A comparative study of regression methods to predict forest structure and canopy fuel variables from LiDAR full-waveform data

Abstract

Regression methods are widely employed in forestry to predict and map structure and canopy fuel variables. We present a study where several regression models (linear, non-linear, regression trees and ensemble) were assessed. Independent variables were calculated using metrics extracted from full-waveform LiDAR data, while the reference data used to generate the dependent variables for the prediction models were obtained from fieldwork in 78 plots of 16 m radius. Transformations of dependent and independent variables with feature selection were carried out to assess their influence in the prediction of response variables. In order to evaluate significant differences and rank regression models we used the non-parametric tests Wilcoxon and Friedman, and post-hoc analysis or post-hoc pairwise multiple comparison tests, such as Nemenyi, for Friedman test. Regressions using transformation of the dependent variable, like square-root or logarithmic, or the independent variable, increased R 2 up to 6% with respect to linear regression using unprocessed response variables. CART (Classification and Regression Tree) method provided poor results, but it may be interesting for categorisation purposes. Square-root transformation of the dependent variable is the method having the best overall results, except for stand volume. However, not always has a significant improvement with respect to other regression methods.Key words: regression models, Random Forest, CART, M5, Wilcoxon, Friedman, forest structure, canopy fuel, LiDAR full-waveform. Estudio comparativo de métodos de regresión para la predicción de variables de estructura y combustibilidad a partir de datos LiDAR full-waveformResumen: Los métodos de regresión se utilizan ampliamente en el ámbito forestal para la predicción y el cartografiado de las variables de estructura y combustibilidad. En este artículo se evalúan diferentes modelos de regresión (lineal, no lineal, árboles de regresión y ensemble). Como variables independientes se utilizaron métricas extraídas de datos LiDAR full-waveform, mientras que los valores de las variables dependientes se generaron a partir de modelos basados en datos de campo obtenidos para 78 parcelas de 16 m de radio. Se llevaron a cabo transformaciones de las variables dependientes e independientes con selección de atributos para evaluar su influencia en la predicción de la variable respuesta. Con el fin de verificar diferencias significativas y ordenar los modelos de regresión se emplearon los tests no paramétricos de Wilcoxon y Friedman, y el análisis post-hoc o los tests de comparación post-hoc por pares, como el de Nemenyi, para el test de Friedman. Las regresiones basadas en la transformación de la variable dependiente,

Evaluación del uso de LiDAR discreto, full-waveform y TLS en la clasificación por composición de especies en bosques mediterráneos

“…This tool computes height and intensity statistics from point clouds, (see Table 3). ALS FW metrics were computed using our own specific software, as reported by (Kimes et al, 2006;Duncanson et al, 2010;Zhang et al, 2011;Ruiz et al, 2016;, and can be divided into seven categories: height, energy, peaks, understory, percentiles, Gaussian decomposition, and others (see Table 4). Lastly, nDSM-derived canopy texture metrics ( As a result, a set of metrics from the four data sets was available for the classification into six classes according to species composition, dominance based on stem density and basal area, and understory vegetation cover.…”

“…In particular, discrete Airborne Laser Scanning (ALS D ) has become an efficient tool for registering information from height distributions in forest stands (Zaldo et al, 2010;Cao et al, 2014;Ruiz et al, 2016). However, ALS D data have restrictions to register different vegetation layers (Crespo-Peremarch et al, 2016).…”

“…In particular, discrete Airborne Laser Scanning (ALS D ) has become an efficient tool for registering information from height distributions in forest stands (Zaldo et al, 2010;Cao et al, 2014;Ruiz et al, 2016). However, ALS D data have restrictions to register different vegetation layers (Crespo-Peremarch et al, 2016). ALS FW tries to fill this gap given that it is a system that registers the complete signal emitted by the sensor and whose point density is much higher (Heinzel and Koch, 2011), being capable of describing the physical properties of the intercepted objects (Ruiz et al, 2014).…”

A Full-Waveform Airborne Laser Scanning Metric Extraction Tool for Forest Structure Modelling. Do Scan Angle and Radiometric Correction Matter?

“…Among other uses, ALS FW data have also been used to estimate forest stand variables [31], forest structure and fuel models [30], to segment trees [32], and to classify tree species [33,34]. Hence, given that ALS FW provides more information from the different vertical layers compared to ALS D , it has a great potential to study the forest structure [31,35,36] and understory vegetation [37,38]. Since ALS D and ALS FW have different data structures (i.e., based on a point cloud and waveforms, respectively), new ALS FW metrics have been proposed in the last decade to carry out the studies mentioned above.…”

Analysis of the Side-Lap Effect on Full-Waveform Lidar Data Acquisition for the Estimation of Forest Structure Variables

“…Eight LiDAR full-waveform metrics were extracted following those proposed by Duong (2010) and further described by Cao et al (2014), and used as independent variables in regression models (Crespo-Peremarch et al, 2016): Height of median energy (HOME), Waveform distance (WD), Number of peaks (NP), Roughness of outermost canopy (ROUGH), Height/median ratio (HTMR), Vertical distribution ratio (VDR), Return of waveform energy (RWE) and Front slope angle (FS). Nine forest structure and canopy fuel variables were employed as dependent variables in regression models (see Table 1).…”