{"id":"a0576a66-885c-4daa-ad8d-bebe0e971cc8","arxiv_id":"2610.32617","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"low","formal_verification":"none","parameter_count":5,"one_line_summary":"Sparse decomposition of annual Earth embeddings yields interpretable features that date fires, forest loss, and landslides to the right year across six continents, and the two leading products retain different phenomena.","lead":"This paper shows that two widely used Earth-observation products, which compress a year of satellite images into short lists of numbers, can be unpacked into meaningful pieces that mark specific events such as fire, forest clearing, and crop type, verified against 3,940 independently mapped events on six continents. The result gives users a practical way to check, before building a monitoring system, whether a product actually contains the phenomenon they care about.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek/deepseek-v4-flash","headline":"The independent-reference premise is load-bearing and unquantified: roughly two-thirds of dated units use Landsat/Sentinel optical composites while embeddings are trained on Sentinel-1/2, and the authors' own caveat leaves no bound on shared-error inflation.","rationale":"The reader's weakest assumption is also the one I would flag: reference-product errors are assumed independent of embedding errors, and the bulk of the validation corpus does not satisfy that assumption in a strong sense. I agree with the reader that this is load-bearing because nearly every headline number—the 85% dated rate, the per-setting bars in Fig. 3b, and the dNBR-based demonstration—compares embedding-derived features against products built from the same optical satellite family. I do not treat this as a new finding: the Discussion states it in one sentence (b171). What is unresolved is magnitude. The paper's claim that 'errors in the reference products mostly reduce agreement' is an assumption, not a demonstrated bound; with shared sensors, correlated errors can instead inflate agreement. The proposed test isolates that magnitude on the largest setting. The code-release and expected-window caveats raised by the reader are real but secondary. Since the reader's CONDITIONAL verdict already rests on this concern, my read does not move the verdict; I would keep CONDITIONAL and make the sensor-stratified revalidation an explicit condition for ACCEPT.","tokens_in":29342,"tokens_out":14762,"duration_ms":132448,"concrete_test":"Re-run the Europe harvest setting (the largest dated setting, n=826) with the identical embedding, dictionary, screening rule and expected window, but replace EFDA labels on a random sample of tiles with harvest-year labels from an independent source that does not use the same optical composites—for example, national forest inventory clearcut records or aerial-photograph/field-based disturbance mapping. If the held-out dating rate stays near 0.832 and far above the 0.301 random-family null, the shared-sensor concern does not land for the bulk of the corpus; if it falls toward the null, the reported rates are inflated by reference-embedding error correlation and the abstract's 'independent reference data' must be qualified.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Load-bearing premise: validation against 'independent reference data' is what turns SAE features into 'records of the land surface.' The Discussion concedes this premise ('This assumes the reference errors are largely independent of the embeddings, which may not hold where both derive from similar satellite data', sentence b171), but the concession is the weak point, not a resolution. The dNBR layer used for screening (Fig. 1d, Supplementary Note S1) is computed from Sentinel-2; EFDA, Hansen GFC and JRC GSW are Landsat/Sentinel optical composites. The units scored against those products—Europe harvest/wind/fire, Amazon/Africa/East Asia forest loss, new permanent water—account for 2,490 of the 3,703 dated units (about 67%). Tessera and AlphaEarth are trained on Sentinel-1/2, so both sides share the same optical reflectance, atmospheric correction, cloud/shadow handling and annual compositing. A feature family could therefore be a detector of the reference products' systematic artifacts (persistent cloud-masked pixels, regrowth lag, compositing gaps) and still pass every held-out dating test in this paper; the temporal hold-out does not remove artifacts that recur in the same locations. The field-checked landslide inventories and agency-derived MTBS/FOR WIND perimeters are more independent, but they are the minority. The authors acknowledge the issue but give no quantitative bound on how much shared-sensor error could inflate the reported 85% rate. That missing bound is the load-bearing concern.