Machine learning prediction of mortality in the common desert shrub Encelia farinosa

Two populations of the common shrub Encelia farinosa in the northern and southern portions of the Mojave Desert have been surveyed each spring for nearly 40 years, providing an opportunity to assess highly variable shrub mortality in an arid ecosystem. Most of the newly established shrubs experienced mortality during the juvenile stage, with median survival time of about three years in both populations yet, a small number of shrubs lived for at least a dozen years or even decades. Applying machine learning techniques, we predicted shrub mortality at different life-history stages using random forest and logistic regression. First, we examined seedling survival to become yearlings (one-year old plants), finding that less than 3% of seedlings in both populations survived to become established yearling shrubs. Second, we predicted whether or not yearlings would die prior to reaching the mature adult stage (four years old). The models achieved an Area Under the Receiver Operating Characteristic (AUC) in the 0.80 range for the Oatman population (southern Mojave Desert) and 0.90 range for the Death Valley population (northern Mojave Desert). We found yearling characteristics of smaller shrub size, low leaf coverage, and location in specific microsites associated with experiencing mortality before reaching the mature stage. Third, using only the average juvenile plant characteristics over the first four years of life, we predicted whether or not new adult shrubs were likely to experience mortality within the next eight years. The performance in this application achieved AUC in the 0.72 range for both populations. We found adult Encelia farinosa shrubs that had juvenile characteristics of smaller size, flowered less frequently, and had smaller interplant distances for the Oatman population were associated with increased mortality within the next eight years. Overall, the size of the shrub was the most important feature for the mortality modeling applications. No significant difference in AUC was found for random forest and logistic regression.

Automated summary

Two populations of Encelia farinosa shrubs in the Mojave Desert were surveyed for nearly 40 years, revealing highly variable mortality rates. Machine learning techniques were used to predict shrub mortality at different life-history stages, with characteristics such as shrub size, leaf coverage, and microsite location influencing mortality risk. Overall, shrub size was found to be the most important factor in mortality modeling.