What reproducing caught.

Reproducing a study means running its open data again and putting the result next to the published number. Most of the time they agree; that is the point. Sometimes they do not, and the difference is usually small: a table row that does not add up, a label that does not match the product, a dataset that sits a little off. Good papers have small slips, and catching them is what reproduction is for. Each item below was found by rerunning the open data with NativeGIS, links to the study page with the full detail and the replayable run, and says how sure we are. We have not contacted the authors or data providers about any of these yet.

In the published papers and statements (7)

R10 · Goodman et al. 2023: roads, clearcuts and debris slides in the Lookout Creek basin Confirmed

Road length: 119 km in the text, 126 km in the paper's own table

Where
Goodman, Segura, Jones & Swanson 2023, Earth Surface Processes and Landforms 48:1103. §2 (reprint p. 2) and Table 2 (p. 6).
Published
"By 1995 there were 119 km of roads" (§2). Table 2 gives road built per period: 94, 11 and 21 km.
What the data gives
94 + 11 + 21 = 126 km, 7 km more than the text. The open road layer holds 122.0 km inside the Lookout Creek watershed.
Suggested correction
Either the text or Table 2 needs a note on which road set each counts.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R10 · Goodman et al. 2023: roads, clearcuts and debris slides in the Lookout Creek basin Likely

Land-use shares that count the road right of way twice

Where
Goodman et al. 2023, §2 (reprint p. 2): 25% clearcut, roads 3.7% of the area, native forest 71.3%.
Published
25% clearcut + 3.7% road right of way + 71.3% native forest = 100%.
What the data gives
The road right of way covers 3.9% of the basin, but 2.0% of it lies inside clearcut units. So 25% and 3.7% overlap, and 71.3% (= 100 − 25 − 3.7) subtracts the overlap twice. Counting each square metre once gives 22.9% harvest units, 1.9% right of way outside them and 75.2% native forest.
Suggested correction
Native forest about 73% on the paper's own 25% and 3.7%, once the overlap (about 2 points) is counted once; 75.2% on the open layers.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R12 · Holiday Farm Fire soil burn severity in HJ Andrews WS1 and WS9 (Bush et al. 2024, Table 1) Likely

Watershed 1's 8% high severity is ground the severity map does not cover

Where
Bush, Johnson, Bladon & Sullivan 2024, Hydrological Processes 38:e15231. Table 1(a)(i).
Published
Watershed 1: 48% unburned or very low, 42% low, 1% moderate, 8% high soil burn severity.
What the data gives
No cell of Watershed 1 has the BAER high class. 8.1% of it is unmapped, and those cells hold the value 15. Reading 15 as "4 or more = high" gives 8.1% high, the published 8%, and leaves Watershed 9 exactly as published. The unmapped 7.9 ha is almost all outside the fire perimeter (only 2.7% inside), and MTBS shows no high severity there.
Suggested correction
Watershed 1 high severity 0%, with about 8% of the watershed outside the mapped burn area. The paper's comparison of high severity between WS1 and WS9 (8% against 0%) would then not hold.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R13 · RAVG basal-area loss over the 2023 Lookout Fire (Dickinson et al. 2026) Likely

RAVG basal-area classes labelled with the map legend's bounds, not the product's

Where
Dickinson et al. 2026, Fire Ecology 22:114, Methods (study area): "54 / 20 / 12 / 14%" in classes 0–25, 25–50, 50–75 and 75–100% basal-area loss.
Published
54% of the fire footprint in 0–25% basal-area loss, 20% in 25–50%, 12% in 50–75% and 14% in 75–100%.
What the data gives
The four numbers match RAVG's four-class product exactly (54.1, 19.6, 11.8, 14.4%). RAVG defines those classes as 0%, more than 0 to under 25%, 25 to under 75% and 75–100% loss. In the bounds the paper names, the same map gives 73.7, 6.2, 5.6 and 14.4%.
Suggested correction
Relabel the classes 0%, >0–25%, 25–75% and 75–100% loss; or keep the labels and use 74 / 6 / 6 / 14%.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R13 · RAVG basal-area loss over the 2023 Lookout Fire (Dickinson et al. 2026) Confirmed

Image source named two ways

Where
Dickinson et al. 2026: the Fig. 1 caption names Sentinel images (2022-10-19 and 2023-10-19); the Methods name Landsat.
Published
Fig. 1: Sentinel. Methods: Landsat.
What the data gives
The RAVG bundle for the Lookout Fire is built from the Sentinel pair the caption names.
Suggested correction
The Methods should read Sentinel-2, as the figure caption and the RAVG product do.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R11 · MTBS Holiday Farm Fire 2020 acreage, and the share of HJ Andrews burned since 2020 Confirmed

"Burned" meaning "inside a fire perimeter"

Where
HJ Andrews Experimental Forest, facts page: "Recent fires since 2020 have burned ca. 70% of the landscape".
Published
About 70% of the forest burned in fires since 2020.
What the data gives
69.7% of the forest lies inside the 2020 or 2023 fire perimeters, which matches. But only 58.4% burned at low, moderate or high severity; the rest is unburned islands inside the perimeters.
Suggested correction
"Fires since 2020 burned within about 70% of the forest; about 58% actually burned."

