Published results you can rerun.

A published number is only as good as anyone's ability to get it again. Here, published results from the HJ Andrews Experimental Forest in Oregon are rerun from their open data with NativeGIS. Each page puts our numbers next to the published ones, says why any of them differ, and gives you the run file to repeat every step yourself.

How we did

20 studies. 17 check published numbers: 123 numbers, 39 exact, 49 within tolerance, 35 differ (each explained on its study page by the published data or wording; none traces to NativeGIS). 3 include a benchmark (R2, R3, R19) and 4 include a new analysis (R8, R19, R20, R6). A study can do more than one: R3, R19, R6 are in more than one group.

39 exact49 within tolerance35 differ, explained

By study (R3, R5, R11, R12, R19, R22, R6 have two checks with different verdicts):Exact 3Close 9Wording 1Matches except 10Benchmark 3New analysis 4

What the verdicts mean
Exact
Every number matches the published one to its printed precision.
Close
Every number is within the stated tolerance.
Wording
Measured right, but the published words describe a different quantity.
Matches except
Matches, except a number whose difference is explained on the study page.
Benchmark
No published answer to match: NativeGIS is scored against a published reference.
New analysis
A new question asked of published data: there is no paper number to match.

Published figures are read from each study's published record and NativeGIS figures from its recorded run; the verdicts are computed from the printed precision and the stated tolerance each time the pages are built. A number counts as exact when ours, rounded the way the paper rounds, equals the printed figure.

What reproducing caught: the small discrepancies these reruns found in published numbers and datasets, with the evidence and how sure we are.

20 studies of 3 kinds: Fire and burn severity (8) · Forest structure from LiDAR (3) · Land use, roads and landslides (9). Browse them as demos on the demos page.

Fire and burn severity

Where a fire burned, how hard, how much of a forest it took, and how that compares with the fires before it.

R1 Exact

Lookout Fire 2023 burn severity, acre for acre

Every severity class on the published MTBS acreage table matches to the acre, and the published class map follows cell for cell from the published dNBR and thresholds.
25,103 of 25,103 ac; 6/6 classes exact; 100.0% of cells re-derived · verified 2026-10-31, 32/32 steps identical
Open the study →
R9 Matches except

How much of the Andrews Forest burned in 2023?

HJ Andrews says the Lookout Fire burned about 70% of the forest. That matches the area the fire covered: 68% of the forest lies inside the incident's final perimeter and inside the BAER soil burn severity map, the two footprints that existed when the figure was first given, and 67% inside the later MTBS perimeter. The ground that actually burned is less: 57% at low, moderate or high severity, about a fifth at high severity.
67%–68% inside the fire's mapped footprint (published ≈ 70%); 57% actually burned; 19% high severity · verified 2026-10-31, 38/38 steps identical
Open the study →
R8 New analysis

Five centuries of fire, then the Lookout Fire

The HJ Andrews fire-history map records 31 fires between 1532 and 1952. Counting every 30 m cell once, ground whose last recorded fire was in the 1500s burned at high severity in the 2023 Lookout Fire at 8–9%, against 28.5% overall. That holds however the fire-history layer is aligned. An earlier claim that ground with no recorded fire after 1482 burned hottest did not hold, and is withdrawn.
per cell: ground last burned in the 1500s 9% high (8–9% aligned) vs 28.5% overall · verified 2026-10-31, 1124/1124 steps identical
Open the study →
R11 Close Wording

Holiday Farm Fire 2020, and 70% of the Andrews Forest burned since 2020

The MTBS acreage table for the 2020 Holiday Farm Fire comes back from the published rasters with the same total, three of six classes to the acre, high severity 1 acre off and two classes 3 acres apart. HJ Andrews says fires since 2020 burned about 70% of the forest. That matches the area the two fires covered: 69.7% to 70.5% of the forest, by MTBS, BAER and NIFC footprints alike. The ground that actually burned is less, 58%. The two fires do not overlap inside the forest.
174,390 of 174,390 ac, 3/6 classes exact; 69.7%–70.5% of the forest inside the 2020 or 2023 fire footprint (published ≈ 70%); 58% actually burned · verified 2026-10-31, 103/103 steps identical
Open the study →
R12 Matches except Close

How badly did the 2020 Holiday Farm Fire burn the soil in Watersheds 1 and 9?

