Search

We evaluated area-based approaches (ABAs) to light detection and ranging (lidar) predictions of plot- and stand-level forest attributes (tree count, height, basal area, volume, aboveground biomass, broadleaf/conifer, and diameter at breast height--'diameter'). ABA methods included post-stratification (PS), ordinary least squares (OLSs) regression, k nearest neighbors (kNN), and random forest (RF). This study was conducted on the Savannah River Site in South Carolina, USA. Plot- and stand-level predictions were validated against fixed-radius 0.04 ha (0.1 acre) plots in 49 [approximately equal to] 2.0 ha (5 acre) stands. Our findings demonstrate that lidar can be incorporated operationally into forest inventory systems to provide stand-level inferences for a wide range of forest attributes. Volume predictions for specific diameter classes, however, often fared poorly (root mean squared error (RMSE) > 100%) for the methods we explored, especially for larger (less common) diameter trees. Stand-level results were consistently better than pixel-level results (10-200+ percentage points). kNN and RF performed similarly and better than OLS and PS, but RF was the most robust to model configurations, while kNN has practical advantages such as simultaneous predictions of many attributes. Key words: lidar, forest inventory, stand-level inference, area-based approach, sampling, 1. Introduction Forest inventories provide information needed by land managers to assess forest conditions with respect to objectives such as sustainable yield, preservation, and wildlife habitat. Under ideal conditions, forest [...]

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 [...]

Ref.: Can. J. For. Res. 53(2): 1-19 (2023) dx.doi.org/10.1139/cjfr2022-0032 In the originally published article, the second row in eq. 5 had a misplaced exponent. The original and corrected versions of [...]

In airborne laser scanning (ALS)-based forest inventories, there is commonly a discrepancy between the plot shape used for model fitting (typically circular) and the shape of population elements (typically square) where predictions are needed. Circular plots are easy to establish, locate and have the smallest number of edge trees on average. Therefore, a circle is the most common plot shape in both traditional and remote sensing-based forest inventories. In contrast, the shape of population elements used for remote sensing-based predictions is nearly always a square because it enables division of the target population into a grid of non-overlapping plots. In this study, we investigate shape effects for ALS-based forest inventories using circular and square plot shapes. This has not been examined earlier. Aboveground biomass was used as the response variable. The sampling design was created in a way that the probability of selection for any location inside a stem-mapped 30 m × 30 m plot was the same for the circular (radius 7.95 m) and square (side length 14.09 m) plot. This configuration enabled us to compare circular and square plots with the same areas and identical sampling probabilities for every tree in the population. Our primary finding is that for equal area square and circular plots, there is no evidence of systematic prediction error when a model fitted to one shape is used to predict for the other shape. Our secondary finding is that root mean square error (RMSE) value is slightly underestimated (1.2 per cent) when a model fitted to circular plots is used to predict for square plots. A small underestimation of RMSE due to plot shape effect has hardly practical significance in stand-level forest management inventories, but the plot shape effect may be problematic in large area forest surveys. [ABSTRACT FROM AUTHOR]

In this study, for five sites around the world, we look at the effects of different model types and variable selection approaches on forest yield modelling performances in an area-based approach (ABA). We compared ordinary least squares regression (OLS), k-nearest neighbours (kNN) and random forest (RF). Our objective was to test if there are systematic differences in accuracy between OLS, kNN and RF in ABA predictions of growing stock volume. The analyses are based on a 5-fold cross-validation at five study sites: an eucalyptus plantation, a temperate forest and three different boreal forests. Two completely independent validation datasets were also available for two of the boreal sites. For the kNN, we evaluated multiple measures of distance including Euclidean, Mahalanobis, most similar neighbour (MSN) and an RF-based distance metric. The variable selection approaches we examined included a heuristic approach (for OLS, kNN and RF), exhaustive search among all combinations (OLS only) and all variables together (RF only). Performances varied by model type and variable selection approaches among sites. OLS and RF had similar accuracies and were more efficient than any of the kNN variants. Variable selection did not affect RF performance. Heuristic and exhaustive variable selection performed similarly for OLS. kNN fared the poorest amongst model types, and kNN with RF distance was prone to overfitting when compared with a validation dataset. Additional caution is therefore required when building kNN models for volume prediction though ABA, being preferable instead to opt for models based on OLS with some variable selection, or RF with all variables together. [ABSTRACT FROM AUTHOR]

Airborne laser scanning, collected in a sampling mode, has the potential to be a valuable tool for estimating the biomass resources available to support bioenergy production in rural communities of interior Alaska. In this study, we present a methodology for estimating forest biomass over a 201,226-ha area (of which 163,913 ha are forested) in the upper Tanana valley of interior Alaska using a combination of 79 field plots and high-density airborne light detection and ranging (LiDAR) collected in a sampling mode along 27 single strips (swaths) spaced approximately 2.5 km apart. A model-based approach to estimating total aboveground biomass for the area is presented. Although a design-based sampling approach (based on a probability sample of field plots) would allow for stronger inference, a model-based approach is justified when the cost of obtaining a probability sample is prohibitive. Using a simulation-based approach, the proportion of the variability associated with sampling error and modeling error was assessed. Results indicate that LiDAR sampling can be used to obtain estimates of total biomass with an acceptable level of precision (8.1 ± 0.7 [8%] teragrams [total ± SD]), with sampling error accounting for 58% of the SD of the bootstrap distribution. In addition, we investigated the influence of plot location (i.e., GPS) error, plot size, and field-measured diameter threshold on the variability of the total biomass estimate. We found that using a larger plot (1/30 ha versus 1/59 ha) and a lower diameter threshold (7.6 versus 12.5 cm) significantly reduced the SD of the bootstrap distribution (by approximately 20%), whereas larger plot location error (over a range from 0 to 20 m root mean square error) steadily increased variability at both plot sizes. [ABSTRACT FROM AUTHOR]