Artificial neural networks as an alternative method to nonlinear mixed-effects models for tree height predictions
Graphical abstract
Introduction
Forest inventories are one of the primary sources for national and international reporting schemes (e.g. FAO, 2020), with wood volume, forest biomass and carbon stock among the most important attributes to be reported (Vidal et al., 2016). Methods for wood volume estimates are usually country-specific relying on diverse volume model types, the most common being the use of different volume functions, taper-curves, and breast-height form factor functions (Gschwantner et al., 2019). Directly and indirectly all of them require diameter at breast height (DBH, or D) and tree height (H) estimates. In addition to volume and biomass estimates, tree heights are of key importance for assessment of forest components such as productivity, site indexes and forest development in general. Consequently almost all growth and yield models require information on tree height to predict forest dynamics (Barreiro and Tomé, 2017), where tree heights could be needed at the tree (e.g. Pretzsch et al., 2002, Buchacher and Ledermann, 2020), plot or stand level (e.g. Härkönen et al., 2019). Therefore, highly accurate estimates of tree heights are crucially important in several forestry subdisciplines, as well in related ecological and environmental disciplines.
Field measurements of tree heights are time-consuming and therefore often measured only for a subsample of trees, with unmeasured heights predicted using height-diameter (H-D) models (Soares and Tomé, 2002, Mehtätalo et al., 2015). There are two types of H-D model; the first requires only DBH to predict tree height (H-D function), while the second incorporates stand-level predictors in addition to DBH. The former are called simple and the latter generalised models (Mehtätalo et al., 2015). In the literature two- and three- parameter H-D functions exists (Kindermann, 2016). Simple models are particularly useful in even aged stands with a small number of species in homogenous stand and site conditions. However, the tendency in European forests and beyond is to promote uneven aged and mixed species stand structures (Bravo-Oviedo et al., 2014, Pach et al., 2018), which require more complex modelling approaches that rely on DBH combined with additional stand and tree characteristics (Temesgen and Von Gadow, 2004).
In more complex forest communities, tree heights are often predicted on the basis of information at the plot level, with the plot index entering the model as a random effect of a mixed-effects model (Zuur et al., 2009, Bronisz and Mehtätalo, 2020). Therefore, we assume the same species-specific H-D curve within each plot. Here the fixed part of the model describes the predicted H-D curve for a typical plot in the used database (fixed-effect prediction) and the random effect provides a calibrated prediction (random-effect prediction), which together describe the plot specific H-D relationship. Consequently, the use of mixed-effects models also enable prediction on new plots (Mehtätalo et al., 2015).
The usefulness of including site or plot effects in H-D relationships has been reported for many species, e.g. Pinus Sylvestris (Lappi, 1991), Quercus pagoda (Lynch et al., 2005), Pinus taeda (Trincado et al., 2007), Pseudotsuga menziesii (Temesgen et al., 2008), and Betula pendula (Bronisz and Mehtätalo, 2020). VanderSchaaf (2014) fitted mixed-effects models for ten different conifer species in the Northwest USA, while Mehtätalo et al. (2015) tested modelled H-D relationships using a dataset representing a wide range of tree species.
The advantage of such approaches is the consideration of local site conditions that importantly affect H-D relationships, while the problems with convergence of model could arise if the number of height measurements per plot is low (Harrison et al., 2018), which is often the case in uneven aged mixed forests, with greater diversity of tree species and diameter distributions.
In recent years, machine learning (ML) has seen increased application in various sciences, including forestry. Besides the decision-tree learning and support vector machine, one commonly used method is the artificial neural network (ANN) (Liu et al., 2018). ANNs have the ability to acquire and maintain information based knowledge and can be defined as a set of processing units, represented by artificial neurons, interlinked by a multitude of interconnections (artificial synapses), implemented by vectors and matrices of synaptic weights (da Silva et al., 2017). The ANN model can be applied to various kinds of problems, from classification, clustering and optimisation to function approximation, and has already been applied in various forestry disciplines, such as forest fire prediction (Safi and Bouroumi, 2013), prediction of insect outbreaks (Park and Chung, 2006), and species distribution models (Scrinzi et al., 2007). Apart from these, the ANN has also been tested in tree height modelling for eucalyptus trees (Vieira et al., 2018), common beech (Fagus sylvatica) from northwestern Spain (Castaño-Santamaría et al., 2013) and Crimean juniper (Juniperus excels) (Özçelik et al., 2013).
