How the search works

Chestly does not ask 'is this pixel an American chestnut?' No satellite currently in orbit can answer that. Sentinel-2 resolves the ground at 10 metres; a single mature tree's canopy is often smaller than one pixel and is always mixed with its neighbours.

The question it does ask is: given everything known about where surviving American chestnuts have been documented, which patches of forest most resemble those places, and which of those are worth a person's time to visit?

That reframing matters. It turns an identification problem, which is currently unsolvable from imagery, into a search-prioritisation problem, which is not.

Thirteen figures, generated directly from the current pipeline's data — regenerated whenever the pipeline runs, never drawn from memory. Panels that use a chosen best-case example say so in their caption; the validation panel is computed only on spatially held-out predictions and is kept deliberately separate from the illustrations.

Documented American chestnut observations are sourced from iNaturalist, retrieved through its public API and refreshed by an ingestion step that updates existing records rather than duplicating them. Each record on the map links back to the original.

These are community-contributed records. They carry two kinds of uncertainty that matter here: whether the identification is correct, and how precisely the location was recorded. Both are stored alongside the observation and shown on the map rather than smoothed away.

iNaturalist assigns each record a quality grade. Research Grade means enough independent identifiers have agreed — a useful filter, and the map's default, but a community consensus rather than a determination. American chestnut is regularly confused with Chinese chestnut and chestnut oak, so a Research Grade record is strong evidence and not proof.

Where iNaturalist obscures a record's coordinates, that decision is respected. Obscured records publish a deliberately imprecise point, are flagged as such in the interface, and no attempt is made to recover the true location. Cultivated and planted trees are excluded by default: a chestnut somebody planted says nothing about where a wild survivor might stand.

Positive examples come from documented observations of Castanea dentata, primarily iNaturalist records, filtered for identification quality and positional accuracy. Cultivated and planted trees are excluded where they can be identified: a chestnut somebody planted says nothing about where a wild survivor might stand.

Every observation carries a confidence value rather than being treated as ground truth. Community identifications are evidence and are sometimes wrong, particularly against Chinese chestnut and chestnut oak.

Where a source obscures a record's location for conservation reasons, that designation is preserved. An obscured point is a random position inside a large box, and treating it as precise would sample satellite features from the wrong forest entirely.

Comparing chestnut observations against randomly chosen forest would produce a model that separates 'places people go' from 'places people do not go'. The negative set is built deliberately instead.

Hard negatives are confirmed observations of species that look or behave similarly: Chinese chestnut, chestnut oak, American beech, and selected oaks and maples. These are the discriminations that actually matter, and they are the ones a naive negative set never teaches.

Geographic negatives are forest near known chestnut habitat with nothing on record. They are stored with low confidence and down-weighted during training, because absence of an observation is not absence of a tree. Treating them as confident negatives would teach the model that undocumented chestnuts are non-chestnuts — precisely backwards for a project trying to find undocumented chestnuts.

Sentinel-2 L2A surface reflectance, discovered through the public Earth Search STAC catalogue and read directly from the open-access archive on AWS. Only the pixels around each study site are fetched — windowed reads from cloud-optimised GeoTIFFs — so no whole scenes are downloaded or stored.

Every scene is masked with its scene-classification layer before use: cloud, shadow, cirrus and defective pixels are discarded, and each seasonal window is reduced to a median composite of what remains. Where masking leaves too little clear imagery, the window is marked missing rather than filled in.

From each composite the pipeline computes vegetation indices and band ratios, and retains the underlying bands — including the red-edge bands, which sit on the steep reflectance transition between red and near-infrared and are among the few places Sentinel-2 can plausibly separate species that look alike in visible light.

Phenology — the timing of a canopy's annual cycle — is likely the strongest available signal. Two species can reflect almost identically in August and still leaf out three weeks apart.

Four windows are composited each year: spring leaf-out, the early-summer flowering period, peak summer canopy, and autumn senescence. American chestnut flowers conspicuously in early summer, with creamy-white catkins covering the crown, and that window carries the most distinctive expected signature — an expectation Chestly has now tested directly, with results reported honestly in the flowering experiment section below.

Window dates are configured per region and shifted with latitude, because spring arrives roughly three days later per degree northward in the eastern United States. A fixed calendar window would compare leaf-out in one place to full canopy in another and call the difference a species signal.