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek/deepseek-v4-flash","summary":"This paper asks whether two publicly available global annual Earth embedding products — Tessera (128-d, 10 m) and AlphaEarth (64-d, 10 m), both trained on Sentinel-1/2 — contain individual, identifiable ecological features. The authors decompose each embedding with a TopK sparse autoencoder (m = 1,024, k = 16 for Tessera; m = 512, k = 8 for AlphaEarth) and screen feature families on a subset of reference-mapped units (typically the earliest year) using a disturbed-minus-undisturbed contrast with an in-window threshold, a pre-event-inactivity condition, and a post-event decay condition. Families are scored only on held-out units: a unit validates when the family's mean contrast peaks inside a pre-fixed expected window. Across 16 settings and 3,940 units (3,703 dated) on six continents, all 14 dated settings beat a same-size random-feature null at P ≤ 0.05, with an aggregate held-out dating rate of 85% (per-setting 0.624–1.000 against null medians 0.14–0.46). Complementary results include: a 19-point drop in burn-probe accuracy when the 32-feature burn family is removed, against a null effect for matched random sets; separation of prescribed fire from wildfire, but not of windthrow from harvest; cross-continental transfer of families (US wheat to Europe, California fire to Europe); features that track corn–soy rotation; and no sub-annual timing resolution or early-warning signal.","tokens_in":29698,"tokens_out":27507,"duration_ms":215777,"significance":"If the results hold, this is a substantial, reusable contribution: it answers the open interpretability question for Earth embeddings [1] with SAE decomposition plus temporal external verification, and shows that two products trained on the same observations are not interchangeable for monitoring small, fast-recovering disturbances. Strengths of the design: families are screened on a subset and scored only on held-out units; every dated setting is compared with a same-size random-feature null that is conservative in 12 of 14 settings (Table S3); rates are stable across a 16-fold threshold range (Table S1) and across seeds; the ablation (Fig. 5a) uses matched random controls; and the RADD and precursor tests turn negative findings into bounds on temporal resolution. The main risk is the shared-sensor dependence between the Sentinel-1/2-trained embeddings and the optical-composite reference products scoring about two-thirds of dated units.","major_comments":[{"comment":"The central claim that the recovered concepts are 'validated against independent reference data' (abstract; b2) is load-bearing, and the independence premise is conceded but not quantified in the Discussion (b171: 'This assumes the reference errors are largely independent of the embeddings, which may not hold where both derive from similar satellite data'). Both embedding products are trained on Sentinel-1/2; the screening severity layer (Fig. 1d; Note S1) is computed from Sentinel-2; and EFDA, Hansen GFC and JRC GSW are Landsat/Sentinel optical composites. Scored against those products are Europe harvest/wind/fire (826 + 352 + 151 units), Amazon/Africa/East Asia forest loss (546 + 140 + 317) and new permanent water (158): 2,490 of the 3,703 dated held-out units, about 67%. The temporal hold-out does not control for spatially recurrent artifacts (persistent cloud-masked areas, regrowth lag, compositing gaps), and the assertion that reference errors 'mostly reduce agreement' fails exactly in the correlated case: if both sides share the same optical compositing, agreement can be inflated. The hand-mapped, field-checked landslide inventories — the most sensor-independent references in the corpus — cover 71 dated units, 1.9%; with FOR WIND added, 160 units, about 4%. To be clear, the temporal hold-out and the random-family null are strong controls against screening artifacts; the residual issue is specifically correlated-error inflation, which they do not address. I request one or more of: (i) validation rates reported separately by reference tier; (ii) a cross-reference agreement test on events carrying two references (Vaia has both FOR WIND and EFDA dates); (iii) a rough bound on shared-error inflation, for example by measuring family-reference agreement where independent references disagree. At minimum, the phrase 'independent reference data' in the abstract should be qualified, since most of the dated validation rests on products that share the embeddings' sensor pipeline.","section":"Abstract; §3 Discussion; §4.6; Table S2","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The most novel practical claim in the abstract — that small fires that recover quickly are detectable in one product but not the other — rests on evidence without uncertainty quantification. The southeastern US wildfire comparison uses 21 events (Fig. 6a caption); the signal-to-noise ratios 1.56 (Tessera) and 1.09 (AlphaEarth, at the noise floor) are point estimates with no interval, and a single identical screen yields 15 features in Tessera and none in AlphaEarth over those 21 events. This contrasts with the paper's otherwise systematic use of tile-clustered bootstraps for validation rates (§4.5). Please add tile-clustered bootstrap or permutation intervals for the Fig. 6a ratios, a permutation null for the 15-versus-0 feature count (for example, the number of features that the identical screen returns on random size-matched subsets of the 21 events in AlphaEarth), and intervals for the retention fractions in Fig. 6b. The 240-megafire peak-offset result (216 vs 84 of 240 peaking in the expected