UTILIZATION OF VARIOUS METHODS AND A LANDSAT NDVI/GOOGLE EARTH ENGINE PRODUCT FOR CLASSIFYING IRRIGATED LAND COVER

Author

Andrew Nemecek, University of Montana, MissoulaFollow

Year of Award

2019

Document Type

Thesis

Degree Type

Master of Science (MS)

Degree Name

Geography (Cartography and GIS Option)

Department or School/College

Geography

Committee Chair

David Shively

Commitee Members

Anna Klene Nathaniel Robinson

Keywords

Irrigation, Landsat, Google Earth Engine, Random Forests, Decision Trees, Remote Sensing

Subject Categories

Agriculture | Databases and Information Systems | Hydrology | Remote Sensing | Spatial Science

Abstract

Methods for classifying irrigated land cover are often complex and not quickly reproducible. Further, moderate resolution time-series datasets have been consistently utilized to produce irrigated land cover products over the past decade, and the body of irrigation classification literature contains no examples of subclassification of irrigated land cover by irrigation method. Creation of geospatial irrigated land cover products with higher resolution datasets could improve reliability, and subclassification of irrigation by method could provide better information for hydrologists and climatologists attempting to model the role of irrigation in the surface-ground water cycle and the water-energy balance. This study summarizes a simple, reproducible methodology using 30-meter resolution Landsat NDVI data for classifying irrigated land cover in semi-arid western Montana by leveraging the rising availability of machine learning algorithms in geographic information systems (GIS) software programs to compare results from models constructed using Decision Trees, Random Forest, and principal components analysis. Finally, this study was an attempt to subclassify irrigated land cover into a geospatial layer that distinguishes center pivot irrigation systems from other methods. The Random Forest model was the best model for classifying irrigated land cover, validating its recent use for classifying irrigated land cover in other studies. Further, the NDVI dataset that interpolates cloud and cloud shadow pixels with a user-specified climatology provided a time-series dataset with sufficient spatial and temporal resolution for time-series irrigated land cover classification at the basin and growing season scales. This dataset provides a viable alternative to coarse resolution products often used for creation of geospatial irrigated area datasets at larger scales and an opportunity to create small-scale irrigated area datasets that provide more detailed information. Finally, subclassification of irrigation by method was unsuccessful, but availability of small-scale evapotranspiration datasets and a time-series green index dataset without cloud contamination could improve models.

Recommended Citation

Nemecek, Andrew. 2019. "Utilization of Various Methods and a Landsat NDVI/Google Earth Engine Product for Classifying Irrigated Land Cover." Thesis, University of Montana, Missoula.