Digital soil mapping typically involves inputs of digital elevation models, remotely sensed imagery, and other spatially explicit digital data as environmental covariates to predict soil classes and attributes over a landscape using statistical models. Digital imagery from Landsat 5, a digital elevation model, and a digital geology map were used as environmental covariates in a 67,000-ha study area of the Great Basin west of Fillmore, UT. A “pre-map” was created for selecting sampling locations. Several indices were derived from the Landsat imagery, including a normalized difference vegetation index, normalized difference ratios from bands 5/2, bands 5/7, bands 4/7, and bands 5/4. Slope, topographic curvature, inverse wetness index, and area solar radiation were calculated from the digital elevation model. The greatest variation across the study area was found by calculating the Optimum Index Factor of covariates, choosing band 7, normalized difference ratio bands 5/2, normalized difference vegetation index, slope, profile curvature, and area solar radiation. A 20-class ISODATA unsupervised classification of these six data layers was reduced to 12. Comparing the 12-class map to a geologic map, 166 sites were chosen weighted by areal extent; 158 sites were visited. Twelve points were added using case-based reasoning to total 170 points for model training. A validation set of 50 sites was selected using conditioned Latin Hypercube Sampling. Density plots of sample sets compared to raw data produced comparable results. Geology was used to stratify the study area into areas above and below the Lake Bonneville highstand shoreline. Raster data were subset to these areas, and predictions were made on each area. Spatial modeling was performed with three different models: random forests, support vector machines, and bagged classification trees. A set of covariates selected by random forests variable importance and the set of Optimum Index Factor covariates were used in the models. The Optimum Index Factor covariates produced the best classification using random forests. Classification accuracy was 45.7%. The predictive rasters may not be useful for soil map unit delineation, but using a hybrid method to guide further sampling using the pre-map and standard sampling techniques can produce a reasonable soil map.
| Identifer | oai:union.ndltd.org:UTAHS/oai:digitalcommons.usu.edu:etd-5556 |
| Date | 01 May 2015 |
| Creators | Fonnesbeck, Brook B. |
| Publisher | DigitalCommons@USU |
| Source Sets | Utah State University |
| Detected Language | English |
| Type | text |
| Format | application/pdf |
| Source | All Graduate Theses and Dissertations |
| Rights | Copyright for this work is held by the author. Transmission or reproduction of materials protected by copyright beyond that allowed by fair use requires the written permission of the copyright owners. Works not in the public domain cannot be commercially exploited without permission of the copyright owner. Responsibility for any use rests exclusively with the user. For more information contact Andrew Wesolek (andrew.wesolek@usu.edu). |
Page generated in 0.0023 seconds