Using Hyperspectral Imagery to Characterize Rangeland Vegetation Composition at Process-Relevant Scales

Abstract

Rangelands are composed of patchy, highly dynamic herbaceous plant communities that are difficult to quantify across broad spatial extents at resolutions relevant to their characteristic spatial scales. Furthermore, differentiation of these plant communities using remotely sensed observations is complicated by their similar spectral absorption profiles. To better quantify the impacts of land management and weather variability on rangeland vegetation change, we analyzed high resolution hyperspectral data produced by the National Ecological Observatory Network (NEON) at a 6500-ha experimental station (Central Plains Experimental Range) to map vegetation composition and change over a 5-year timescale. The spatial resolution (1 m) of the data was able to resolve the plant community type at a suitable scale and the information-rich spectral resolution (426 bands) was able to differentiate closely related plant community classes. The resulting plant community class map showed strong accuracy results from both formal quantitative measurements (F1 75% and Kappa 0.83) and informal qualitative assessments. Over a 5-year period, we found that plant community composition was impacted more strongly by weather than by the rangeland management regime. Our work displays the potential to map plant community classes across extensive areas of herbaceous vegetation and use resultant maps to inform rangeland ecology and management. Critical to the success of the research was the development of computational methods that allowed us to implement efficient and flexible analyses on the large and complex data.

Keywords

• plant community composition
• hyperspectral
• grasslands
• HPC computing
• NEON AOP
• machine learning