window, Fig. 6c) is the most robust piece of the dissociation and would be a more appropriate primary support for the claim than the n = 21 comparison.","section":"§2.6; Fig. 6; §4.7","authors_acknowledge":null,"authors_acknowledge_p":null}],"minor_comments":[{"comment":"For the two settings closest to their nulls, the reported bounds are coarse: SE Australia (observed 0.624, null median 0.369, null 95% interval [0.201, 0.639]) and Europe fire (0.669 vs [0.159, 0.543]). Because the observed SE Australia rate lies within the null's reported 95% interval, please report the exact one-sided P-value or the number of the 200 null draws reaching the observed rate for every dated setting, so the claim 'P ≤ 0.05' is checkable per setting.","section":"Table S2; §4.5","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The abstract's 'date abrupt canopy loss' (b3) is stronger than the operational definition in §4.4, where a held-out unit validates when its family's mean activation contrast is highest inside the expected window, computed with the reference mask and reference year given. The validation therefore establishes that within reference-defined disturbance areas the family contrast follows the reference year; it does not establish that the embeddings localize new events in previously unmapped areas. A sentence in §3 or §4.4 making the mask-conditional nature of the dating metric explicit would prevent the headline from being read as event detection.","section":"§4.4; Abstract","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The 19-point ablation drop is measured on the Dixie tile, where the burn family was screened; the text states this, but the claim that the family is 'how the embedding represents fire' rather than a correlate would be more strongly supported by repeating the ablation on the held-out fires of Fig. 1f (North Complex, Caldor, Mosquito), where the family was not screened.","section":"§2.5; Fig. 5a; Note S10","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Data availability list omits several products used in the main validation: JRC Global Surface Water, the three landslide inventories, the JRC EU crop maps, ESA WorldCover, the Copernicus DEM, and the mining and solar datasets. As written it falls short of the 'All data products used are publicly available' sentence in §4.6; please complete the list.","section":"Data availability; §4.6","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"Presentation points: in Note S3, 'the corpus does not depend on the choice' should say 'the unit corpus', since Table S1 shows the feature families themselves change size with the multiplier. The Fig. 6c caption phrase 'both screened' should read 'screened in both models'. Data availability spells the product 'Geotessera' while §4.1 uses 'GeoTessera'.","section":"§4.3; Fig. 6; Note S3","authors_acknowledge":null,"authors_acknowledge_p":null},{"comment":"The Discussion's coherence argument — that features peaking before fires or separating causes with identical surfaces would have pointed to artifacts — is reasonable corroboration, but it is not an artifact control of the same strength as the random-family nulls and the held-out design; consider presenting it explicitly as corroborative rather than as decisive evidence.","section":"§3 Discussion","authors_acknowledge":null,"authors_acknowledge_p":null}],"recommendation":"major_revision"},"author_rebuttal":null,"desk_editor":{"model":"deepseek/deepseek-v4-flash","letter":"Read it twice. Bottom line: this is a real contribution, and the central claim holds up better than I expected, but the one assumption that worries me is exactly the one the authors flag in the Discussion: reference-product errors are treated as largely independent of the embeddings, and that is left unquantified.\n\nWhat's new: applying sparse autoencoders to annual Earth embeddings and testing the resulting features by when they activate, not just where. That temporal test—contrast must be quiet before the event, peak inside a pre-fixed window, decay afterward—is a genuinely better check that a feature means something than correlating with a static covariate. The validation design is careful: families screened on earliest-year units, scored on later years, never-seen storms, never-seen continents, always against same-size random feature sets from the same dictionary. The nulls are fair, and the sensitivity sweeps over the screening threshold and dictionary seeds are reassuring. The cross-model dissociation on identical events is strong evidence, and it partly offsets the fact that one of the two products is the authors' own. The citation pattern is fine; the external references carry the argument.