The above image was captured by the Copernicus satellite, from Leonardo
In a recent study funded by a NASA grant to the Fire Ecology Program, Tall Timbers’ Jenny Rogers, supervised by Kevin Robertson, used cutting edge technology to expand the capacity to accurately assess conservation value from afar.
This study utilizes next-generation “hyperspectral” satellite imagery to map twenty-one different land cover types in parts of Florida and Georgia, including our most imperiled and biodiverse natural community types.
To first visualize how satellite imagery works, think of a standard camera photo.
Digital photos are made up of pixels that combine three basic colors—red, green, and blue—at different numerical intensities.
Multispectral imagery, like LANDSAT and Sentinel, takes this a step further by recording a few additional color bands, ranging from visible blue light to infrared as numerical values. The new satellites that use hyperspectral imagery go even further by capturing over a hundred narrow color bands, providing vastly more detail.
This level of detail is subtle enough to distinguish plants and soils by their chemistry and physical structure. This increased detail allows for increased accuracy.
The hyperspectral satellite used in this study was launched in 2019, and since then, scientists have been exploring how to make the most of the vast amount of data that was previously unavailable.
Through a stimulating collaboration with colleagues at the USGS and San Diego State University, we tested a fresh, new way to extract as much information as possible.
The method removes variability linked to broad patterns from dominant land cover types, like green foliage, soil, and shadows, leaving the finer details, like leaf waxiness or soil moisture, that otherwise might be drowned out. Then, we use a machine learning tool called random forest models (nothing to do with real forests) to classify image pixels as different land cover types.
The results? The new approach produced models with consistently better accuracy (5% on average across all cover types) than using the hyperspectral data directly. The team was able to clearly differentiate between upland, sandhill, and flatwoods pine savannas, and even between native and old field pine savannas in the Red Hills.
These results were recently published in the Journal of Geophysical Research – Biosciences. This research is already proving valuable for expanding Tall Timbers’ map of native, undisturbed groundcover in the region. These findings have led to a new project funded by the USDA’s Natural Resources Conservation Service to use hyperspectral imagery to identify pristine pine savannas throughout the South.
