Riverside Fire 2020: Soil Severity and Map Accuracy

What happened in the Riverside Fire?

The Forest Service’s 2026 account dates the Riverside Fire’s start to 8 September 2020, during the wider Pacific Northwest Labor Day fire episode. The BAER catalogue identifies the Oregon event with Mt. Hood National Forest and a listed area of 138,723 acres. That is an attributed catalogue figure, not the new area calculation below. [1] [4]

This record focuses on soil-severity evidence. It makes no original ignition determination, current closure claim or property-by-property loss reconstruction. A regional study covering several fires also remains a regional study when discussed alongside Riverside.

The BAER program describes satellite-derived BARC as an input to field assessment and development of a soil burn severity map. Its workflow helps explain why a classified product is a measurement object with a particular method, rather than a complete reconstruction of the fire. [5]

Original raster audit: Riverside’s mapped soil-severity classes

Tecnobosque calculation, 12 October 2026. We counted the native cells in Riverside_SBS_final.tif, retaining class codes 1–4 and excluding background code 127. The embedded grid is EPSG:26910, with 20 × 20 metre cells. Its metadata identify Sentinel-2 pre/post imagery dated 4 and 29 September 2020 and publication on 13 November 2020. [1]

One cell covers approximately 400 square metres in the projected grid, or 0.04 hectares. The table uses complete native cells, without terrain correction, ownership filtering or a separate perimeter intersection. Rounded class percentages may not sum to exactly 100%.

These are newly computed map statistics. They do not replace the catalogue area, quantify actual terrain-surface area or validate the map’s classes against an independent ground survey. Decimal precision in the arithmetic should not be interpreted as field accuracy.

The file contains another 1,543,618 cells with code 127. In the reader used for this audit, no NoData value is declared and the automatic mask includes the full rectangle. We therefore select the four verified severity classes explicitly, using the supplied palette and legend. The background is not counted as an unburned class.

How does the mapped composition compare with Holiday Farm?

We processed the official Holiday Farm raster using the same native-cell counting method. Its classified footprint contains 1,750,770 cells, with 70,030.80 projected hectares. Each map retains its own denominator; the two fires are not pooled into one population. [2]

Riverside’s high-only share is 12.3%, compared with 9.3% for Holiday Farm. Combining moderate and high reverses that ordering: 52.2% for Riverside and 71.3% for Holiday Farm.

Interpretation: the aggregation is part of the question. High-only area and moderate-plus-high area are different quantities. Neither should silently replace the other in a headline or be described as a universal measure of destruction.

Nor do these proportions explain why the maps differ. Weather, vegetation, terrain, mapped extent, field adjustment and class definitions may matter. This descriptive comparison does not isolate any one factor or estimate a treatment effect.

The same class names do not mean identical thresholds

Both packages identify the same image dates, projected coordinate system, nominal cell size and metadata publication date. Their listed revised SBS thresholds nevertheless differ. The table preserves the source metadata’s boundary values; it does not apply new thresholds to either raster. [1] [2]

Those numbers are classification boundaries in the processing metadata, not temperatures, percentages of soil loss or rainfall thresholds. A repeated label such as “moderate” does not prove that the underlying spectral interval or its ground interpretation is identical between products.

Comparing the supplied class distributions is therefore a description of the products. A common-calibration experiment would require additional inputs and an explicitly justified reclassification method. We have not performed that experiment and do not treat our comparison as its result.

The catalogue labels the products final SBS, while their supplied metadata retain preliminary/draft limitations. Both statements remain attached to the reviewed version; this audit does not assign a new validation status.

What did later field research find about map accuracy?

Busby, Klock and Fried’s 10 October 2023 study compares mapped products with field observations across six Labor Day fires. Its soil-comparison figure uses 130 plots meeting a specified single-forested-condition criterion. About 40% of plots in the field unburned/low soil category were overestimated by BAER-SBS; about 15% of plots with no organic matter cover were underestimated. These are conditional findings for the study’s selected groups. [3]

The paper describes 30 metre products and a 0.81-hectare majority-filter neighborhood for plot comparison. Our native 20 metre file count does not reproduce that processing or plot matching. The paper’s findings cannot simply be assigned as an error rate to every cell in this download. [3]

The denominator matters again. “40% of a specified field category” does not mean 40% of Riverside’s mapped area is wrong. Nor can that proportion be subtracted from the table’s high-severity hectares to create a corrected area. A valid correction would need matching definitions, locations, sampling weights and an appropriate error model.

These research results address agreement with observations. Our file hashes and independent pixel counts address computational reproducibility. Passing one set of checks cannot establish the other property.

The Forest Service’s 2026 Science Findings 278 returns to the research and its implications for interpreting remotely sensed burn maps. It is a later agency communication of the research, not a new independent validation of every Riverside pixel. [4]

How to use this comparison in a stronger investigation

Begin by specifying the decision and the outcome needed. Mapping soil effects, estimating tree mortality and locating erosion hazards involve related but distinct measurements. A color class can help identify a question; it cannot supply every answer about that location.

For a matched map-and-field comparison, retain the product version, grid geometry, class definitions, sample design and observation dates. Specify the footprint used to connect each field observation to the raster. A point extraction and a neighborhood majority rule can describe different local conditions.

Also record intervening changes. Field observations made later may reflect erosion, vegetation recovery or management after the satellite image. Agreement or disagreement requires interpretation in light of that timing, rather than assuming both datasets describe the same moment.

To evaluate a mitigation measure, add its location, prescription, timing and completion status, together with comparable exposure information. Differences between two whole-fire class distributions do not establish a causal effect of a treatment.

Finally, do not treat the large pixel count as millions of independent ground observations. Adjacent cells can share conditions and classification processes. Our descriptive count comes with no new statistical confidence interval for classification accuracy.

Known At: the map, field study and later explanation

Fire DNA and remaining evidence gaps

• Identity: Riverside remains one incident within a multi-fire research setting.
• Mapped effects: native cells are counted with an explicit class mask.
• Calibration: source-specific threshold values stay attached to each product.
• Validation: field agreement is a different test from arithmetic consistency.
• Remaining work: no new matched field validation, ownership subset or causal mitigation analysis.

No severity ranking or TB-ECI score is assigned. The calculated shares do not quantify household damage, current debris-flow risk or water safety. We have not converted a regional research finding into a Riverside-only accuracy rate.

Sources and reproducibility

1. Riverside official BAER product record and SBS package. USDA Forest Service GTAC data distributed through the USGS portal; supplied raster, auxiliary palette, legend and metadata.
2. Holiday Farm official BAER product record and SBS package. Same counting method applied; each product’s thresholds and denominator retained.
3. Busby, Klock and Fried: Inventory analysis of fire effects wrought by wind-driven megafires. Fire Ecology 19:58, 10 October 2023. Methods and map-comparison results, especially Figure 8. DOI: 10.1186/s42408-023-00219-x.
4. Watts, Fried and Busby: Extreme Weather, Forest Structure, and Burn Severity. USDA Forest Service Science Findings 278, 2026. Later agency explanation of the research; not an independent field inventory.
5. BAER imagery support: Background, Products and Applications. Role of satellite BARC and field-adjusted soil products.

The reproduction package retains both unmodified source archives, SHA-256 hashes, reader versions, complete code, numerical output and the original comparison graphic. Rasterio and Pillow decoding counts agree; pixel and projected-area totals reconcile. No resampling, ownership clipping, rethresholding or independent classification validation was performed.

Cite: Tecnobosque, Riverside Fire 2020, record TB-WBB-US-2020-RIVERSIDE, version 1.0, evidence review and raster audit 12 October 2026.