The flowering hypothesis was attractive enough to deserve a real test rather than an assumption. In summer 2026 Chestly ran one: can Sentinel-2 detect American chestnut flowering strongly enough to help identify trees? The design used every Pennsylvania iNaturalist record with a community flowering annotation, habitat-matched forest controls a few hundred metres from each tree, Chinese chestnut records as a specificity control, and five years of imagery per site — with features measuring both change over time and contrast against the immediately surrounding canopy.

The honest answer is no — not at this resolution, with this method. Individual seasonal changes at chestnut sites are real and measurable, but they do not separate chestnuts from ordinary forest: the flowering-window brightening the hypothesis predicts shows up no more often at confirmed flowering chestnuts than at their paired controls, and nothing tested distinguishes American chestnut from Chinese chestnut. Adding flowering features to a baseline habitat model made spatially held-out classification slightly worse, not better.

The failure modes were instructive. The largest apparent flowering-window brightening in the whole dataset traced to thin cirrus haze that the standard cloud mask passes — an artefact that would have looked like a discovery under a less suspicious analysis. And the strongest 'signal' found — chestnut sites differing from surrounding canopy — is static across seasons, which means it reflects where these trees grow, not the flowering event.

A null result changes the architecture, not the mission. Flowering evidence is retired as an identification signal at 10 m; habitat and general phenology carry the model until higher-resolution imagery or a better method reopens the question. The figures below show the experiment's raw material at its most favourable: three confirmed flowering chestnuts, chosen deliberately as the strongest measured cases with verified-clear imagery. Even here, the flowering window looks like forest.

Imagery is not the only input. A separate habitat score is built from elevation, slope, aspect, forest cover and structure, distance from development, and soil characteristics where they are available.

Aspect is decomposed into northness and eastness before it reaches a model. Raw aspect in degrees is circular: 1 and 359 describe adjacent slopes but sit at opposite ends of the numeric range, which a tree-based model would happily split on.

An eligibility mask runs before anything is scored, removing water, dense development, highways, large agriculture and terrain outside the elevation and slope ranges the species tolerates. This is both a correctness measure and what makes state-scale inference affordable.

Three models compete on identical spatially blocked folds: logistic regression, a Random Forest, and gradient boosting (LightGBM). Not a neural network: with hundreds rather than millions of positive examples, a deep model would mostly memorise them, and there is no way to tell that apart from success without honest validation first.

The Random Forest won on held-out geography and calibration, and the simpler logistic baseline was close enough to confirm the signal is real rather than an artefact of model capacity. Selection preferred the simplest model within a hair of the best score.

The output is a probability that a location is worth investigating. It is not a probability that a chestnut is present, and the site does not present it as one.

The first gated release of the general model was trained on 898 grouped chestnut sites — every Research Grade, precisely located, wild Pennsylvania record, with repeat observations of the same tree collapsed into one — against 2,860 tiered negatives: community-confirmed lookalike species, habitat-matched forest a few hundred metres from each chestnut, and randomly sampled forest as the weakest tier.

Every number below is out-of-fold from spatial cross-validation: scored by a model that never saw that site or its 25 km block. Before any statewide inference was allowed, the model had to clear a five-part gate fixed in code in advance — materially above chance on spatial holdouts, ranking precision double the base rate, no regional collapse, and probabilities that beat the base-rate predictor. It cleared all five.

The figures that follow separate two kinds of content deliberately: the metrics table is unbiased validation; the comparison table, seasonal profile and aerial pairs are explanatory illustrations, chosen to be clear rather than average.

What a released candidate looks like: the top-ranked site in the first experimental set is a forest cell on an upper slope in eastern Pennsylvania, 6.8 km from the nearest documented chestnut, scored 0.82 by the general model with habitat and phenology channels agreeing, and its flowering-window 2022 aerial imagery was independently consistent with known chestnut crowns (a weak signal, weighted accordingly). What rejection looks like: three otherwise high-scoring clusters were down-ranked because their leaf-on aerial imagery showed open ground where coarser layers saw forest — the refinement stage doing exactly its job.

Splits are geographic, never random. Observations of the same stand share pixels, weather, terrain and often the same photographer; a random split puts near-copies of the training data into the test set, and the model scores beautifully while finding nothing in the field.

Training, validation and test geographies are separated by county or by spatial block. A final holdout geography is not used for training, feature selection or hyperparameter choice, and is scored once at the end. Its entire value is that nothing about the model was chosen with knowledge of it.

Accuracy is not reported as a headline. Against a landscape of forest holding a few hundred known chestnuts, predicting 'none' everywhere scores above 99% accurate and is worthless. What is tracked is precision at the top of the ranking — precision@10, @50, @100 — alongside PR-AUC, Brier score and a calibration curve, because a stated 85% confidence should mean 85%.

