PS 19-176 - Developing and extrapolating allometric models for aboveground biomass of Wyoming big sagebrush

Background/Question/Methods

Sagebrush ecosystems cover 430,000 km² of the Western United States, with the most common shrub being Artemisia tridentata subsp. wyomingensis, or Wyoming big sagebrush. Management of Wyoming big sagebrush communities can be enhanced by methods to estimate biomass of individuals, as these estimates can be used to calculate carbon storage and primary production of sagebrush communities. Allometric models are a non-destructive method to estimate biomass. They use morphological characteristics and mathematical formulas to model the relationship between shape and biomass of an individual, and are commonly used for biomass estimation in forests and shrublands. We measured morphological traits and then harvested Artemisia tridentata subsp. wyomingensis individuals from across the state of Wyoming. The shrubs were dried, separated based on biomass type, and weighed. We used these data to develop our allometric models. We used three forms of allometric models in our development: non-linear power, log-transformed linear, and logistic curve models. We used several criteria to identify the most accurate models, including K-fold cross-validation. In addition, we investigated which forms of allometric models performed best for data extrapolation and applicability outside the range of the data we used to estimate their parameters.

Results/Conclusions

Four allometric models for Artemisia tridentata subsp. wyomingensis were created; one each for leaf, small branch (diameter < 1 cm), large branch (diameter > 1 cm), and total biomass. Crown volume was typically the best predictor of sagebrush biomass. Power models outperformed logistic curve and log-transformed linear models for biomass estimation. Power and log-transformed linear models tended to have better cross-validation scores than logistic curve models, while logistic curve and power models had better AIC and RMSE values. When extrapolating trained models to data from sites outside the range of training data, log-transformed linear models performed significantly better than power or logistic curve models. While they have acceptable scores for typical criteria like RMSE and AIC, logistic curve models appear to overfit the data, not performing well in cross-validation or extrapolation. Our conclusions are that when working with limited datasets, it is advisable to develop log-transformed linear allometric models, while non-linear power models are ideal for more comprehensive datasets. Our new models outperformed established global and regional models for shrub biomass estimation in this region, indicating that they could be used for future applications of shrub biomass estimation in this region.