Watershed 9 and the whole-fire shares match the paper to the percent. Watershed 1 matches to within half a percent, except the 8% the paper lists as high severity. That 8% equals the part of the watershed BAER did not map, mostly outside the fire. It is exactly what reading the raster's no-data value as high severity gives: most likely a coding slip.
WS9 2/88/10/0% and fire 24/62/9% as published; WS1's 8% "high" is unmapped area · verified 2026-10-31, 64/64 steps identical
Open the study →
R13 Exact

How much basal area did the 2023 Lookout Fire take?

The paper's 54 / 20 / 12 / 14% split comes out exactly from the RAVG map it cites, but only from that map's four-class product, whose classes are 0%, under 25%, 25–75% and 75–100% loss. Grouped into the 0–25 / 25–50 / 50–75 / 75–100% classes the paper names, the same map gives about 74 / 6 / 6 / 14%, which is RAVG's own table.
BA4 codes 54/20/12/14% = published 54/20/12/14%; the paper's stated classes give 74/6/6/14% · verified 2026-10-31, 54/54 steps identical
Open the study →
R17 Close

RAVG burn severity for the 2023 Lookout Fire, re-derived from its Sentinel-2 images

A burn-severity map is only as trustworthy as the steps behind it. Starting from the two Sentinel-2 images in RAVG's own download, NativeGIS rebuilds NBR, dNBR, RdNBR and the basal-area loss model, and puts 99.88% of the fire's 117,334 cells in the same loss class as RAVG. The class shares come out at 54.2 / 19.5 / 11.9 / 14.4% against RAVG's 54.1 / 19.6 / 11.8 / 14.4%.
99.88% of cells in RAVG's class; shares 54/20/12/14% (RAVG 54/20/12/14%) · verified 2026-10-31, 49/49 steps identical
Open the study →
R20 New analysis

What the Carpenter Mountain fire lookout can see: a viewshed to Mount Hood

A fire lookout is only as useful as the ground it overlooks. A viewshed from the Carpenter Mountain lookout on a 30 m terrain model, with the earth's curvature, sees all seven Cascade summits it is said to see, Mount Hood 126 km away included, and 37.7% of the HJ Andrews Experimental Forest below it.
7/7 summits in line of sight (Mount Hood at 126.4 km); 37.7% of HJ Andrews seen from a 10 m cab · verified 2026-10-31, 41/41 steps identical
Open the study →

Forest structure from LiDAR

Tree tops, heights and canopy from airborne LiDAR point clouds, checked against published maps of the same trees.

R2 Benchmark

The HJ Andrews LiDAR stem map, benchmarked

The published file holds exactly the 1,328,242 trees it reports. On one 750 m tile of the same flight, NativeGIS finds the tall trees the stem map finds, to within a few tenths of a metre of height. Where the two disagree is the 10 to 20 m trees.
1,328,242 trees confirmed; F 0.75 / 0.83 on one tile; heights within 0.2–0.4 m · verified 2026-10-31, 18/18 steps identical
Open the study →
R3 Close Benchmark

The HJ Andrews biomass map (TV080), rebuilt from the LiDAR with the published equation

Seven LiDAR tiles run through the published equation, with nothing fitted and on TV080's own 5 m cells, come back at TV080's mean to within 1% (632.5 against 637.4 Mg/ha over 134,255 cells) and match it cell by cell: r = 0.9998, RMSE 12.1 Mg/ha, 99.8% of cells within 10%. What is left is a small, steady shortfall of about 5 Mg/ha.
mean 632.5 vs 637.4 Mg/ha; r 0.9998; RMSE 12.1 Mg/ha per 5 m cell · verified 2026-10-31, 155/155 steps identical
Open the study →
R5 Close Matches except

