Wildfire GIS · Reproducible methods
A severity percentage is only as reproducible as its pixel accounting. This guide explains how to convert a classified raster to area, identify boundary effects and avoid turning class codes into artificial measurements.
Start with the classified product, not its colours
MTBS distributes continuous NBR/dNBR/RdNBR products separately from thematic severity images. Its thematic layer includes severity classes, increased greenness and potentially a mask for problematic features. Metadata accompanies each fire. [1]
Read the actual file’s value definitions. A colour palette can make the same numeric value appear differently in two viewers. Conversely, matching colours do not prove matching class definitions. Save the provider’s class names and distinguish its masked class from the file’s NoData setting.
For this calculation, choose one classified layer, one analysis boundary and one observation period. Do not silently merge a preliminary assessment with a later one. This guide addresses raster accounting, not whether a classification accurately represents ecological conditions.
Convert pixel counts to hectares
GDAL describes an affine geotransform using six coefficients. For a north-up grid, its rotation terms are zero and its horizontal and vertical pixel sizes are separate coefficients. [2]
Geometric derivation: the two pixel-edge vectors are (GT1, GT4) and (GT2, GT5). Their parallelogram area in coordinate units squared is |GT1 × GT5 − GT2 × GT4|. For an unrotated grid it simplifies to |pixel width × pixel height|. This is planar coordinate area; it does not automatically equal undistorted ground area.
Use an appropriate equal-area analysis grid when a constant cell area is required for ground-area reporting. A grid in metres can still have projection distortion. Longitude and latitude are angular units: multiplying a degree-sized pixel’s dimensions does not produce square metres. Record the chosen projection and any transformation.
| Cell dimensions | One cell | 2,000 selected cells |
|---|---|---|
| 30 × 30 m | 900 m² = 0.09 ha | 180 ha |
| 10 × 10 m | 100 m² = 0.01 ha | 20 ha |
| 20 × 30 m | 600 m² = 0.06 ha | 120 ha |
The conversion is 10,000 square metres per hectare. These dimensions are synthetic examples, not claims about every MTBS file. A filename or sensor name is insufficient: inspect the raster’s own grid.
If an existing 30-m grid is subdivided into an aligned 10-m grid with replicated labels, each original cell can become nine output cells. The example’s 2,000 cells become 18,000, but 18,000 × 0.01 remains 180 hectares. Keeping the old 0.09-ha multiplier would incorrectly report 1,620 hectares. Extra cells do not create extra observations.
Choose how boundary cells count
A polygon often crosses raster cells. Three transparent rules answer different accounting questions: select cells whose centres fall inside; select any cell touched by the polygon; or weight cells by the fraction of their footprint inside. None should be left implicit.
| Rule | Included contribution | Area |
|---|---|---|
| Centre rule | 10 selected whole cells | 0.90 ha |
| Any-touch rule | 16 selected whole cells | 1.44 ha |
| Fractional intersection | 10 full cells plus six cells with 25% overlap each | 11.5 cell equivalents = 1.035 ha |
This constructed geometry makes the centre rule select only the ten full cells. The fractional method measures geometric overlap more closely, but still assumes a categorical label can represent each intersected portion. It does not reveal the distribution of real fire effects within a pixel.
Small, narrow or fragmented polygons can have a large boundary contribution relative to their interior. Report the rule and, where it materially affects interpretation, compare plausible alternatives. A difference produced by a clipping rule is not evidence of a change in the fire.
Keep class labels categorical during resampling
GDAL documents nearest-neighbour, average and mode among its resampling methods. Average combines valid contributors by a weighted average; mode selects the most frequent sampled value. It also notes that output-grid alignment does not necessarily align that grid to the input. [3]
Editorial example: invent a four-cell block with labels 1, 1, 4, 4. Its arithmetic mean is 2.5. If a downstream step rounds that to 3, it has created a label absent from all four inputs. Numeric class identifiers are labels, not quantities for interpolation.
Nearest-neighbour keeps a sampled label; mode keeps a selected category. These are category-preserving choices, but they do not guarantee that the total area of every category survives reprojection or aggregation. A coarse cell with mixed inputs must still be represented by a selected class, unless a separate fractional-area product is deliberately created.
For quantitative reporting, document the analysis grid and compare totals before and after any necessary transformation over a comparable footprint. Keep the original input. Use a separate display copy if smoothing is needed for presentation, and never calculate areas from a screenshot or its interpolated colours.
A compact calculation audit
| Field | What to preserve |
|---|---|
| Input identity | Provider, file, version, original class definitions and observation dates |
| Geometry | Coordinate reference system, transform, grid origin, dimensions and boundary version |
| Accounting | Class counts or fractional contributions, mask/NoData treatment and boundary rule |
| Transformation | Reprojection, resampling, target grid and software version |
| Output | Area by class, denominator, excluded area and rounding |
A useful statement is: “Within boundary X, product Y contains Z hectares in its high-severity class, calculated on grid G with boundary rule R. Masked and NoData areas were handled as M.” Provide the underlying totals so readers can reproduce any percentage.
Correct arithmetic cannot establish that a satellite severity class measures soil damage, predicts erosion or proves treatment effectiveness. Those are separate evidence questions. Read the Wildfire Black Box method to keep reported data, derived calculations and interpretations distinct.
From hectares to percentages: define what the result describes
Once class areas are calculated, the next decision is the denominator. A technically correct hectare total can still produce a misleading headline if the reader cannot tell which land or incidents the percentage represents.
| Question | Calculation | Report alongside it |
|---|---|---|
| What share of classified land is high severity? | High-severity area ÷ classified area × 100 | Unclassified area excluded from that denominator |
| How much of the whole boundary is mapped high severity? | Mapped high-severity area ÷ total boundary area × 100 | Missing share; this does not assign a class to missing land |
| What is the pooled share across incidents? | Sum of high-severity areas ÷ sum of corresponding analysis areas × 100 | Consistent denominators and a rule for overlapping footprints |
| What is the mean incident percentage? | Sum of individual incident percentages ÷ number of included incidents | Incident count; each fire receives equal weight regardless of size |
Worked example: 2,400 high-severity hectares in 8,000 classified hectares gives 30%. If the entire boundary is 10,000 hectares, the mapped high-severity contribution is 24% of that boundary. The remaining 2,000 unclassified hectares do not become unburned simply because they are omitted from the first denominator. Follow the full missing-data example and accounting bounds.
When combining fires of different sizes, a mean incident percentage can differ substantially from a pooled area percentage. The simple versus area-weighted comparison shows why the same input table can yield 50% or 26%, and how to label each result correctly.
These are separate analytical choices: pixel accounting establishes the areas; denominator selection and weighting establish the question answered by their summary. Keep all three choices in the calculation record.
Official documentation and method
- MTBS: Product Descriptions — thematic products, index layers and metadata.
- GDAL: Geotransform Tutorial — affine coefficients and north-up grids.
- GDAL: gdalwarp documentation — resampling and target-grid alignment.
Method: documentation review, geometric derivation and independently checked synthetic arithmetic. No incident raster was processed, no new severity map was produced and no field validation was performed. Documentation reviewed 9 October 2026; preserve the software version used in an actual workflow. Revision 1.0.