Allometric equations are commonly used to estimate tree biomass; however, actual measurements of tree biomass are required to develop these allometric models. Terrestrial laser scanning (TLS) can provide significant improvements in this regard. Here, we propose a TLS-based approach to enhance conventional destructive measurements and model fitting by selecting subsamples that mitigate prediction bias. The approach involves scanning a large number of trees to estimate their above-ground volume, fitting a model to multiple tree subsamples, and selecting suitable subsamples through cross-validation. We compared simple random sampling (SRS), systematic sampling (SYS), and stratified sampling (STR). Our goals were to (i) quantify the influence of the tree selection on the fit of allometric models; (ii) assess the utility of TLS-based estimates in building an optimal tree sample for destructive measurement, particularly to identify samples that potentially lead to biased predictions and large mean prediction errors; (iii) compare subsampling methods and determine key characteristics contributing to model fit quality or bias; and (iv) evaluate the impact of sample size on the different sampling methods. To test the proposed method, we scanned and harvested 184 hybrid poplar trees to measure their volume and biomass, then we established subsamples to develop a model and assess the predictive capacity of each fit. Results showed that the samples leading to a high prediction bias could be detected and avoided using the TLS-based volume estimates. Bias was the most discriminant statistic in comparing subsamples. However, certain trees with a high influence on model fit may not be easily identified. Sample size affects prediction error, aiding in optimal sample size selection. SRS was identified as an efficient sampling method. We conclude that TLS offers a viable solution for optimizing tree sampling strategies.
Dănilă et al. (Wed,) studied this question.