Canopy-height profiles and harvest-edge distance bins in HJ Andrews' unharvested forest (Bell, Spies & Pabst 2017)

From open HJ Andrews layers and all 122 LiDAR tiles over the study area, 38,333 pixels of unharvested forest fall in the paper's three harvest-edge distance bins (published 37,900), split 11.2%, 10.7% and 78.1% (published 10.6%, 11.0% and 78.4%). The canopy-height profiles come from NativeGIS; the PCA on them, computed outside NativeGIS, puts 91.5% of the variance on four axes on the covariance matrix and 72.4% on the correlation matrix, either side of the published 85.2%.
PCA 4 axes 91.5% (pub. 85.2%); within 75 m of a harvest 21.9% (pub. 20%); 38,333 pixels (pub. 37,900) · verified 2026-10-31, 659/659 steps identical
Open the study →

Land use, roads and landslides

How much of a basin was logged and roaded, and where the landslides came from.

R10 Matches except

Roads and clearcuts in Lookout Creek, 1950–1995

From the open HJ Andrews layers, 13.3% of the Lookout Creek basin was cut by 1964 (published 13%), it holds 122 km of road (published 119 km), and the road right of way covers 3.9% (published 3.7%); the road construction history puts 89.9 km of road in place by 1964 (published 94 km). The harvest total comes out 2 points lower, the published 71.3% native forest is the paper's own subtraction, and the slide counts differ because only an older inventory is open.
13.3% cut by 1964 (pub. 13%); 122 km roads (pub. 119); right of way 3.9% (pub. 3.7%); slides from an older inventory · verified 2026-10-31, 69/69 steps identical
Open the study →
R14 Close

HJ Andrews gauged watersheds: areas and elevations, delineated from a terrain model

Ten gauged watersheds traced from a public terrain model and the gauges' GPS positions come back at the published areas to within 4% (Lookout Creek 6,240 ha against 6,242), and at the published gauge elevations to within 4 m.
8/10 within 2% of the published area; Lookout 6,240 ha (pub. 6,242); gauges within 4 m · verified 2026-10-31, 30/30 steps identical
Open the study →
R15 Matches except

Where the 1964 storm moved the soil: roads, logging and tuff

After the December 1964 storm, Dyrness tied 72% of the Andrews' mass movements to roads and 17% to logged areas, and put 94% of them on tuffs and breccias that cover 37% of the forest. From the open slide inventory and geology map, the road share comes out 71.7%, logged areas 17.4% and the tuff and breccia share of the area 37.3%. The share of movements on tuff and breccia comes out lower, 89.1%, on the open geology map.
roads 71.7% (pub. 72%); logged 17.4% (pub. 17%); tuff area 37.3% (pub. 37%); on tuff 89.1% (pub. 94%) · verified 2026-10-31, 49/49 steps identical
Open the study →
R16 Matches except

Debris slides after the February 1996 flood

The February 1996 flood started 36 debris slides larger than 75 m³ in the Andrews, against 72 after the 1964 flood. The open inventory gives exactly 36 for 1996, with the same split between roads, plantations and natural forest to within one slide. For 1964 the 2022-checked inventory gives 70 slides dated to the 1964–65 storms, against 72. 21 of the 36 slides of 1996 are flagged as becoming debris flows, against 16 published, and the open data does not resolve that difference.
36 slides in 1996 (pub. 36); 21 became debris flows (pub. 16); 70 slides in 1964 (pub. 72) · verified 2026-10-31, 81/81 steps identical
Open the study →
R18 Matches except

