Application of Botanical Composition Identification and Accuracy Assessment for Grassland Mixtures and Brush from Unoccupied Aerial System Imagery

Abstract

Grassland comprises approximately 40.5% of terrestrial land area, which is around 5.25 billion ha in total. Grasslands are vital, multi-functional ecosystems for effective nutrient cycling, and habitat and forage for livestock and wildlife. Regardless of its importance, grassland area is declining due to various reasons, mainly human activity. Invasive species are one example of human activity which leads to grassland loss. To provide adequate information to facilitate grassland preservation, mapping botanical composition of grasslands can provide crucial information for land management and further quantitative analyses, such as herbage mass estimations or herbicide application to control brush. Remotely sensed data has been used in numerous studies to map land use and cover at various spatial scales to better understand the Earth’s surface. Unoccupied Aerial System (UAS) imagery can facilitate classifying different grass and brush species to estimate botanical composition and associated characteristics. The main objective of this study is to find the optimal way to classify forages within a grassland and brush within a rangeland in South Texas. Grassland mixture classification had overall accuracy from 70 – 99% through the growing season for both classification algorithm Support Vector Machine (SVM) and Random Forest (RF) The ideal time, determined by the greatest user’s, producer’s, and overall accuracy, to differentiate grasses from each other in thsi experiment is in late July and August for identifying Bermudagrass and Johnsongrass, after mid-October for identifying Bermudagrass and KR bluestem. and before March for identifying Bermudagrass and burr medic. Brush classification had overall accuracy from 83 – 96%, and object-based classification using RF classifier had better results than pixel-based classification since object-based classification had the highest overall accuracy in both sit at 96%. In Site 1 with band combination of stacking all bands but NDVI band had 96% overall accuracy using object-based classification. In Site 2 with band combination of stacking all bands also had 96% overall accuracy with object-based classification. The UAS image was useful to assess herbicide efficacy by calculating canopy change after herbicide treatment. In conclusion, UAS derived multispectral imagery can be used to of identify grass species in a mixture and identify brush species in rangelands.