# R1: MTBS Lookout Fire 2023 burn-severity acreage, reproduced with NativeGIS

Draft, not published. Run 2026-10-30. NativeGIS 0.2.0 (nativegis-mcp, Node 24.13.0, win32-x64).

## What was reproduced

**Source:** the Monitoring Trends in Burn Severity (MTBS) program (USGS EROS and USFS GTAC), fire `OR4422212214620230806` (LOOKOUT, Oregon). Ignition 2023-08-06. Extended assessment: pre-fire image 2023-07-31 (Landsat 8), post-fire image 2024-08-26 (Landsat 9).

**Published number:** the "Acreage of Burn Severity" table on the MTBS fire map, `or4422212214620230806_map.pdf`.

**Data:** the MTBS fire bundle, re-hosted by the HJ Andrews LTER at https://andrewsforest.oregonstate.edu/sites/default/files/lter/fsdbdata/gis/LF_or4422212214620230806.zip (8.3 MB, no form needed). It holds the `dnbr`, `dnbr6` and `rdnbr` GeoTIFFs, the `burn_bndy` perimeter shapefile, the metadata and the map PDF. MTBS products are US federal data in the public domain. The local copy is `samples/lookout/x/`.

**Two claims checked:**
1. The published acres in each severity class (a tally).
2. That the published class map follows from the published dNBR raster and the published thresholds (the method). The thresholds are in the perimeter attributes: `IncGreen_T` -150, `Low_T` 74, `Mod_T` 308, `High_T` 573, `dNBR_offst` 0.

## Method (NativeGIS calls; step numbers are those in `run.json`)

1. Open `dnbr6.tif` [2] and `dnbr.tif` [3]. Both are 30 m grids in EPSG:5070 (metres).
   - The perimeter `.prj` is the non-EPSG "USA_Contiguous_Albers_Equal_Area_Conic_USGS_version", and NativeGIS does not map it to an EPSG code [1]. It is opened again with EPSG:5070 declared [4]. Its parameters are identical to EPSG:5070: NAD83/GRS80, standard parallels 29.5 and 45.5, origin 23 N 96 W, no false easting or northing. So it is opened with `epsg: 5070` declared [4]. No reprojection is done.
2. Read the thresholds and the published perimeter acreage from the perimeter attributes [5]. `BurnBndAc` is 25,103.
3. Clip `dnbr6` to the perimeter, keeping cells whose centre is inside it [6]. Polygonize the result (4-connectivity) [7], dissolve by class [8], measure the areas [10] and read them [12].
4. Clip `dnbr` to the perimeter [13]. Reclassify it with the thresholds, first with lower-inclusive bins (`T(k-1) <= dNBR < T(k)`) [14]. Compare cell by cell with the published `dnbr6` [15]. Tabulate the disagreements by transition [19-23].
5. Reclassify again with upper-inclusive bins (`T(k-1) < dNBR <= T(k)`) [21] and compare cell by cell [22].
6. Render the map as SVG over a 10 m USGS 3DEP hillshade, with the HJ Andrews boundary and streams [31], render the same map as a 1600 px PNG for the page preview [33], and export the re-derived class raster (COG) [34] and the class areas (GeoPackage) [35].

Acres are m² / 4,046.8564224 (international acre). Each cell is 900 m².

## Result 1: acreage by class (n = 112,877 cells of 30 m inside the perimeter)

| Class | MTBS map (ac) | NativeGIS cells | NativeGIS m² [10] | NativeGIS ac | Difference |
|---|---:|---:|---:|---:|---:|
| Unburned to low | 4,457 | 20,043 | 18,038,700 | 4,457.46 | 0 |
| Low | 9,309 | 41,858 | 37,672,200 | 9,309.00 | 0 |
| Moderate | 5,576 | 25,073 | 22,565,700 | 5,576.11 | 0 |
| High | 5,750 | 25,855 | 23,269,500 | 5,750.02 | 0 |
| Increased greenness | 11 | 48 | 43,200 | 10.67 | 0 (rounds to 11) |
| Non-processing mask | 0 | 0 | 0 | 0 | 0 |
| **Total** | **25,103** | **112,877** | **101,589,300** | **25,103.26** | **0** |

Every class matches to the whole acre.

