Remote sensing-based time-series analysis of cheatgrass (Bromus tectorum L.) phenology.

The western United States is under invasion from cheatgrass (Bromus tectorum L.), an annual grass that alters the pattern of phenology in the ecosystems it infests. This study was conducted to investigate methods for monitoring this invasion. As a result of its annual phenology, cheatgrass is not only an extremely competitive invader, it is also detectible from time series of remotely sensed data. Using the MODerate resolution imaging spectro-radiometer (MODIS) normalized difference vegetation index (NDVI) and spatially interpolated precipitation data, we fit splines to monthly observations to generate time series of NDVI and precipitation from 2001 to 2005 in the state of Utah. We generated a variety of existing metrics of phenology and developed several metrics to describe the relationship between the NDVI and the precipitation time series. These metrics not only describe the pattern of response to precipitation in ecosystems of various infestation levels, but they are predictive of cheatgrass infestation. We tested several popular data mining algorithms to investigate the predictive ability of the time series-based metrics. Our results show that presence-absence can be predicted with 90% accuracy, and four categorical levels of infestation can be predicted with 71% accuracy. The results show that time series-based metrics are effective in prediction of cheatgrass abundance levels, are more effective than metrics based only on NDVI, and provide more information that existing approaches to cheatgrass mapping using phenology. These results are important for designing strategies to monitor ecosystem health over long periods of time at a landscape scale.