Statistical analyses and techniques for forest volume estimation and stand structural classification using airborne laser scanner data
Author(s)
Torresan, Chiara
Date Issued
June 24, 2014
Type
Doctoral Thesis
Abstract
Forests represent an important environmental and economic resource and an accurate knowledge of the amount of timber volume and their structural features allows efficient forest management strategies on which decision making processes depend.
Airborne laser scanning (ALS) is widely used for the retrieval of detailed information on height and cover of forests at different levels, whether for individual trees, plots, stands or on a nationwide forest scale. Coupling aerial laser scanner data with inventory data provides information on timber volume and forest structural characterization, and improves the precision of forest inventory estimates.
This thesis attempted to provide specific solutions to some problems and objectives encountered by the Forest Service of Trento Province (Northern East of Italy) pertaining to both forest volume estimation at the forest compartment level and area of classification of structure of forest stands.
Volume estimation at stratum level can be obtained from field plots located within, on the contrary volume estimation in portions of strata may be problematic due to the small number or even the absence of plots in those portions. But if a Canopy Height Model (CHM) is available for the whole area under planning, height at pixel level can be used as auxiliary information. A ratio model presuming a proportional relationship between heights and volumes at pixel level was adopted to lead to an estimation. From this model, the volume within any portion was estimated as the proportionality factor estimate multiplied by the total of heights in that portion. This volume estimator was considered from the model-based, design-based and hybrid perspectives, and variances and their estimator were derived under the three approaches. Volume estimator and the model-based, design-based, and hybrid variance estimators were checked by a simulation study performed on a real forest in Val di Sella (South-Eastern of Trento Province), considered as a fixed population in which plots were randomly allocated at each simulation run. Furthermore, volume estimator and the hybrid variance estimator were applied to infer the volume of four forest compartments in the public forest estate of San Martino di Castrozza (North-Eastern of Trento province). The simulation and the case study showed that the bias of the volume estimator derived from the ratio model adopted depends on the similarity of the whole stratum and the portion of stratum in which estimation is performed. In presence of negligible bias, the results from simulation and case study are satisfactory in terms of relative root mean square error and hybrid variance estimates, respectively.
The Forest Service of Trento Province presently uses a structural classification system based on the percentage of basal area in small (or understory), medium (or mid-story), and large (or
overstory) trees. The classification in forest structural types is one of the primary goals along with species composition, for forest managers to identify the strata in the process of forest inventory based on stratified sampling. Moreover, the forest structural classification is used during the decision making process.
This dissertation examined the use of CHM-derived metrics to predict one of the twelve forest structure classes of the classification system utilized by the Forest Service. Two approaches to predict size-based forest classifications were evaluated: in the first, supervised classification with both linear discriminant analysis (LDA) and random forest (RF) was attempted; in the second, basal areas of small, medium, and large trees from CHM-derived variables by k-nearest neighbor imputation (k-NN) and parametric regression were predicted, and then classified observations based on their predicted basal areas. Leave-one-out cross-validation was used to evaluate the ability to predict forest structure classes from CHM data and in the case of prediction-based classification approach the performances in predicting basal area were evaluated. The strategies proved moderately successful, being the LDA the techniques with best performance. In general, the prediction of basal areas of understory and overstory trees was the most successful.
An additional study was carried out to assess whether metrics extracted this time from lidar raw point cloud can be exploited to discriminate different forest types through machine learning techniques such as classification trees. This study suggests that the proportion of basal area in diameter size-classes is a valuable parameter to distinguish between different forest structural clusters through clustering analysis. The five airborne laser scanner metrics selected for this study were able to confirm that at least one of the five forest structure patterns identified by the clustering analysis (pole-stage, young, adult, mature, and old forests) is different from the other patterns. The classification tree model built into this study provided moderately satisfactory results in term of classification performance. The binary tree representing the model is clearly interpretable and easily usable by forest technicians in operational contexts.
This dissertation contributed to the body of knowledge that is related to the application of statistical analyses and techniques for forest resources inventory using airborne laser scanner data. Nevertheless, more work is suggested. Future research emphasis could be on how reflectivity from a laser of different wavelengths could be used to discriminate between different types of forest cover. Moreover fusion between lidar data and other image sources could be investigated.
Additional information
Dottorato di ricerca in Ecologia forestale
File(s)![Thumbnail Image]()
Name
ctorresan_tesid.pdf
Size
4.85 MB
Format
Adobe PDF
Checksum (MD5)
1251523237e7abefd6dde515413ebd96