Wed, Aug 04, 2021:On Demand
Background/Question/Methods
We developed a non-native invasive grass monitoring tool for operational use to produce maps of herbaceous cover (%) on a biennial basis over the last decade and for upcoming years for southern California at 30 m resolution. Beta regression and random forest models were tested using Landsat 5, 7, and 8 data, including a variety of vegetation indices and spectral bands, as well as a predictor representing the time elapsed since the most recent fire.
Results/Conclusions Comparing models created in six different years for which high resolution training data were available, we found that a random forest model trained using 2009 data performed better than models trained in other years in terms of variance explained by predictors (70% variance explained), inter-annual temporal transferability, cross-validation predictions for withheld data (root mean square error = 4.14), and minimal bias in burned v. unburned areas. Predictors included the maximum spring value for red-green angle (RGA), mean summer RGA, and standard deviation in the spring for the normalized burn ratio (NBR); these predictors leverage the distinct phenology of annual grasses as compared to evergreen shrubs that dominate the study area. Feasibility considerations for maintenance of an operational tool were an important component of model development, and we discuss 1) leveraging data processing capabilities using Google Earth Engine, 2) a cost-benefit comparison of model re-training for future years, and 3) an assessment of training methods that relied primarily on imagery-based assessments, while reserving field visitation for targeted assessments when conditions warrant further validation. Model outputs predicted to the full raster scale indicated that non-native invasive grass spread has occurred at a moderate level over the last decade, with increases observed mainly in burned areas, but with expansion and density likely suppressed by recent drought.
Results/Conclusions Comparing models created in six different years for which high resolution training data were available, we found that a random forest model trained using 2009 data performed better than models trained in other years in terms of variance explained by predictors (70% variance explained), inter-annual temporal transferability, cross-validation predictions for withheld data (root mean square error = 4.14), and minimal bias in burned v. unburned areas. Predictors included the maximum spring value for red-green angle (RGA), mean summer RGA, and standard deviation in the spring for the normalized burn ratio (NBR); these predictors leverage the distinct phenology of annual grasses as compared to evergreen shrubs that dominate the study area. Feasibility considerations for maintenance of an operational tool were an important component of model development, and we discuss 1) leveraging data processing capabilities using Google Earth Engine, 2) a cost-benefit comparison of model re-training for future years, and 3) an assessment of training methods that relied primarily on imagery-based assessments, while reserving field visitation for targeted assessments when conditions warrant further validation. Model outputs predicted to the full raster scale indicated that non-native invasive grass spread has occurred at a moderate level over the last decade, with increases observed mainly in burned areas, but with expansion and density likely suppressed by recent drought.