The primary aim of our study was to explore the application of mixed-effects H-D models for height predictions within predominantly uneven aged mixed forests using forest inventory plot data. Slovenia has a long tradition of close to nature forest management with strong emphasis on natural regeneration, and consequently, a high proportion of uneven aged and mixed stands (Diaci, 2006). Secondly, focused on less representative tree species, we explore the application of grouping plots based on site factors derived from principal component analysis (PCA) (Jolliffe and Cadima, 2016), which are later used as (nested) random effects and categorical independent variables. Finally, we test and compare a new methodological approach based on artificial neural networks (ANN) for tree height predictions for variety of different tree species specific to central Europe. Inclusion of additional explanatory variables, namely competition, was tested for the ANN. Competition is often among the most effective in explaining forest stand dynamics (Jevšenak and Skudnik, 2021, Vospernik, 2021) and studies have reported a significant effect of competition on the modelling of height (Temesgen and Von Gadow, 2004) and height growth (Sharma and Brunner, 2017).
Section snippets
Data
To compare different modelling approaches, we used 5450 tree height measurements from 685 plots from the Slovenian national forest inventory (Skudnik et al., 2021). The data used in this study is from the fourth cycle of the Slovenian nation-wide survey, which was carried out in 2018. The sample trees were measured on permanent concentric sampling plots on a 4 × 4 km grid arranged systematically across the country (Kušar et al., 2010, Skudnik and Hladnik, 2018). At each plot located within the
Comparison of mixed-effects models and ANN at the plot level
The most representative tree species was common beech (Fagus sylvatica), followed by Norway spruce (Picea abies), silver fir (Abies alba) and sessile oak (Quercus petraea) (Table 2). Tree height measurements per species are described in more detail in Supplementary Table 3.
At the plot level comparison (Table 3A), mixed-effects models showed generally more accurate predictions than the ANN models. Out of 16 tree species, ANN1 was more accurate than the mixed-effects models for 7 species.
Use of artificial neural networks for tree height predictions
With our study, we directly compared the ANN approach against the current golden standard, i.e. mixed-effects models (for example Mehtätalo et al., 2015). While there are numerous types of ANN, we decided to use the Bayesian regularised ANN, which is robust to overfitting and often results in an S-shaped curve, similar to growth functions. Nevertheless, users must optimise the complexity of the neural net, which is defined by the selected number of hidden layers and associated neurons (Gardner
Conclusions
Reliable models for tree heights are needed for the estimation of growing stock and biomass, understanding forest dynamics, and assessing site quality. In more complex forest communities, mixed-effects models are the current golden standard for tree height predictions, in which plot-level effects are included as random effects. In this study we present that also ANN can be reliably used to predict tree heights. For Slovenian NFI data using only plot IDs, the mixed-effects approach showed the
Funding
Funding for this study was provided by the Slovene Research Agency: Program and Research Group “Forest biology, ecology and technology” (P4-0107) and Target research project “Development of models for forest management in Slovenia” (V4-2014B). The collection of data used in this study (Slovenian NFI Data) was financed by the Ministry of Agriculture, Forestry and Food in the scope of the “Public Forest Service” programme. Jernej is grateful for the support by the World Federation of Scientists,
Declaration of Competing Interest
The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.
References (70)
- et al.
Projection Systems in Europe and North America: Concepts and Approaches
- et al.
On the use of cross-validation for time series predictor evaluation
Inf. Sci.
