FOREMA Quarterly Colloquia

formerly RU Colloquium


The research unit (RU) “Forest Resources and Management” holds a colloquium four times a year. Besides an update on organisational news, this bilingual event allows the exchange of cutting edge forest research, useful skills and advantageous techniques used for excellent science.

Research notes

Abstract of presentations given at the FOREMA Quarterly Colloquia


Daniel Kükenbrink: Progress towards an operational inclusion of close range remote sensing in the Swiss NFI.

National forest inventories (NFIs) are an important source of information to assess the state and dynamics of forest ecosystems. Traditional statistical forest inventory procedures are labour intensive, while the actual coverage of evaluated plot area is limited. Many forest parameters rely on assessments based on expert knowledge. These expert assessments can be subject to observer bias and hence may lack in robustness. Close range remote sensing (CRS) shows large potential for quantitative forest assessment. With recent sensor developments, the possibilities for an operational inclusion of CRS within an NFI increased. Especially mobile laser scanning (MLS) showed good performance both in terms of acquisition time and level of accuracy for the derivation of forest inventory relevant parameters (e.g. tree position, diameter at breast height (DBH)). However, extensive analysis on the robustness of data acquisition and feature extraction is needed to evaluate their suitability for an operational inclusion within the framework of an NFI.

In this presentation I will talk about the progress on the operational inclusion of CRS for the next cycle of the Swiss NFI. I will present most recent results on the robustness of CRS acquisitions as well as the next steps to be taken to assure a successful implementation of MLS in the Swiss NFI. Finally, I will talk about the potential applications and benefits of acquiring 3D point cloud data within an NFI.

Marco Ferretti: ICP Forests: facts, figures, and science.

European forests represent an invaluable asset for wood production, biodiversity conservation and to combat climate change. Concern for the possible effects of transboundary air pollution on ecosystems, health and materials prompted the United Nations to establish the Air Convention in 1979 – the first international multi-lateral treaty to protect the environment. Within the Air Convention, the International Co-operative Programme for the Assessment and Monitoring of Air Pollution Effects on Forests (ICP Forests) was launched in 1985. The objectives were, and largely remain, to monitor spatio-temporal changes in forest condition across the UNECE region and to identify driver-response relationships. The methods adopted (i) cover from plot design to field and laboratory measurements, with comprehensive Quality Assurance/Quality Control; (ii) include a multi-level monitoring concept, with a probabilistic component (ca. 6000 Level I plots) and an intensive component (ca. 600 Level II plots); (iii) were all internationally agreed; and (iv) are continuously updated. Such a monitoring system remained internationally unrivalled to detect the impact of climate extremes and the combined effects of air pollution, climate change, forest management and biotic factors on key ecological features and processes of European forests. Here I will present the ICP Forests program together with some results.

Francesco Rota: Nationwide forest edge structure characteristics derived from multitemporal airborne laser scanning data

Soil properties influence plant physiology and growth, playing a fundamental role in shaping species niches in forest ecosystems. Here, we investigated the impact of soil data quality on the performance of climate-topography species distribution models (SDMs) of temperate forest woody plants. We compared models based on measured soil properties with those based on digitally mapped soil properties at different spatial resolutions (25m and 250m). We first calibrated SDMs with measured soil data and plant species presences and absences from plots in mature temperate forest stands. Then, we developed models using the same soil predictors, but extracted from digital soil maps at the nearest neighboring plots of the Swiss National Forestry Inventory. Our approach enabled a comprehensive assessment of the significance of soil data quality for 41 Swiss forest woody plant species. The predictive power of SDMs without soil information compared to those with soil information, as well as those with measured vs digitally mapped soil information at different spatial resolutions was evaluated with metrics of model performance and variable contribution. On average, performance of models with measured and digitally mapped soil properties was significantly improved over those without soil information. SDMs based on measured and high-resolution soil maps showed a higher performance, especially for species with an ‘extreme’ niche position (e.g. preference for high or low pH), compared to those using coarse-resolution (250m) soil information. Nevertheless, globally available soil maps can provide important predictors if no high-resolution soil maps are available. Moreover, among the tested soil predictors,  pH and clay content of the topsoil layers improved the predictive power of SDMs for forest woody plants the most. Such improved model performance informs biodiversity modelling about the relevance of soil data quality in SDMs for species of temperate forest ecosystems. In conclusion, the incorporation of accurate soil information into SDMs becomes indispensable for making well-informed forecasts for guiding decisions in forest management, also when addressing the potential distribution shifts of woody plant species due to climate change.


Moritz Bruggisser: Nationwide forest edge structure characteristics derived from multitemporal airborne laser scanning data

Forest edges represent the transition zone between the forest interior and the open countryside and provide several ecological functions. They contribute to biodiversity as habitats for plant and animal species, regulate fluxes of nutrients and pollutants between surrounding agricultural areas and the forest, or regulate the microclimate. To ensure maintenance of these edge functions, a reliable and frequent assessment and monitoring of their conditions is required, particularly for degraded edges which should be restored.

Regular, nationwide, publicly available airborne laser scanning (ALS) acquisitions (15-20 points/m2) as scheduled for Switzerland in the future are ideal to monitor forest edge conservation and renaturation efforts. To this means, we provided a map of forest edge structure characteristics for entire Switzerland (total forest edge length ~187,000 km). We derived canopy height variability, presence or absence of the shrub layer and forest edge slope from the latest nationwide ALS acquisition, thus information about the horizontal and vertical structure in the forest edge zone. These ALS structure metrics were chosen such that they are as close as possible to the definition of the forest edge metrics which are collected as part of the Swiss National Forest Inventory (NFI). Additionally, we derived light regime information and the horizontal visibility into the forest interior which are two metrics that go beyond standard parameters.