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R4 · Zald et al. 2016, whole-forest carbon Open

Aboveground live carbon, 302 ± 201 Mg/ha: not reproducible as printed

Where
Zald, Spies, Harmon & Twery 2016, Forest Ecology and Management 366:193. Table 2 and §3; open copy from the Forest Service research library.
Published
A LiDAR model, cube root of carbon = 1.2496240 + 0.0219194 × H95 − 0.0000049 × INDEX3 (H95: 95th-percentile first-return height), mapped at 25 m over the forest, giving a mean of 302 ± 201 Mg/ha.
What we found
Read as printed, the model cannot reach the published mean. The INDEX3 term only lowers the result, so the most it can give is (1.2496 + 0.0219 × H95)³: 18.7 Mg C/ha at a 64 m canopy, or 200 Mg/ha even if heights were in feet. Under the 400× per-cell scale another HJ Andrews biomass map uses, 302 Mg/ha would need INDEX3 near 250,000. The map is also said to hold 213,570 cells of 25 m (13,348 ha), about twice the forest's 6,369 ha.
What is missing
The scale of the modelled response, the definition and units of INDEX3 (a point-density index), and the area the 302 ± 201 was taken over. Any one of these, or the authors' 25 m map, would let it be run. Choosing them so that the mean comes out at 302 would be fitting to the answer, so the study was not built.

Checked from the paper itself, before any data was run: there is no study page because the result could not be reproduced as printed. We have not contacted the authors.

In the datasets (3)

R8 · HJ Andrews fire history (1532–1952) against 2023 Lookout Fire severity Confirmed

HJ Andrews fire-history map (DF019) sits up to about 200 m south of the forest

Where
HJ Andrews dataset DF019, firehistory_region.shp (knb-lter-and.3990.5).
Published
Fire-history polygons in NAD83 UTM zone 10N, converted from NAD27 in 2009 according to the file's processing history.
What the data gives
Along their shared ridgelines the layer's outer edge sits about 200 m south of both the HJ Andrews boundary and the 2018 LiDAR-based Lookout watershed, which agree with each other to 11 m. Registered on streams and summits only, Teensma's own 1987 maps put his study-area outline within about one registration error of the boundary, and DF019 182.6 m south of it on Lookout Ridge. At the north edge the two agree. The gap grows about 20.8 m per km southward, so DF019 is stretched about 2% north to south, not shifted; a NAD27/NAD83 mislabel would be a uniform shift with an east–west part. The 2004 preview already shows DF019 157.2 m south of the boundary, before the 2009 datum conversion. The most likely source is the 1997 step in which the episodes were "manuscripted onto HJA base maps and digitized".
Suggested correction
Re-register DF019 to Teensma's base maps (a north–south scale of about 0.979 about northing 4,903 km matches them), or note the offset in the metadata. The R8 study tests its result at three uniform offsets.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R8 · HJ Andrews fire history (1532–1952) against 2023 Lookout Fire severity Likely

A study-area outline stored as a fire

Where
DF019 metadata: "AGE = 0 means that a fire occurred in the polygon".
Published
One AGE = 0 polygon with EPISODE 1482, spanning the whole reconstruction area.
What the data gives
Its stair-stepped edge traces the study area, not a burn: it covers all of the forest but 71 ha. Treated as a fire, it makes every cell look burned at least once since 1482.
Suggested correction
Mark the 1482 polygon as the reconstruction's base or study area, not a fire episode.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

R1 · MTBS Lookout Fire 2023 burn-severity acreage Confirmed

MTBS thresholds published without saying which side a boundary value falls

Where
MTBS fire bundles for the Lookout (2023) and Holiday Farm (2020) fires: perimeter attributes IncGreen_T, Low_T, Mod_T, High_T.
Published
The four dNBR thresholds for each fire, with no rule for values exactly on a threshold.
What the data gives
Re-deriving the class map with the usual rule (a value on a threshold goes up) disagrees in 0.38% of cells; putting it in the lower class reproduces the published map exactly, for both fires.
Suggested correction
State in the metadata that a dNBR value equal to a threshold belongs to the lower class.

Found by rerunning the open data with NativeGIS. We have not contacted the authors or data providers yet. Full detail and the replayable run →

NativeGIS reproductions · every NativeGIS number on this page is read from the recorded run file it links to.