Driver for Projections Tillage Suitability 2016

The tillage suitability product is a per-crop, per-pixel (30 square-meters) model representation of the predicted probability (0.00-1.00) that an area can support commodity crop development for a suite of crop types commonly grown in the LCD landscape. The values for each grid cell are interpreted as a probability, with any value greater-than 0.50 suggesting an area should be suitable for crop development based on observations of thousands of farmed areas around the LCD. To demonstrate composite suitability (“tillage”) for all crops, we added the individual probabilities for our modeled from cover classes (cereals, corn, cotton, and beans; described below), which represents the overall proportion of votes for “crop” vs. “not-crop/other” for trees in our full Random Forest model. To model tillage suitability, we used 2.5 million observations of predominant crop cover types found throughout the LCD region as training data, sampled from the annual USDA-NASS Cropland Data Layer dataset (2016 CDL; XXX). USDA-NASS CDL is a 30m composite gridded model representation of crop production published for the continental United States that represents common crop types found throughout the region. We excluded all observations of crop production outside of the LCD area from consideration during model building. We aggregated individual crop type designations from NASS-CDL into seven broad categories thought to represent crops that occur most frequently in the LCD geography. These were 1.) corn, 2.) cotton, 3.) cereals, 4.) beans, 5.) native grasses and shrubs, 6.) crop grasses (e.g., switchgrass), and 7.) other. We fit our model using a Random Forests classifier to a suite of landscape variables thought to be limiting factors governing suitability for crop development derived from common climate (ClimateWNA), soils (SSURGO), topographic (NED), and groundwater datasets (Ogalalla). To account for potential model overfit and remove unimportant variables from consideration, we used a threshold-based optimization procedure that sought to minimize Random Forests’ out-of-bag (OOB) error statistic using an importance threshold (parsimony=0.3) applied to each candidate variable in an iterative model building process (Murphy et al., 2010). Variables with an OOB importance consistently in the lower (30th) percentile were excluded from the final model. Our final candidate variable series was : Available Water Storage (25-150 CM classes; SSURGO), Growing Degree Days > 5 degrees celcius (ClimateWNA), elevation (NED), std deviation of elevation (8100 m2 window; NED), Frost Free Period (ClimateWNA), Hydrological Group Category (SSURGO), Irrigated Capability Class (SSURGO), % of Irrigated Capacity Class (SSURGO), Mean Annual Precipitation (ClimateWNA), Precipitation Tenths (ClimateWNA), Soil Capability Class without Irrigation (SSURGO), Topographic Roughness (0.66 km2 moving window; NED), Current Aquifer Saturated Thickness (Ogalalla), Slope Gradient - Dominant Component (SSURGO), Slope Gradient - Weighted Average (SSURGO). We then projected suitability across the LCD region as a continuous raster surface and used quantile-based contour-delineation to extract relevant suitability hotspots.