Destructive sampling-based generalized additive model (GAM) prediction of Dendrocalamus asper (Schult. & Schult.f.) Backer root–shoot ratios

The above-ground biomass reported for Dendrocalamus asper is impressive. However, there is a dearth of studies on the below-ground biomass of this species . This knowledge gap may leave farmers and other stakeholders unable to fully realize the financial benefits of participating in voluntary carbon credit initiatives. In this study, empirical estimates of below-ground biomass and root-to-shoot ratios (RSR) were generated for Dendrocalamus asper using destructive sampling (n=27). I found an RSR range of 0.18–2.20. Of the eight Generalized Additive Models (GAMs) that I evaluated, the predictive model involving fixed-effect bamboo class size, a smooth for the number of culms per clump, and random-effect smooths for location and site within location was selected for its robust, stable performance. In-sample and cross-validation Leave-One-Out (LOOCV) and K-fold R 2 performance of the selected model were 0.743, 0.512 and 0.466, respectively. Post-hoc power estimates, based on the terms of the chosen GAM, yielded 0.997 and 0.86 for the model terms, number of culms smooth, and intercept, respectively. All other terms in the model were near 0, except for one class parameter, which was at 0.64. A Monte Carlo simulation was conducted to determine whether 27 samples were sufficient to predict RSR using the chosen GAM. Increasing the sample size from 1 to 20 per site or from 27 to 540 overall did not increase the power estimates for low-powered terms. However, caution should be exercised as simulated data cannot substitute for biological replication. Future work with larger, more diverse datasets across ecological gradients will be necessary to refine and assess the broader generalizability of this study's findings. • Generalized Additive Model (GAM) in-sample and cross-validation predictive metrics of purely additive and single smooth models performed better in predicting Dendrocalamus asper Root Shoot ratio (RSR) than complex models. • Two measured bamboo physical attributes, the circumference of a clump and the number of culms in a clump, were used to predict the root shoot ratio (RSR). • Predictive in-sample and out of sample metrics of the GAM model with either the circumference of a clump or the number of culms in a clump practically performed equally well compared to when both are included in a model. • Monte Carlo simulations of an increased sample size from 1 to 20 per site did not improve the estimated power of the selected model.