(2012) - Bravo-Oviedo, A., Pretzsch, H., Ammer, C., Andenmatten, E., Barbati, A., Barreiro, S., Brang, P., Bravo, F., Coll, L.,...
- et al.
Mixed-effects generalized height–diameter model for young silver birch stands on post-agricultural lands
For. Ecol. Manage.
(2020) - et al.
Interregional Crown Width Models for Individual Trees Growing in Pure and Mixed Stands in Austria
Forests
(2020) - et al.
Bayesian regularization of neural networks
- et al.
Tree height prediction approaches for uneven-aged beech forests in northwestern Spain
For. Ecol. Manage.
(2013) - et al.
Model evaluation using grouped or individual data
Psychon. Bull. Rev.
(2008) Height-Diameter and Height-Diameter-Age Equations For Second-Growth Douglas-Fir
For. Sci.
(1967)
Nature-based silviculture in Slovenia: origins, development and future trends
A multivariate simultaneous prediction system for stand growth and yield with fixed and random effects
For. Sci.
A review of robust clustering methods
Adv. Data Anal. Classif.
Artificial neural networks (the multilayer perceptron)-a review of applications in the atmospheric sciences
Atmos. Environ.
A Flexible Height-Diameter Model for Tree Height Imputation on Forest Inventory Sample Plots Using Repeated Measures from the Past
Forests
Harmonisation of stem volume estimates in European National Forest Inventories
Ann. Forest Sci.
A climate-sensitive forest model for assessing impacts of forest management in Europe
Environ. Model. Software
Nitrogen concentrations in mosses indicate the spatial distribution of atmospheric nitrogen deposition in Europe
Environ. Pollut.
Heavy metal and nitrogen concentrations in mosses are declining across Europe whilst some “hotspots” remain in 2010
Environ. Pollut.
A brief introduction to mixed effects modelling and multi-model inference in ecology
PeerJ
A random forest model for basal area increment predictions from national forest inventory data
For. Ecol. Manage.
Principal component analysis: a review and recent developments
Philos. Trans. A Math. Phys. Eng. Sci.
Field calibration of merchantable and sawlog volumes in forest inventories based on airborne laser scanning
Can. J. For. Res.
Evaluation of growth functions for tree height modelling
Austrian J. For. Sci.
Chapter 7 - Linear Mixed Effects Models
I. Gozdna inventura
Methodological bases of the forest and forest ecological condition survey
Mixed linear models for analyzing and predicting stem form variation of scots pine
Calibration of Height and Volume Equations with Random Parameters
For. Sci.
A longitudinal analysis of height/diameter curves
For. Sci.
Application of machine-learning methods in forest ecology: recent progress and future challenges
Environ. Rev.
A Random-Parameter Height-Dbh Model for Cherrybark Oak
South. J. Appl. For.
Cited by (15)
Uncertainty quantification for structural response field with ultra-high dimensions
2024, International Journal of Mechanical SciencesShifting potential for high-resolution climate reconstructions under global warming
2024, Quaternary Science ReviewsMachine Learning Forest Simulator (MLFS): R package for data-driven assessment of the future state of forests
2023, Ecological InformaticsPrediction of tree crown width in natural mixed forests using deep learning algorithm
2023, Forest EcosystemsIndividual tree detection and estimation of stem attributes with mobile laser scanning along boreal forest roads
2022, ISPRS Journal of Photogrammetry and Remote SensingCitation Excerpt :For instance, Kolendo et al. (2021) used a large-scale reference dataset to parameterize ITD algorithms in coniferous forests, reaching tree count RMSEs varying from approximately 6 to 13%, depending on the forest type. Skudnik and Jevšenak (2022) found that, in the presence of sufficient reference data for calibration, artificial neural network-derived tree height predictions can outperform predictions derived from mixed effect models. Generally, deep learning methods require large datasets for calibration to be used at their full potential (Hamraz et al., 2019; Xi et al., 2020).