\n\nSoft spots, in proportion. The main one is the independent-reference premise. Roughly two-thirds of the dated units rely on Landsat/Sentinel optical composites (EFDA, Hansen GFC, JRC GSW), while the embeddings are trained on Sentinel-1/2. The authors concede the issue but do not bound it. A feature family could in principle be tracking recurring artifacts of the shared optical archive—persistent cloud masking, regrowth lag, compositing gaps—and still pass the temporal hold-out, because those artifacts repeat in the same locations. The temporal screen and random nulls rule out chance, not shared systematic error. I don't think this is fatal: event-dated and radar/field references (MTBS dates, FOR WIND, landslide inventories, RADD) behave consistently, and the cross-model comparison uses raw embeddings on identical events rather than reference agreements. But the paper would be much stronger with validation rates reported separately for the subgroups whose references do not share the optical Sentinel/Landsat lineage, or a concrete argument for why shared artifacts cannot reproduce the cross-continent pattern. The sentence \"Errors in the reference products mostly reduce agreement\" is asserted; correlated errors can inflate agreement, so that sentence needs support.\n\nTwo minor things. No code or configurations ship, so the central pipeline is rebuildable from the Methods but not immediate. And the headline 85% absolute rate should be quoted with the caveat that expected windows are pre-fixed and, for detection-dated data, reasonably generous at {Y, Y+1}; the null comparisons remain fair, so this is a presentation issue, not a statistical one.\n\nWho this is for: anyone building monitoring, carbon, or compliance systems on annual Earth embeddings, and the interpretability community. It deserves a serious referee. I would send it out and ask for the quantitative independence check plus code release; my own read is that the claim that annual embeddings are decomposable records of the land surface is supported, with that caveat attached.","headline":"A serious paper with a real contribution: the temporal screening design holds up, but the unquantified shared-satellite-data dependence of the reference products is the spot to push before trusting the 85% headline rate.","tokens_in":30312,"tokens_out":3955,"duration_ms":37434},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek/deepseek-v4-flash","headline":"Annual Earth embeddings are decomposable records of the land surface, this paper argues, with features that name fires, clearings, and crops, date disturbances to the right growing season, and differ between embedding products.","keywords":["Earth embeddings","sparse autoencoders","interpretability","land cover","forest disturbance","crop type","satellite remote sensing","foundation models"],"falsifier":"Re-run the dating test using only reference events mapped from non-satellite sources—field inventories, drone or aerial surveys, camera traps—for the same fires, clearings, and storms. If feature families dated those events at rates near the random-family nulls (0.14 to 0.46) instead of the reported 85%, the conclusion that the embeddings encode the surface itself, rather than agreeing with satellite-derived reference products, would collapse. A second check: inject a synthetic disturbance with known timing into otherwise identical embeddings and verify that the screening rule recovers it at the predicted window.","tokens_in":30600,"feed_emoji":"🛰️","tokens_out":6060,"duration_ms":50515,"temperature":0.7,"pith_summary":"Annual Earth embeddings—learned vectors that compress a year of satellite imagery into a few dozen or hundred numbers for every ten-meter pixel—are now consumed as standard inputs for ecological monitoring, carbon accounting, and compliance, but what they actually contain had not been established. This paper decomposes two public global embedding products with sparse autoencoders and shows that individual directions in the space carry specific, verifiable surface phenomena: fire, forest clearing, crop type, new water, mining land, and solar farms. Feature families screened on a subset of events date 85% of 3,703 held-out disturbances to their expected calendar window, against random-feature baselines of 0.14 to 0.46, in all 14 dated settings on six continents. Removing the burn family from an embedding costs a fire-detection probe 19 accuracy points while a matched random set does nothing, which the authors read as evidence the features are how the embedding represents fire rather than a correlate of it. The two products, trained on largely the same Sentinel observations, retain different phenomena: Tessera captures small fast-recovering fires that AlphaEarth nearly misses, while AlphaEarth separates corn from soy far more cleanly.","feed_headline":"Earth embeddings decompose into datable fire and crop features","feed_subtitle":"Sparse decoding turns annual satellite vectors into verifiable, dated surface records, checked on six continents.","key_machinery":"The TopK sparse autoencoder ($m=1024$ features, $k=16$ active per pixel on Tessera, matched $8\times$ expansion on AlphaEarth) rewrites each annual embedding as a weighted sum of a few non-negative features, giving each feature a near-monomorphic meaning, in contrast to a raw embedding dimension whose two extremes are different classes, such as corn at one tail and developed land at the other. The load-bearing screening rule is the timing contrast: a feature joins a family only if its disturbed-minus-intact activation peaks inside a setting's expected window, stays quiet in every earlier year, and decays from the third year onward. Applying that 'when it appears' test—families screened on a subset