A Multi-Scale Approach to Predict the Fractional Cover of Medusahead (Taeniatherum Caput-Medusae)

Medusahead is an aggressive, winter annual that is of dire concern for the health and sustainability of western rangelands in the United States. Medusahead reduces plant diversity, alters ecosystem function, and reduces carrying capacities for both livestock and wildlife. The species has competitive advantages over cheatgrass and native grasses that causes an increased amount of fine fuels deposited on western rangelands. The Channeled Scablands of eastern Washington in the United States represent a typical example of a region being challenged by the expansion of this weed. The costs of the invasion are high and financial constraints can limit successful management. Managers need the ability to identify medusahead across entire landscapes, so they can work towards effective and efficient management approaches. Remote sensing offers the ability to measure vegetation cover at large spatial scales, which can lead to a better understanding of the invasive characteristics of problematic species like medusahead. For instance, research has been successful in creating large-scale distribution maps of cheatgrass over western rangelands. Many applications rely on the phenological characteristics of a target plant which can present problems in separating two species with similar phenologies (i.e. cheatgrass & medusahead). A medusahead-specific map gives managers the flexibility to prioritize and direct management needs when attempting to control the spread of medusahead into non-invaded areas. This study integrated GPS acquired field locations from three study sites (Sites S, C, & N) and imagery from two remote sensing platforms (1-m aerial imagery & 30-m Landsat), to model and predict fractional cover of medusahead over 37,000+ ha of rangelands in the Channeled Scabland region of eastern Washington. Using a multi-scaled approach, this research showed that regression tree algorithms can model the complex spectral response of senesced medusahead using late summer Landsat scenes. The predictive performances resulted in a R² of 0.80 near the model's training site (Site S) and an average R² of 0.68 away from the training site (Sites C & N). This research provides a non-phenological approach to produce accurate large-scale, distribution maps of medusahead which can aid land managers who are challenged by its invasion.