Phenology-based, Multi-Scale Classification of Invasive Annual Grasses to the Species Level using Hyperspatial UAV Data

The spread of invasive plant species severely alters wildfire regimes, degrades critical habitat for native species, and has detrimental impacts upon ecosystem function, rangeland productivity, and dynamics of long-term carbon storage. Remote sensing technology has greatly improved our understanding of invasive plant ecology, and hence our ability to manage invasive species. Imagery obtained from airborne or space-borne platforms can provide spatially explicit estimates of plant population size, extent, and spread. However, it has proved quite challenging to remotely detect and monitor weed invasions at the species level, as even the most detailed satellite imagery is commonly greater than one meter in resolution and is too coarse to identify isolated individuals or small patches of invasion. There is a growing need to map the spread of invasive grasses at the species level to facilitate precision management of invasive weeds. Controlling emerging and individual infestations is critical for slowing the rate of invasion and promoting rangeland biodiversity in regions that are potentially at risk.

By capitalizing on species-specific differences in plant phenology and using high resolution Unmanned Aerial Vehicle (UAV) imagery we are able to collect detailed data emphasizing the spectral differences between invasive plants at the species level, even where different species co-occur in a fine-grained mosaic. UAVs can produce images at the centimeter scale, avoiding the 'mixed-pixel problem' where larger pixels encompass multiple cover types and plant species, confounding classification efforts. Pixel-based landcover classifications at this scale frequently contain excessive spatial detail caused by variations in features such as shadows and canopy gaps, often resulting in misclassification, inaccuracy, and a “salt-and-pepper” effect. This study addresses this challenge and refines a novel combination of spectral, textural, contextual, object-based, and multitemporal plant phenology-based classification techniques that employs the full range of available information to differentiate invasive annual plants to the species level. A detailed vegetation classification can be used to train classifications at larger resolutions, identifying invasion patterns at landscape and regional scales. By comparing classifications across spatial resolutions, we can better characterize the landscape context of annual grass invasions. Our approach distinguishes invasive plant species from one another and from the dominant species of native vegetation within which they are embedded, increasing the utility of remote sensing data in invasive species management. Analyzing alternative resolutions contributes to the management of invasive species and deepens our understanding of invaded plant communities at multiple scales.