of events and scored on the held-out remainder—is what converts a learned dictionary of directions into verifiable records of dated surface events.","core_discovery":"The central claim is that annual Earth embeddings are not opaque blobs but decomposable, interpretable records of the land surface. A sparse autoencoder rewrites each embedding as a weighted sum of a few active dictionary features, and when features are screened by the timing of their disturbed-minus-intact contrast, the survivors behave like named concepts: a burn family appears only after a fire, decays afterwards, transfers without modification to other fires and continents, and—as ablation shows—carries the information a linear probe uses to detect fire. Disturbances are dated to the calendar year containing the first growing season that shows the disturbed surface, which is why western conifer fires peak the year after burning while southeastern understory regrows within weeks and peaks in the event year, and why three European windstorms on either side of a calendar boundary peak in the predicted post-storm growing season. Severity is resolved well enough for a single threshold to separate prescribed fire from wildfire, but cause is not: windthrow, bark-beetle damage, and harvest share most features because they leave the same surface. Tessera and AlphaEarth, observing the same planet, record different phenomena on identical events, consistent with their training objectives treating time differently—one event-indexed, one state-indexed.","pith_inferences":["If the event-indexed versus state-indexed distinction is causal rather than correlational, it predicts a general trade-off in Earth embedding objectives between retaining brief abrupt events and tracking persistent state; training one model family on a fixed Sentinel subsample with varying objectives would test the mechanism the paper could only observe in the two available products.","The screening machinery could be run inverted as an unmapped-event detector: decompose each new annual embedding and flag any feature whose activation appears strongly outside all mapped events, giving an independent audit channel against reference-record omissions.","Because the corn and soy features follow field rotation (82% to 5% and 5% to 83% on CDL-recorded switches), the decomposition is a candidate mechanism for crop mapping from embeddings that tracks the crop in the field year by year, which may matter for small-field agriculture where parcel-level maps have been unreliable.","Feature families screened in one region and transferred to another hint that interpretability built on one product's dictionary could transfer to other products trained on similar data, which would make decomposition-based audits portable rather than per-product efforts."],"forward_implications":["A user can screen any embedding product, open or closed, for a target phenomenon at the required temporal resolution before building a monitoring or compliance system, needing only the embeddings and a reference product for that phenomenon.","Disturbances are placed in the calendar year of the first post-disturbance growing season, not the event year, so downstream models that assume event-year dating will systematically misplace western conifer fires by a year.","The two embedding products are not interchangeable: for small fast-recovering fires, Tessera carries a signal at 1.56 times pre-event noise while AlphaEarth sits near its noise floor at 1.09, whereas AlphaEarth separates corn from soy at 3.19 within-class deviations against Tessera's 1.43.","Benchmarks that score only the tasks they include cannot certify that a phenomenon absent from the benchmark is present in a model, and the two products' divergence on identical events shows that benchmark parity can mask real differences in what an embedding retains.","The embeddings record what a disturbance did to the surface, not what caused it, so causes that leave identical surface signals—windthrow, beetle damage, harvest—cannot be separated by this decomposition."],"supporting_citations":[{"why":"Supplies the TopK sparse autoencoder method that turns each embedding into a few active features, the machinery everything else rests on.","marker":"[42]"},{"why":"Tessera is the primary embedding product analyzed, providing the 128-dimensional annual embeddings on which most settings are tested.","marker":"[8]"},{"why":"AlphaEarth Foundations is the second product, enabling the cross-model comparison and the claim that training objective shapes what an embedding retains.","marker":"[9]"},{"why":"MTBS fire perimeters and ignition dates supply the independent event dates for the western and southeastern US fire settings.","marker":"[29, 30]"},{"why":"Hansen Global Forest Change loss year provides the reference dates for Amazon, Africa, and East Asia clearing settings.","marker":"[35]"},{"why":"European Forest Disturbance Atlas supplies disturbance year and agent labels for the harvest, wind, and beetle settings across 36 countries.","marker":"[32, 33]"},{"why":"USDA Cropland Data Layer gives the class labels for the crop atlas and corn-soy rotation validation.","marker":"[37]"},{"why":"JRC Global Surface Water yearly history