.
03
Train a per-species ensemble under spatially blocked cross-validation, with effort-corrected background weighting and isotonic calibration.
04
Build a debiased monthly fruiting curve per species, then shift it using the current season's regional soil-moisture anomaly.
05
Freeze suitability, uncertainty, habitat masks, and metadata into versioned artifacts so serving never depends on a live model.
06
Render the suitability raster in the browser and apply elevation, host, and accessibility gates client-side for instant interaction.
Features
Every prediction cell carries covariates drawn from sources that disagree about resolution by three orders of magnitude. Reconciling them — without pretending coarse data is fine-grained — is most of the feature engineering.
Approach
Seven learners with deliberately different inductive biases — bagged trees, boosted trees, and linear models including a MaxEnt-style formulation — are combined by a meta-learner trained on out-of-fold predictions. When a species has too little data to train the stacker honestly, it degrades to a performance-weighted average rather than overfitting the blend.
Random k-fold leaks badly on spatial data: a held-out point sitting ▓▓▓ from its training neighbor is not a real test. Block size is derived from the empirical variogram of elevation, soil moisture, and water distance together, rather than assumed or tuned by hand.
Roughly ▓▓▓ of background weight is concentrated on forested cells within ▓▓▓ of known finds. This is the single highest-leverage decision in the system — it forces the model to discriminate within habitat instead of learning the trivial forest-versus-desert boundary.
Presence-only data reflects where people look, not where mushrooms grow. Background is reweighted by a distance-decayed count of all target-group observations, and presences are thinned so no two retained points sit within ▓▓▓ of each other — which strips out roadside and trailhead clustering.
Climate normals arrive on a coarse grid far wider than the prediction cells. Monthly normals are shifted by the elevation difference between each point and its grid cell at ▓▓▓, then temperature aggregates are recomputed — a large effective resolution gain for almost no compute.
Per-pixel 64-dimensional AlphaEarth vectors encode vegetation texture, moisture signature, and seasonality at a resolution no land-cover class can express. These are reduced by PCA for the ensemble, and separately scored against multi-modal per-species centroids to produce a lookalike mask.
Parasitic and mycorrhizal species are gated on the modeled suitability of their hosts rather than treated as independent. Each species carries an affinity profile over forest community types ▓▓▓, plus a soft substrate preference, so predictions respect the biology instead of only the correlations.
A static seasonal average is the obvious trap — it is right on median and wrong every actual year. Monthly curves are nudged by the current season's soil-moisture anomaly with a lag of ▓▓▓, computed per region so a dry east-Cascades autumn and a soaked Olympic one diverge on the same map.
Modeling
Rigor
Full-background AUC reads about 0.99 for every species, which is meaningless — it measures forest versus desert. The reported metric compares presences against the forested serving grid, which is the question a forager actually asks.
Two independent holdouts: train on older years and test on recent ones, and train on one archive while testing on others. Together they measure genuine forecasting skill rather than memorization of a single platform's sampling quirks.
Isotonic regression on balanced class weights makes scores comparable across species. Brier score and calibration curves are tracked alongside discrimination metrics, because a well-ranked but badly calibrated surface produces confident nonsense.
Confidence intervals come from a block bootstrap over spatial folds, and permutation importance is computed on held-out folds — in-sample permutation importance on tree ensembles is close to meaningless.
Trade-offs
Models train across the wider Pacific Northwest and predict only onto Washington. State lines are political, not ecological, and truncating training data at one multiplies the effective sample size for rare species.
Every feature is either snapped from a keyless REST endpoint or reduced server-side before it crosses the network. The pipeline stays deployable and reproducible without a data-warehouse step.
Optional feature families self-prune when their upstream source is unreachable. A soil API outage costs a few features and some accuracy; it does not fail the build.
Public-land polygons post-filter the surface so results only appear where harvesting is actually legal. A prediction the user cannot act on is not a feature.
Built with
This write-up names the techniques but not the tuning. The redacted values above, and the following, stay in the private repository:
Happy to walk through any of it in depth — get in touch.