Many real-world modeling problems are hindered by limited data availability. In such cases, transfer learning leverages related source domains to improve predictions in a target domain of interest. We extend the classical gradient tree boosting paradigm to a regression transfer algorithm by modeling the weak learner as a sum of two regression trees. The trees are fitted on source data and target data, respectively, and jointly optimized for the target data. We derive optimal coefficients for the model update under the least-squares, the least-absolute-deviation, and the Huber loss functions. We benchmark our approach against boosting-based regression transfer methods in twelve transfer scenarios. The results indicate that our approach constitutes a competitive alternative within the realm of boosting-based regression transfer. Moreover, we provide a theoretical justification as well as empirical validation that our approach is robust under larger domain shifts.

Remote sensing techniques are widely used for mapping and monitoring forest attributes, providing valuable information on forest cover, biomass, and overall forest health. In recent years, national airborne laser scanning (ALS) campaigns have been conducted in several countries to map forest resources. When combining ALS with field inventory data, these datasets enable the development of nationwide models for prediction of forest attributes. In this study, we explore the potential of machine learning (ML) to enhance existing modeling approaches for nationwide forest attribute mapping in Sweden. We achieve this by relating ALS data from the most recent ALS campaign of Sweden with field data from the Swedish National Forest Inventory (NFI). By aggregating laser metrics from surveyed areas (NFI plots), as well as over surrounding areas to the plots, we investigate (1) if ML approaches can outperform existing linear regression baseline models and (2) if further enhancements of the predictive capacity can be achieved by including surrounding, spatially correlated ALS data. To this end, we used extreme gradient boosting (XGBoost), as well as a convolutional neural network (CNN), specialized to handle tabular data and spatially correlated data, respectively. The models were evaluated on five forest variables: basal-area weighted mean tree height, basal-area weighted mean stem diameter, basal area, stem volume, and above-ground biomass. All models were evaluated on several nested datasets to assess the robustness, showcasing consistent results across datasets. We achieved significant improvements in prediction accuracy across all investigated forest variables. Furthermore, incorporating surrounding information to the modeling rendered further improvements for diameter, basal area, and biomass predictions. The approaches tested and developed here thus form a promising basis for flexible modeling approaches that can be transferred globally for large-scale forest monitoring and management.

The appearance of the log cross-section provides important information when assessing the quality of the log, where properties to consider include pith location and density of annual rings. This makes tasks like estimation of pith location and annual ring detection of great interest. However, creating labeled training data for these tasks can be time-consuming and subject to misjudgments. For this reason, we aim to create generated training data with controlled properties of pith location and amount of annual rings. We propose a two-step generator based on generative adversarial networks in which we can completely avoid manual labeling, not only when generating training data but also during training of the generator itself. This opens up the possibility to train the generator on other types of log end data without the need to manually label new training data. The same method is used to create two generated training datasets; one of entire log ends and one of patches of log ends. To evaluate how the generated data compares to real data, we train two deep learning models to perform estimation of pith location and ring counting, respectively. The models are trained separately on real and generated data and evaluated on real data only. The results show that the performance of both estimation of pith location and ring counting can be improved by replacing real training data with larger sets of generated training data.

Birch has long suffered from a lack of active forest management, leading many researchers to use material without a detailed management history. Data collected from three birch (Betula pendula Roth, B. pubescens Ehrh.) sites in southern Sweden were analyzed using regression analysis to detect any trends or differences in wood properties that could be explained by stand history, tree age and stem form. All sites were genetics trials established in the same way. Estimates of acoustic velocity (AV) from non-destructive testing (NDT) and predicted AV had a higher correlation if data was pooled across sites and other stem form factors were considered. A subsample of stems had radial profiles of X-ray wood density and ring width by year created, and wood density was related to ring number from the pith and ring width. It seemed likely that wood density was negatively related to ring width for both birch species. Linear models had slight improvements if site and species were included, but only the youngest site with trees at age 15 had both birch species. This paper indicated that NDT values need to be considered separately, and any predictive models will likely be improved if they are specific to the site and birch species measured.

This study investigated whether improved downy birch could perform as well as improved silver birch, and whether there was sufficient genetic variation and control for non-destructive testing (NDT) values to include them as selection traits in breeding programs. NDT tools were applied to a 15-year-old downy birch family trial intermixed with improved silver birch. Average diameters, fissured bark height, and grain angle were higher for silver than downy birch. The genetic analysis for downy birch provided estimates of narrow-sense heritability (h2) for acoustic velocity and Pilodyn penetration depth that were above 0.3 but had low genetic variation. Grain angle had relatively high genetic variability (18%) and an h2 of 0.20. A subsample of 49 trees had 4 mm cores x-rayed for wood density estimates, and 34 stems had 12 mm cores macerated for cell measurements. t-tests revealed that average wood density and cell measurements were not significantly different between species. For silver and downy birch, fiber length and vessel length increased between inner and outer measurement positions, and fiber length was reasonably correlated with acoustic velocity. Silver birch tended to have denser and stiffer wood, while downy birch had less rough bark and straighter grain, and these results are in agreement with existing knowledge. The h2 values were similar to those observed in other birch species and indicate there is potential to breed for improved wood density and grain angle in downy birch.

