Mapping Western Rangelands Year by Year, 1985–2021

The Rangeland Condition Monitoring Assessment and Projection project (RCMAP), run by the U.S. Geological Survey with the Bureau of Land Management (BLM), produces yearly maps of what covers the ground across the western United States. Each map gives the percentage cover of a vegetation or ground component, for every year from 1985 to 2021 — 37 years in all.

What it maps

Nine components:

• annual herbaceous, perennial herbaceous and total herbaceous;
• sagebrush, non-sagebrush shrub and total shrub;
• litter;
• bare ground;
• tree canopy.

The full 1985–2021 time series can be downloaded from the Multi-Resolution Land Characteristics Consortium.

Uses. Land managers and scientists can track changes in plant makeup, judge past management and target future improvements, find critical wildlife habitat, study the effects of climate change and year-to-year variation, and assess landscape health and fragmentation.

Healthy sagebrush (Artemisia tridentata) steppe near Granite Mountain, central Wyoming. Photograph by Hua Shi, USGS.

Training data

The model learns from ground observations and from field data scaled up with imagery and machine learning:

Distribution of RCMAP training data. USGS figure.

What the model sees

The training data are related to independent data layers, and those relationships drive the maps:

• Terrain: slope, aspect, position and elevation.
• Landsat composites: two median images a year — leaf on, for peak growth, and leaf off, for dry, brown conditions — with dates set for each mapping region. Pixels with fewer than three clear views, which tend to have odd dates or cloud, shadow or snow, are cleaned using synthetic images.
• Synthetic imagery: six months a year of images from Continuous Change Detection and Classification (CCDC), modeled from each pixel's history, weighted toward the early growing season to sharpen differences between components.
• Indices: water, built-up and soil-adjusted vegetation indices, plus tassel cap greenness, wetness and brightness, for both leaf-on and leaf-off composites.

RCMAP mapping regions, each with its own imagery dates — from Mediterranean California to the Northern Great Plains and the Mojave, Sonoran and Chihuahuan Deserts. USGS figure.

The model

Each mapping region has one neural network that predicts all components: 4 layers deep, 128 neurons wide, with 20 percent dropout between layers, tuned with KerasTuner. Compared with earlier RCMAP versions, which used Cubist regression-tree models, error fell by 5–7 percent with all else equal. Post-processing rules then limit noise and capture cover after fires, using fire-recovery equations grouped by ecosystem resistance and resilience classes.

How accurate is it?

The maps were checked against field data not used in training, where R² measures how much of the field variation the maps explain (precision) and RMSE the expected difference between field and mapped values (accuracy):

Against the BLM plots, tree cover had an R² of 0.66. These are strict, single-pixel comparisons; work at broader scales, or averaged over management units, would tend to show lower error.

Caveats

• Synthetic images come from models fit to clear Landsat views, so their quality depends on how many clear views there are, models near the end of the series tend to be simpler, and changes can show up with a lag. They are not treated as real extra observations, only as a record of seasonal patterns.
• The approach tends to underestimate change between periods rather than overestimate it.
• Most training data are themselves model predictions from high-resolution imagery, and carry error: average correlation of 0.90–0.97 and absolute error of 1.81–4.65 percent by component. Tree cover training came from a convolutional neural network rather than Cubist. The other training sources contain error too.
• The maps aim to show peak growing-season conditions, but image availability and the timing of field observations mean they may not always do so.

The authors advise comparing the largest units an analysis allows — two pastures' worth of pixels, say, rather than two single pixels — to keep error down.

Sagebrush flat in Grand Teton National Park, Wyoming. Photograph by Hua Shi, USGS.

Sources

• U.S. Geological Survey: "Rangeland Condition Monitoring Assessment and Projection, 1985–2021," Fact Sheet 2023–3004.
• Works cited by the fact sheet include Shi and others (2022) and Rigge and others (2020, 2021, 2022) on RCMAP; Herrick and others (2017); LANDFIRE (2022); Zhu and others (2015) on CCDC; and Maestas and others (2016).