Choosing a Map of Invasive Annual Grasses

Invasive annual grasses — among them cheatgrass (Bromus tectorum), medusahead (Taeniatherum caput-medusae) and ventenata (Ventenata dubia) — are a serious problem on western rangelands. They change plant communities, disrupt how ecosystems work, reduce forage for wildlife and livestock, and raise the risk of fire. Many regional and national maps of them now exist, and a U.S. Geological Survey guide explains how to choose the right one for the job.

How these maps are made

Most maps start with remotely sensed data — satellite or aerial images taken over time. These are paired with field measurements of the grasses and other data such as elevation to build a model that predicts where the grasses are.

That is hard to do well. Cover swings from year to year with rain and fire; the grasses often make up only a small share of a map pixel; and they are easily confused with other plants. Some models average several years to cut error, but lose information about the current landscape and how it is changing. Models that don't use remote sensing generally map only the potential for invasion.

A: building a map from remote sensing. B: a map showing change over time, with larger error in one area. C: a multiyear map of maximum or average cover — no change over time, but smaller error. USGS figure.

Every map has errors, but accuracy metrics show how far to trust it, and free aerial imagery (such as Google Earth) and local monitoring can reduce uncertainty. The USGS offers three aids for comparing more than 20 products published since 2010: a sortable product database, a compendium of 2-page summaries, and a review article.

Seven things to weigh, in order

Five steps to a final choice

Use considerations A–D to cut the field to five products or fewer, then:

1. Compare the details — how each was built and used (E and F). The database describes more than 40 attributes per product.
2. Compare evaluation results. Scores come from different methods and can themselves be uncertain. A moderate score from a rigorous test on fully independent data may be as good as a high score from a weaker test, and a test that included your region earns more confidence.
3. Look at the map in your management area, overlaid with known grass locations, other vegetation maps and high-resolution imagery. Each source is incomplete, but together they show where the map is more or less reliable.
4. Pair it with local knowledge. Regional maps are general and shouldn't be expected to map small places accurately; field data and monitoring reveal mismatches.
5. Reassess its limits — read the original documentation, contact the developer, or look at how others have used it.

A predicted distribution (left; Dahal and others, 2020) checked against high-resolution aerial imagery (right). Invasive annual grasses appear as light green splotches in early spring that turn grayish green by summer; native vegetation's look varies locally, and some native grasses are hard to tell apart, but shrubs and trees show as individual plants. USGS figure.

Reading accuracy scores

A higher score does not by itself make a better map. What matters is which measure was used and on what data.

Measures differ by output type.

• Continuous maps (percent cover, biomass): R² — the share of variation the model explains (0.340–0.980 among reviewed products); MAE and RMSE — average error, smaller is better (0.87–14.00%); normalized versions, from 0 (perfect) to 1 (no better than chance) (0.130–0.700).
• Categorical maps (presence, suitability, binned cover): percent correctly classified (0.520–0.975) and Cohen's kappa, which adjusts for chance (0.511–0.945), among others. These depend on the probability threshold chosen to call a pixel "present." Some models aim to catch every real infestation, which matters for early detection; others aim to avoid predicting grass where there is none.

Scores are often reported only for the whole map, although accuracy varies locally.

Rigor of the test data, from weakest to strongest:

1. Within-sample — compares the model with the data used to build it; can overstate accuracy.
2. Cross-validation or bootstrapping — refits the model on subsets and predicts the held-out data; less likely to overstate.
3. Fully independent — compares predictions with data never used in building the model; the most rigorous.

Trade-offs: two hypothetical maps

The guide compares a percent-cover map of any invasive annual grass across the western deserts at 30 m, averaged over 2016–2020 and scored within-sample, with a cheatgrass presence-absence map covering the whole West at 250 m, yearly from 1990 to 2016 and cross-validated.

• For early detection and rapid response (such as herbicide treatment), the percent-cover map can point to low-to-moderate infestations to target and supports finer targeting, while the presence map suits planning further surveys.
• For judging wildlife habitat loss, both can work: percent cover reflects degradation at the patch scale, presence at the landscape scale. The broader map can span a species' whole range; the finer resolution matters only for small animals choosing habitat at fine scales; and the time series can reveal long-term trends linked to population change.

In both cases the cross-validated map's accuracy inspires more confidence, but users should check whether either was tested with data from their own region.

Sources

• U.S. Geological Survey: "A user guide to selecting invasive annual grass spatial products for the western United States," Fact Sheet 2022–3001.
• Works cited include D'Antonio and Vitousek (1992); Vaz and others (2018); Sofaer and others (2019); Funk and others (2020); Smith and others (2019); Rocchini and others (2015); Uden and others (2015); He and others (2015); Liu and others (2009); and Dahal and others (2020).