**A note on reading the PDF:** text extracted from the map PDF in layout mode puts "11" on the row for the non-processing mask. Extracted in raw order (`pdftotext -raw`), the values list as 4,457 / 9,309 / 5,576 / 5,750 / 11 / 0 for the six classes in table order. So increased greenness is 11 ac and the mask is 0, which agrees with the raster: 48 cells of class 5 and none of class 6.

**Why this matches exactly:** MTBS counts 30 m cells inside the perimeter, and so does this run. NativeGIS's centre-in-polygon clip gives the same 112,877 cells as the published class raster, which has nodata outside the perimeter. The total also equals the perimeter attribute `BurnBndAc` (25,103).

## Result 2: re-deriving the classes from dNBR (n = 112,877 cells)

| Threshold rule | Cells agreeing with MTBS dNBR6 | Agreement |
|---|---:|---:|
| Lower-inclusive, `T(k-1) <= dNBR < T(k)` [14, 15] | 112,453 | 99.624% |
| **Upper-inclusive, `T(k-1) < dNBR <= T(k)`** [21, 22] | **112,877** | **100.000%** |

With the lower-inclusive rule, all 424 differing cells fall one class higher than MTBS [19-23]:

| Lower-inclusive class vs MTBS class | Cells |
|---|---:|
| Low vs Unburned to low | 250 |
| Moderate vs Low | 102 |
| High vs Moderate | 66 |
| Unburned to low vs Increased greenness | 6 |

These are cells whose dNBR equals a threshold exactly (dNBR is stored as integers). With the upper-inclusive rule, the published map is reproduced cell for cell. The `cellsPerRule` counts [21] equal the published class counts exactly.

**Finding:** the MTBS metadata gives the thresholds but not which side of each threshold is closed. This run shows from the data that MTBS assigns a value equal to a threshold to the lower class. Anyone re-deriving MTBS classes with the more common `>=` rule will be off by about 0.4% of cells, all biased one class upward.

## Replay

The run is recorded in `run.json`: 32 steps, paths relative to `NATIVEGIS_DATA`, sha256 9d1f0af75d27b57cab3cec335befcc2aa790988a40e5092de5d642a491bfd91a. It was replayed in a fresh session into `replay-1/`:

- **32 of 32 steps identical.**
- 0 differ and 0 failed.
- Response numbers were compared exactly and output files by sha256 (the GeoPackage by content).

Command for anyone with the data folder:

```
gis_replay {action:"replay", file:".../run.json", inputRoot:"<your NATIVEGIS_DATA>", outputRoot:"<new empty folder>"}
```

## Limits and honesty notes

- **This reproduces a published tally and a published classification rule.** It is a check of open data and arithmetic, not of the science. Whether dNBR thresholds of 74/308/573 are the right severity breaks for Lookout Creek forests is an MTBS analyst judgement that this does not test. Field CBI plots would be needed for that, and none are public for this fire.
- **The MTBS bundle is used as re-hosted by HJA.** The national MTBS download was not fetched to confirm the two copies are byte-identical.
- **The perimeter CRS was declared, not detected.** `epsg:5070` was declared on the strength of identical projection parameters. This is a NativeGIS gap: ESRI "USGS_version" Albers is not recognised as EPSG:5070.
- **`gis_report` could not write the cited report** (`GIS_REPORT_NOTHING`: it only reports point-cloud runs), so this file is written by hand with step citations instead. This is a NativeGIS gap.
- **The exported GeoPackage records EPSG:5070 with an undefined WKT** (export warning). Other GIS software may then not recognise the CRS. This is a NativeGIS gap.
- **Two dissolved class polygons report self-touching rings** (validity warning at step 8). These are cells that touch only at a corner (diagonal pixel contact). Areas are unaffected: they sum exactly to the cell count × 900 m².
