2018 Irrigated Lands for the Eastern Snake Plain Aquifer: Machine Learning Generated

Eastern Snake River Plain Irrigated Lands 2018 was created for use in water budget studies in the Eastern Snake River Plain. The area of interest was determined by Hydrology Section staff at IDWR, and a study boundary was given to GIS staff and used to clip the model output. The random forest (RF) model is a type of supervised machine learning algorithm requiring GIS staff to provide manually labeled training data. GIS staff also provide the RF model with several input features, typically raster datasets that help distinguish characteristics of irrigated lands. Eastern Snake River Plain Irrigated Lands 2018 used the following as input features: • Seasonally averaged Landsat 8 [2] and Sentinel-2 surface reflectance imagery [3] (bands: SWIR 2, NIR, Blue, and calculated NDVI) • 10-meter digital elevation model [4] (including slope and aspect) • PRISM [5] 800 meter seasonal averaged climate data • IDWR METRIC [7] evapotranspiration dataset • Height Above Nearest Drainage (HAND) [6] • Topographic Wetness Index, derived from the digital elevation model For additional information on the averaging process for Landsat and Sentinel-2 imagery, please see below. Additional datasets used only for labeling training data include Mapping EvapoTranspiration at high Resolution with Internalized Calibration (METRIC) [7], IDWR-provided Active Water Rights Place of Use, and the Cropland Data Layer [8] for 2018. The accuracy of the Eastern Snake River Plain Irrigated Lands 2018 dataset was verified by several methods. Firstly, a validation test is done by withholding a subset of the training data to evaluate how well the model classifies unseen information. Second, GIS staff will run several iterations of the model with variations of training data, with the goal of improving classification for areas consistently misclassified. This process requires GIS staff knowledge, aided by supplementary datasets, to review the area and make decisions. Once a model iteration is determined as ‘final’, a manual mask is created to correct any remaining misclassification in the dataset. Consistent areas of misclassification for Eastern Snake River Plain Irrigated Lands 2018 include the area between Ashton and Lamont, pastures within Chester, dry and fallow fields in southern Twin Falls county, and separation between wetlands and irrigated fields in Wood River and the Lost River Valleys. Misclassification largely occurred in areas with active water rights, but no visible irrigation or irrigation infrastructure in the satellite and aerial imagery available. Final decisions made for these areas were conservative, relying heavily on both an active water right and clear indication of diverted pressurized water and purposeful application of water to a given field. References: [1] https://developers.google.com/earth-engine/apidocs/ee-classifier-smilerandomforest [2] https://developers.google.com/earth-engine/datasets/catalog/NASA_HLS_HLSL30_v002 [3] https://developers.google.com/earth-engine/datasets/catalog/NASA_HLS_HLSS30_v002 [4] https://developers.google.com/earth-engine/datasets/catalog/USGS_3DEP_10m [5] Daly, C., Halbleib, M., Smith, J.I., Gibson, W.P., Doggett, M.K., Taylor, G.H., Curtis, J. & Pasteris, P.A. (2008). Physiographically sensitive mapping of climatological temperature and precipitation across the conterminous United States. International Journal of Climatology, 28, 2031-2064. [doi:10.1002/joc.1688](https://doi.org/10.1002/joc.1688) [6] Donchyts, G., Winsemius, H., Schellekens, J., Erickson, T., Gao, H., Savenije, H., & van de Giesen, N. (2016). Global 30m height above the nearest drainage (HAND). Geophysical Research Abstracts, 18, EGU2016-17445-3. EGU General Assembly 2016. [7] https://data-idwr.hub.arcgis.com/documents/776cfc545e0944fc89a75d4777031bb4/about [8] https://developers.google.com/earth-engine/datasets/catalog/USDA_NASS_CDL Information on averaged imagery: GIS staff average Landsat and Sentinel-2 imagery in two month increments to fill gaps of missing data. Images are averaged using the HLSS and HLSL datasets for March 1st through May 1st, May 1st through July 1st, July 1st through September 1st, and September 1st through November 1st. Averaging results in 4 total images that are entirely spectral data.