Get reliable science

Accurate mapping of rangeland vegetation is critical in studies of livestock spatial distribution. Our objective was to develop a detailed vegetation map to help model the drivers of beef heifer grazing behavior on Northern Great Plains rangeland at the USDA-ARS Fort Keogh LARRL, Miles City, MT. Remote sensing technologies for mapping vegetation are less time-intensive than traditional on-the-ground techniques and cover larger areas. However, map classes should correspond with vegetation classes of interest, and the accuracy of a final map product should be known. Unsupervised classification of National Agricultural Imagery Program (NAIP) imagery guided ground data collection for training / testing supervised classification. Two line point transects were read at 15 sites within each class to determine plant cover and species composition. A 2 m belt transect was read along each line point transect to determine shrubs density. We discriminated four classes. Class 1: Badlands - sparsely vegetated hills; 57.7 � 3.8 % bare ground (BG) and 998.7 � 151.4 shrubs/ha. Class 2: Mixed grass prairie, dominated by cool season grasses; 13.7 � 2.5 % BG and 227.5 � 65.1 shrubs/ha. Class 3: Shrublands with mixed grass prairie understory; 27.1 � 3.3 % BG and 765.1 � 124.8 shrubs/ha. Class 4: Cool season grasses/legumes and shrubs; 8.5 � 3.1 % BG and 1098.6 � 268.4 shrubs/ha. Training data class separability was evaluated using Jeffries-Matusita and transformed divergence methods (ENVI software). Values of 1.6 � 1.9 indicated that classes were sufficiently separable. We used ENVI�s maximum likelihood algorithm to classify combined NAIP, Landsat 8 Operational Land Imager and digital elevation datasets (spatial resolution sampled to 15 m). Testing data were used to develop a confusion matrix and to calculate Kappa. Overall accuracy was 89.1 % with Kappa = 0.8. We conclude that these results are acceptable for the scope of our study.