Search

Semi- and nonparametric models are popular in the area-based approach (ABA) using airborne laser scanning. It is unclear, however, how many predictors and training plots are needed to provide accurate predictions without overfitting. This work aims to explore these limits for various approaches: ordinary least squares regression (OLS), generalized additive models (GAM), least absolute shrinkage and selection operator (LASSO), random forest (RF), support vector machine (SVM), and Gaussian process regression (GPR). We modeled timber volume ([m.sup.3]*[ha.sup.-1]) for four boreal sites using ABA with 2-39 predictors and 20-500 training plots. OLS, GAM, LASSO, and SVM overfitted as the number of predictors approached the number of training plots. They required [greater than or equal to] 15 plots per predictor to provide accurate predictions (RMSE [less than or equal to] 30%). GAM required [greater than or equal to] 250 plots regardless of the number of predictors. The number of predictors only mildly affected RF and GPR, but they required [greater than or equal to] 200 and [greater than or equal to] 250 training plots, respectively. RF did not overfit in any circumstances, whereas GPR overfit even with 500 training plots. Overall, using up to 39 predictors did not generally result in overfit, and for most model types, it resulted in better accuracy for sufficiently large datasets ([greater than or equal to] 250 plots). Key words: LiDAR, machine learning, remote sensing, area-based approach, sampling size. Les modeles semi et non parametriques sont populaires avec l'approche territoriale (AT) qui utilise le balayage laser aeroporte. On ignore cependant combien de predicteurs et de parcelles d'entrainement sont necessaires pour generer des predictions justes sans surapprentissage. Cet article vise a explorer ces limites pour differentes approches : regression des moindres carres ordinaire (MCO), modeles additifs generaux (MAG), operateur de selection et reduction par moindres valeurs absolues (LASSO); foret aleatoire (FA), machine a vecteurs de support (MVS), et regression de processus gaussien (RPG). Nous avonsmodelise le volume de bois ([m.sup.3]*[ha.sup.-1]) dans quatre stations boreales a l'aide de l'AT avec 2 a 39 predicteurs et 20 a 500 parcelles d'entrainement. Les approches MCO, MAG, LASSO et MVS surajustaient lorsque le nombre de predicteurs approchait le nombre de parcelles d'entrainement. Elles exigeaient au moins 15 parcelles par predicteur pour generer des predictions justes (EMQ [less than or equal to] 30%). L'approche MAG necessitait au moins 250 parcelles peu importe le nombre de predicteurs. Le nombre de predicteurs influencait seulement legerement les approches FA et RPG mais elles exigeaient respectivement au moins 200 et 250 parcelles. L'approche FA ne surajustait pas quelles que soient les circonstances, tandis que l'approche RPG surajustait meme avec 500 parcelles d'entrainement. Dans l'ensemble, l'utilisation de 39 predicteurs n'engendrait generalement pas de surapprentissage et, avec la plupart des types de modeles, cela se traduisait par une precision accrue pour des jeux de donnees suffisamment importants (au moins 250 parcelles). [Traduit par la Redaction] Mots-cles : LiDAR, apprentissage machine, teledetection, approche territoriale, taille de l'echantillon., 1. Introduction The development of prediction models is an essential step when performing forest assessments through airborne laser scanning (ALS). In the area-based approach (ABA) to forest inventory with ALS [...]

While the analysis of airborne laser scanning (ALS) data often provides reliable estimates for certain forest stand attributes-such as total volume or basal area-there is still room for improvement, especially in estimating species-specific attributes. Moreover, while the information on the estimate uncertainty would be useful in various economic and environmental analyses on forests, a computationally feasible framework for uncertainty quantifying in ALS is still missing. In this paper, the species-specific stand attribute estimation and uncertainty quantification (UQ) is approached using Gaussian process regression (GPR), which is a nonlinear and nonparametric machine learning method. Multiple species-specific stand attributes are estimated simultaneously: tree height, stem diameter, stem number, basal area, and stem volume. The cross-validation results show that GPR yields on average an improvement of 4.6% in estimate root mean square error over a state-of-the-art $k$ -nearest neighbors (kNNs) implementation, negligible bias and well performing UQ (credible intervals), while being computationally fast. The performance advantage over kNN and the feasibility of credible intervals persists even when smaller training sets are used. [ABSTRACT FROM AUTHOR]

In this paper, we propose an approach to quantify the plot-level uncertainty in species-specific growing stock volume estimated from airborne laser scanning data and aerial imagery. This is accomplished by adopting the framework of Bayesian inference in the area-based estimation of stock volume. The results show that the proposed approach performs well in quantifying the estimate uncertainty and produces optimal interval estimates for species-specific volumes when sufficient training data are available. Also the point estimate accuracy is competitive with current state-of-the-art methods. Furthermore, we demonstrate how the quantified uncertainties of the stand attributes can be utilized to determine the uncertainty in classification done using the estimated stand attributes. [ABSTRACT FROM AUTHOR]