This study investigated heritability of stem and wood traits to improve Swedish silver birch (Betula pendula Roth.) through breeding. Birch is 12% of Sweden’s forest area but mainly used for low value pulp or firewood. This paper applied non-destructive test (NDT) methods, and estimated traits’ heritability (h2), to help breed birch for high value solid wood products. Two trials of 22 families were assessed at age 19 for stem diameter (DBH), stem straightness, rough brown bark height (BH), grain angle (GA), Pilodyn penetration depth (Pilo) and acoustic velocity (AV). X-ray densitometry was performed on a subsample of radial cores taken at 1.3 m from the ground to get an average benchmark density. The h2 values were moderate for GA (0.20 and 0.21) and Pilo (0.53 and 0.48) at the two sites, but the h2 values for AV were low (0.05 and 0.30). There were moderate genotypic correlations between BH and DBH (0.51–0.54). There were low genotypic and phenotypic correlations between NDT measurements and other traits so including NDT in birch breeding efforts should not inadvertently reduce size, stem or wood quality. The high genetic correlations between sites suggest that GA, Pilo and AV values were determined more by genotype than by environment.

Non-destructive testing of wood density, as is already done for Eucalyptus, can allow for earlyselection of birch trees for breeding programmes and stands for harvesting. In this work, external stemquality traits were visually assessed in a family trial of downy (Betula pendula) and silver (B.pubescens) birch in southern Sweden. A subsample of trees was measured for wood density using thePilodyn resistometer portable NDT tool. An X-ray microdensitometric analysis of the subsample oftrees was completed using the Itrax X-ray machine for increment cores taken from the south face,through the pith to the north bark at 1.3 m stem height. The Pearson’s r value for Itrax density andPilodyn density was high (0.580 for downy birch and 0.795 for silver birch), and this correlationmeans Pilodyn should provide a good estimate of average birch wood density. Neither species hadstable wood density values at age 13 and both species’ density increased over time from pith to bark.Ring width influence on stem density was minor or non-existant, and may vary between birch species

In breeding Norway spruce, selection for improved growth and survival is performed at age 10-15 years in order to optimize genetic gain per year. We investigated whether a selection based on wood traits such as density and grain angle, measured under bark in the field at the same age would be informative enough with respect to structural quality traits of sawn boards. To achieve this objective, a sawing study was conducted on the butt logs of 401 trees from a 34-year-old Norway spruce progeny trial situated in southern Sweden. Stem discs were excised from the top of the logs and radial profile data of grain angle, and wood density was recorded for specific annual rings. The sawn and dried boards were assessed for structural traits such as twist, board density, bending stiffness (static modulus of elasticity, sMoE) and bending strength (modulus of rupture, MoR). Additive genetic correlations (r (a)) between single annual ring density measurements and board density, sMoE and MoR were consistently strong (r (a)> 0.7) for annual rings 5-13. Genetic correlations of similar magnitude between grain angle and board twist were estimated for all investigated annual rings (from 2 to around 26 under bark). Consequently, it was found that indirect selection for wood density and grain angle at the tree age 10-16 years would result in more genetic gain per year than selection at later ages. This makes it feasible to perform simultaneous selection of progeny in the field for both growth and wood traits at similar ages.

Thirteen wood parameters were predicted using near infrared (NIR) spectra in the range 780-2380 nm modelled by biorthogonal partial least squares regression. The analysis of parameters and NIR measurements was done on clear-wood samples from the base and midstem of Scots pine (Pinus sylvestris L.) from trees at two sites. Calibrations based oil the measured parameters at seven growth rings (cambial age ranging between 6 and 42 years) could be divided into three groups: (i) the best accuracy was found for longitudinal modulus of elasticity (r > 0.9) followed by bending, compression, and cell length (0.8 < r < 0.9); (ii) microfibril angle, longitudinal hardness, proportion of latewood, and creep with correlations in the range of 0.7-0.8; and (iii) tangential hardness, cell diameter, and cell wall thickness with 0.4 < r < 0.7. It was also shown that juvenile (cambial age : 20 years) and mature wood can be classified using NIR techniques.

In this thesis, two x-ray based dendro-analyses (batch-wise microdensitometry and energy dispersive x-ray fluorescence (EDXRF)) and the conditions under which these methods could be used on the two typical boreal conifers Norway spruce (Picea abies (L.) Karst.) and Scots pine (Pinus sylvestris L.) were evaluated. For density measurements using batch scanning x-ray densitometry, sample preparation and density calibration are vital to acquire high resolution and precision in densitometric measurements. Thickness and alignment should be adapted to give optimal resolution without loss of precision. Samples should be extracted to remove resins before wood density measurements. Ca, Mn, Fe, and Sr and in most cases K and Zn could thus be measured in an efficient way using EDXRF. Significant differences in concentrations between tree rings were found for all of these elements except Sr, indicating that tree ring concentrations of these elements could be correlated to changes in the tree environment. For Mn a correlation between soil pH and Mn concentration in tree rings was found. Other correlations between tree ring element concentration and tree environmental factors were also found. For most elements in tree-rings of Scots pine and Norway spruce it seems that the concentration is due to environmental conditions during the year the tree-ring was developed and several years afterwards. It is concluded that microdensitometry and EDXRF analyses on increment cores are cost-efficient and nondestructive analyses of wood properties. Batch-wise, x-ray based dendrochemical analysis provides opportunities for more property-based use of wood raw material and for environmental monitoring.