1992 Irrigated Lands for the Raft River Valley: Machine Learning Generated

This raster file represents land within the Raft River Study Area classified as either “irrigated” with a cell value of 1 or “non-irrigated” with a cell value of 0 at a 30-meter spatial resolution. These classifications were determined at the pixel level by a Random Forest supervised machine learning methodology. Random Forest models are often used to classify large datasets accurately and efficiently by assigning each pixel to one of a pre-determined set of labels or groups. The model works by using decision trees that split the data based on characteristics that make the resulting groups as different from each other as possible. The model “learns” the characteristics that correlate to each label based on manually classified data points, also known as training data.A variety of data can be supplied as input to the Random Forest model in making its classification determinations. Irrigation produces distinct signals in observational data that can be identified by machine learning algorithms. Additionally, datasets that provide the model with information on landscape characteristics that often influence whether irrigation is present are also useful. This dataset was classified by the Random Forest model using top-of-atmosphere reflectance data from Landsat 5, Mapping Evapotranspiration with Internalized Calibration (METRIC) data, United States Geological Survey National Elevation Dataset (USGS NED) data, and Height Above Nearest Drainage (HAND) data. Landsat 5, METRIC, and HAND data are at a 30-meter spatial resolution, and the USGS NED data are at a 10-meter spatial resolution. The National Land Cover Dataset (NLCD) from USGS, Bureau of Reclamation (BOR) Land Use and Land Cover data, as well as Digital Ortho Photo Quadrangle (DOQQ) data were also used in determining irrigation status for the manually classified training data points but were not used for the machine learning model predictions. The final model results were manually reviewed prior to release, however, no extensive ground truthing process was implemented. “Speckling”, or small areas of incorrectly classified pixels, in the mountain areas was reduced by masking all pixels with a slope value of 15% or greater as “non-irrigated”, regardless of the status they were assigned by the Random Forest model. Speckling within irrigated areas was reduced by a majority filter smoothing technique using a kernel of 8 nearest neighbors.