| Abstract
|
This layer was created to map the 2013 distribution of Russian Olive and other landcover types in ten large river valley bottoms in eastern Montana known to be infested with Russian olive: the Bighorn, Clark Fork of the Yellowstone, Judith, Milk, Marias, Missouri, Musselshell, Powder, Tongue, and Yellowstone rivers. Other classes mapped include Forested, Upland Emergent, Riparian Emergent, Shrub-Scrub, Sand Bar, Water, Hay, Irrigated Agriculture, Developed, Road, and Railroad.
|
| Purpose
|
This dataset is a polygon layer generated through the segmentation of NAIP 2013 imagery in eCognition 9.0 and its classification using RandomForest in Weka. Ten large rivers in eastern Montana known to be infested with Russian olive were selected (Bighorn, Clark Fork of the Yellowstone, Judith, Milk, Marias, Missouri, Musselshell, Powder, Tongue, and Yellowstone) and their valley bottoms were delineated using contours derived from a 10m DEM and NAIP imagery. Four band, 2013 unprocessed (but georeferenced) NAIP imagery at 1m resolution was extracted for each valley bottom and a pseudo-NDVI was calculated. Agricultural lands were masked out using the most recent (2013) Department of Revenue FLU database; segmentation was conducted in non-agricultural lands using eCognition at a scale of 100, an arbitrary number selected through trial-and-error but aiming to map even small stands of Russian olive. Resulting polygon outlines were displayed on top of NAIP and training points were manually digitized within polygons representative of each of the seven following classes: Forested, Russian Olive, Riparian Emergent, Upland Emergent, Shrub-Scrub, Sand Bar, and Water. By default, developed and barren areas (roads, roofs, paved surfaces) were classified as Upland Emergent or Sand Bar and later updated through post-modeling. Classes poorly represented within a given valley bottom did not enter the classification process but were instead manually recoded during post-modeling. Thirty-five attributes were extracted for each polygon and classification was done using RandomForest in Weka, with number of trees set at 200 and a 10-fold internal cross-validation; results were extrapolated to all segments within each valley bottom. Post-modeling was conducted by scanning each valley bottom and correcting misclassifications manually, giving particular attention to Russian olive. An independent validation performed after post-modeling gave overall accuracies above 90% for all rivers, with accuracy for Russian olive greater than 95%. Agriculture polygons were merged to the final segmentation results. Upland Emergent and Sand Bar polygons overlapping tructures from the most recent Framework geodatabase (04/2015) were reclassified to "Developed". Roads and railroads were extracted from the most recent Framework database (02/2015); roads were buffered based on their width, 10m per lane up to 40m maximum and railroads were buffered by 20m. Both layers were "burned in" the final landcover layer.
|