dates newly permanent water, supporting the claim that families are not specific to vegetation loss.","marker":"[41]"},{"why":"RADD radar alerts date Amazon clearings to the day, probing whether the embeddings carry any sub-annual timing information.","marker":"[36]"}],"fun_headline_variants":["Earth embeddings decompose into named land-surface events","Satellite vectors split into fire, clearing, and crop features","Annual Earth embeddings are decodable records of disturbance","Embedding decomposition dates fires to the first regrowth season","Sparse decoding turns satellite vectors into dated surface events"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The validation rates assume errors in the reference products are largely independent of the embeddings, but the embeddings and most reference layers derive from the same Sentinel and Landsat optical archives, so shared systematic blind spots could make the true agreement with ground reality lower than the reported figures.","fun_headline_variants_meta":{"raw":{"variants":["Earth embeddings decompose into named land-surface events","Satellite vectors split into fire, clearing, and crop features","Annual Earth embeddings are decodable records of disturbance","Embedding decomposition dates fires to the first regrowth season","Sparse decoding turns satellite vectors into dated surface events"]},"model":"deepseek/deepseek-v4-flash","effort":"low","cost_usd":9.3e-05,"raw_usage":{"cost":9.28564e-05,"is_byok":false,"cost_details":{"upstream_inference_cost":9.28564e-05,"upstream_inference_prompt_cost":2.71544e-05,"upstream_inference_completions_cost":6.5702e-05},"total_tokens":1341,"prompt_tokens":980,"completion_tokens":361,"prompt_tokens_details":{"audio_tokens":0,"video_tokens":0,"cached_tokens":852,"cache_write_tokens":0},"completion_tokens_details":{"audio_tokens":0,"image_tokens":0,"reasoning_tokens":297}},"tokens_in":980,"tokens_out":361,"cost_source":"openrouter_usage","duration_ms":4662,"image_pages":[],"temperature":1.0,"billed_attempts":1,"reasoning_tokens":297,"transport_provider":"openrouter","cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-10-26T07:34:19.376430+00:00","model_set":{"reader":"deepseek/deepseek-v4-flash"},"pipeline_version":"pith-pipeline@v1.4.1-daily-deepseek-or-medium","input_kind":null,"input_chars":null,"input_total_chars":null,"passes_missing":null,"published_standing":null,"published_confidence":null,"objection_audit":null,"run_record":null,"basis":{"kind":"not_recorded","source":null,"chars_given":null,"chars_total":null},"receipt":null,"falsifier":"Re-run the dating test using only reference events mapped from non-satellite sources—field inventories, drone or aerial surveys, camera traps—for the same fires, clearings, and storms. If feature families dated those events at rates near the random-family nulls (0.14 to 0.46) instead of the reported 85%, the conclusion that the embeddings encode the surface itself, rather than agreeing with satellite-derived reference products, would collapse. A second check: inject a synthetic disturbance with known timing into otherwise identical embeddings and verify that the screening rule recovers it at the predicted window.","supporting_citations":[{"cited_title":"Scaling and evaluating sparse autoencoders","cited_arxiv_id":null,"evidence_quote":"Supplies the TopK sparse autoencoder method that turns each embedding into a few active features, the machinery everything else rests on."},{"cited_title":"Lisaius, Markus Immitzer, Toby Jackson, James Ball, David A","cited_arxiv_id":null,"evidence_quote":"Tessera is the primary embedding product analyzed, providing the 128-dimensional annual embeddings on which most settings are tested."},{"cited_title":"AlphaEarth Foundations: An embedding field model for accurate and efficient global mapping from sparse label data","cited_arxiv_id":"2507.22291","evidence_quote":"AlphaEarth Foundations is the second product, enabling the cross-model comparison and the claim that training objective shapes what an embedding retains."},{"cited_title":"High-Resolution Global Maps of 21st-Century Forest Cover Change","cited_arxiv_id":null,"evidence_quote":"Hansen Global Forest Change loss year provides the reference dates for Amazon, Africa, and East Asia clearing settings."},{"cited_title":"Monitoring US agriculture: the US Department of Agriculture, National Agricultural Statistics Service, Cropland Data Layer Program","cited_arxiv_id":null,"evidence_quote":"USDA Cropland Data Layer gives the class labels for the crop atlas and corn-soy rotation validation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"JRC Global Surface Water yearly history dates newly permanent water, supporting the claim that families are not specific to vegetation loss."},{"cited_title":"Weisse, Fred Stolle, Amy Pickens, Gennadii Donchyts, Nicholas Clinton, Noel Gorelick, and Martin Herold","cited_arxiv_id":null,"evidence_quote":"RADD radar alerts date Amazon clearings to the day, probing whether the embeddings carry any sub-annual timing information."}],"lane_accuracy":{"status":"not yet measured","lane":{"model":"deepseek/deepseek-v4-flash","input_kind":null},"url":"https://pith.science/accuracy"},"review_version":1}