Sentinel-2 cannot identify an individual tree. A candidate is a patch of forest whose seasonal behaviour resembles places chestnuts are known to survive — nothing stronger.

Training data is biased towards where people walk. Roads, trails and popular parks are over-represented among documented observations, and a model fit to them can learn accessibility as much as habitat. This is a known weakness, not a solved one.

Blighted chestnuts usually persist as understory sprouts beneath a closed canopy, effectively invisible from above. The trees this method can plausibly find are the large, canopy-reaching survivors — which are also the rarest and the most valuable.

Mixed pixels, cloud gaps, terrain shadow and year-to-year weather all inject noise. A signal that does not recur across multiple years should not be trusted, and multi-year stability is a planned filter rather than an implemented one.

The question the whole system exists to answer, asked retrospectively: if the trees documented in eastern Pennsylvania were not yet documented, would this pipeline have pointed a volunteer at them? The eastern longitude band — 211 known chestnut groups — was held out entirely, along with anything within 1 km of one of them, and its 839,462 eligible forest cells were re-screened by models that had never seen the band.

At candidate-release depth, no model rediscovered a known tree: zero of 211 recovered in any top-25. A 25-candidate release is a ranked bet on plausible habitat, not a tree detector — which is exactly what this site claims it is, now with a measured number behind the claim.

Deeper in the ranking the models do beat chance: roughly 2× enrichment over a computed random baseline at the top 0.1–0.5%, and the pooled model's top 100 reached 24× the random rate. But at 1% depth both models fell below random. The most plausible reading is the accessibility bias measured from the other side — documented trees sit where people walk, while the model's deep ranking concentrates on remote ridge habitat nobody has surveyed. That is an interpretation offered as such, not an excuse; the numbers stand either way.

Two attempts to improve the deployed model were built, validated and then rejected against criteria fixed in code before any result was seen. Both rejections are reported here because a project that only publishes its successful experiments is not reporting anything.

A six-state regional retrain (v2.0) improved mean ranking accuracy across the new states, and still failed: Pennsylvania precision@100 fell from 0.83 to 0.72, past the pre-registered 0.05 tolerance. The improvement was real and the cost landed exactly where field planning lives, at the top of the ranking.

An ecoregion-adaptive architecture went further — EPA Level III background baselines built from randomly sampled forest, percentile and deviation forms of soil and terrain variables, latitude-residualized phenology, and per-ecoregion calibration. Contextual features did not improve transfer into unseen states (leave-one-state-out ROC 0.725 with context vs 0.733 without), and every pooled variant repeated the same Pennsylvania regression. Neither was promoted; the deployed model is unchanged.

What survived is the framework rather than a bigger model: a configuration-driven onboarding pipeline that ingests a new state, measures how far its environment has drifted from the training domain, tests transfer, and either unlocks inference or publishes an explanation of the failure.

Nothing here is confirmed from a screen. A candidate becomes meaningful only when somebody walks to it and looks, and the metric that matters in the long run is field-confirmed chestnuts divided by field investigations. The workflow below is how a model output becomes — or fails to become — a verified tree.

Approved volunteers claim a candidate before visiting it. A claim reserves the search area, unlocks its exact location for that volunteer only, and expires if unused; it is operational bookkeeping and changes nothing scientific. Candidates on private land stay in the system as scientifically interesting but are not field-trip destinations — nobody is ever pointed onto land they have no right to enter.

A field check is structured observation, not identification. The report asks for photographs — whole tree, leaves, twig and buds, bark, canopy, plus flowers and burs if present — measurements, and simple observations. The visitor is deliberately not asked to decide the species; the photographs carry that burden later, in front of people qualified to judge them.

A report of a possible chestnut moves the candidate to 'possible chestnut — awaiting expert review'. It can never move it to verified: verification is a reviewed decision recorded with its source and date, made by a human examining the evidence, and usually worth an independent confirmation such as an iNaturalist submission or a herbarium sample.

Every outcome feeds the model as a stored label — confirmed chestnuts, confirmed non-chestnuts, and inconclusive visits alike. False positives are the most instructive records in the system and are preserved, never deleted. Nothing retrains automatically: field labels accumulate until a human decides a retraining run is warranted, and any retrained model faces the same spatial-validation gate as its predecessor before it can publish a single candidate.

Confirmed trees are reported to researchers and landowners rather than published. Exact coordinates of unverified candidates and newly confirmed trees are withheld from public maps as a standing rule, not a case-by-case judgement.