Where the Andrews holds its soil: the stable zone, clearcuts and roads

Swanson and Dyrness split the Andrews into a stable zone of 33.4 km² on lava above a contact at 900 to 1,000 m, with only two small road slides from 1950 to 1974, and an unstable zone with 139 slides, where clearcuts lost 6,130 m³/km² to landslides, 2.8 times forest, and road rights of way 30 times. With the 2022-checked open slide inventory, the stable zone is 53.4% of the forest with the contact at 900 m and 44.9% at 1,000 m, against 52.0% published; it holds exactly 2 small road slides at 1,000 m and 5 at 900 m. Over the paper's own areas the clearcut rate comes out 6,356 m³/km² and the road rate 68,669; the forest rate comes out lower, so the ratios to forest come out higher than published.
stable zone 53.4% at 900 m, 44.9% at 1,000 m (pub. 52.0%); 2 / 5 road slides in it (pub. 2); clearcut 6,356 m³/km² on the paper's area (pub. 6,130); ×4.4 / ×40 forest on our areas (pub. 2.8 / 30) · verified 2026-10-31, 120/120 steps identical
Open the study →
R19 Close Benchmark New analysis

The Lookout Creek stream network, derived from a terrain model and checked against HJ Andrews' stream maps

A stream network traced from a public 5 m terrain model makes Lookout Creek a 5th-order basin, as both HJ Andrews stream maps do, when channels start at 5 to 10 ha of contributing area. At the 5 ha threshold, 94% of the 2008 LiDAR map's channel length lies within 30 m of a derived channel.
5th order at 5–10 ha (pub. 5th); 93.6% of the 2008 layer within 30 m of a derived channel; density 3.53 at 5 ha against 3.27 · verified 2026-10-31, 53/53 steps identical
Open the study →
R21 Matches except

How much of the Andrews is old growth? Checking the 40% against two open maps

HJ Andrews says about 40% of the forest was old growth before the 2020 fires. Neither open map gives 40% inside the forest boundary: the 1988 Landsat stand-age map puts 46.1% in its old-conifer class, close to the 45% a 1984 introduction gave for trees over 400 years, while the Forest Service stand-origin map puts 30.0% in stands that began before 1600 and 59.3% in stands that began before 1800. Counted as 400 years or more by 2023 instead of by 2000, the same map gives 49.6%, so 40% falls between the two reference years.
old conifer 1988 46.1%, origin before 1600 30.0%, 1623 or earlier 49.6% (published: 40%, 1984: about 45%) · verified 2026-10-31, 170/170 steps identical
Open the study →
R22 Matches except Close

Rain gauges against an elevation-aware climate surface: how much a plain interpolation misses in the HJ Andrews mountains

Against the PRISM grid for 1971-2000, the 13 long-record HJ Andrews gauges agree to 0.5% on average and a distance-only interpolation of them misses only 9 mm (0.4%) over the forest, though 181 mm (7.1%) above 1,300 m. The first PRISM grid, for 1980-89, puts 9.2% more rain on the forest and rises 1.10 mm per metre of elevation against 0.61: most of the 20% storage-gauge shortfall an earlier version of this page reported came from that grid.
13 gauges +0.5% vs PRISM 1971–2000 (2,297 mm/yr forest mean); IDW −181 mm above 1,300 m; first grid 2,509 mm · verified 2026-10-31, 183/183 steps identical
Open the study →
R6 Close Matches except Exact New analysis

Ward et al.'s Lookout Creek river corridor network, derived again from a terrain model and checked against their published geometry

Channels traced from a public 4 m terrain model at Ward et al.'s published 5 ha threshold give 689 links and a 5th-order Lookout Creek, against their 686 segments and 5th order. 90% of their published stream points lie within 30 m of a derived channel, and 99% and more on 3rd to 5th order streams.
689 links against 686 published segments, both 5th order; 89.9% of the published points within 30 m of a derived channel · verified 2026-10-31, 56/56 steps identical
Open the study →

What comes next

More HJ Andrews studies are queued. Each waits on something named, rather than being approximated:

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