However, acquisition characteristics of different ALS flight campaigns differ regarding flight plans, time of acquisition, and sensor specifications. Variations in these acquisition characteristics result in differences in the point clouds (point densities and height distribution). Therefore, we put a particular emphasis on the evaluation of the robustness of the derived ALS metrics. The goal of our study is to verify that the detected differences in the ALS structure metrics reflect real structure changes in the forest edge instead of different acquisition characteristics.

Nataliia Rehush: TreeDetector – a fast and accurate approach for tree position and diameter retrieval from terrestrial LiDAR data

Diameter at breast height (DBH) serves as a key attribute for a qualitative and quantitative evaluation of a tree and a forest, in general. Thus, DBH belongs to the most frequently derived tree attributes in forestry and is included in the measurement protocols of every national forest inventory (NFI). Recently, terrestrial LiDAR data have been actively investigated to extract various forest characteristics, including DBH. So far, various approaches have been developed. However, they often require extensive preprocessing and filtering of the point cloud and rely on expert-based iterative parameter optimisation. This results in a low level of automation and high time consumption. Recent advances in the field of computer vision show that deep learning methods are often highly efficient in extracting meaningful features from complex real-world data. 

We propose a fully automated deep learning-based approach for a fast and accurate tree position and tree diameter retrieval from dense terrestrial LiDAR data. The approach relies on plot-level point clouds and requires very little data pre-processing. The backbone of the approach is an existing deep learning model (YOLOv5) retrained on our data. In addition to real-world terrestrial (TLS) and mobile (MLS) laser scanning data, we enhanced our training dataset with synthetic point clouds of simulated forest stands. The final model was tested on a “blind” dataset that included TLS and MLS point clouds with available reference data (DBH and tree position measured by experts during the field survey using calipers and a total station). 

The results reveal a high performance of our approach in both speed and accuracy. The processing time for a forest plot was reduced to several minutes and a high tree detection rate was achieved. Thus, our approach can be used as an efficient tool for tree position and tree diameter retrieval from terrestrial LiDAR data.

Peter Stoll: Sampling with Partial Replacement in the Swiss National Forest Inventory (NFI 6)?

National forest inventories (NFI) need to be cautiously adapted to new needs without jeopardizing the continuity of long time series. Up to NFI 5 (2018 - 2026), terrestrial forest sample plots in Switzerland were permanent. Permanently marked plots have been used since NFI 1 (1983 - 1985) raising questions about the representativeness of the original sample. Sampling with partial replacement (SPR) has repeatedly been proposed as alternative in order to ensure that samples remain representative. Moreover, replacing some permanent with temporary plots may make it easier to improve public accessibility of plot positions.

However, SPR was never adopted in Switzerland because of a potentially substantial loss in precision of change estimates and the added complexity in managing, handling and the estimation steps of the inventory.

Using re-sampling of Swiss NFI data and simulations, we found that 30 or even up to 50% of the permanent plots could be replaced by temporary plots without losing much precision in change estimates. On the other hand, bias of change estimates increased with increasing fraction of replaced plots and an anticipated gain in precision of state estimates was marginal.

Cautious recommendation: 10% of the permanent plots could be replaced by temporary plots to see how representative the original sample still is without jeopardizing the continuity of a long (> 40 years) time series. Even better would be to add 10% temporary plots to the existing set of permanent plots and thoroughly evaluate potential consequences of sacrificing 10% permanent plots after completion of NFI 6.


Meinrad Abegg: Challenges and Opportunities in Retrieving Forest Allometric Traits Using Close Range Sensing for Allometric Modeling

Allometric models play a crucial role in forest monitoring, e.g., converting diameter measurements into biomass. Many studies show the potential of close-range remote sensing (CRS) technologies in forest ecosystem assessment. However, allometric models are still based on traditional measurements, and there is no established procedure to include CRS in developing new allometric models. Within the Cost Action 3DForEcoTech, a wiki will be developed to document the challenges and opportunities in collecting and analyzing forest allometric data using different CRS techniques for comparing with classical approaches and develop new advanced models.
The wiki's main objective is to provide practitioners with a hands-on guideline on the steps to be taken combining existing allometric models with CRS technologies. The wiki contains a general introduction to approaching a CRS survey. Additionally, an overview of data, collections and models for sharing are compiled to facilitate collaboration. Specific information about scanners, including TLS, MLS, photogrammetric, and UAVLS point clouds, is documented and reviewed in detail.
The link between CRS features and scanner types provides a comprehensive overview of the potential connections between the two. Firstly, traditional allometries are evaluated regarding input datasets, tree components or statistics considerations and goodness of fit statistics. Secondly, the way forward with CRS data is discussed in terms of sample types, ground truth, and new CRS input data, emphasizing the importance of explanatory variables and their connection with the acquisition approach.
Validation considerations are also discussed in terms of the conditional use case. This open document will support future internationally standardized forest data collection and model development for a broad range of purposes.

Hristina Hristova: Viewing the Forest in 3D: How Stereo Spherical Videos Enable Low-Cost Reconstruction of Forest Plots

Understanding and monitoring the surrounding environment depend increasingly on its 3D representation. However, the high costs of 3D data equipment limit its wide usage. At the same time, low-cost sensor solutions have increased in popularity thanks to their lightweight setup and relatively user-friendly interfaces. Here, we present a novel low-cost approach for 3D point cloud reconstruction based on stereo spherical videos. Traditionally, visually coded targets, set up during data acquisition, provide the scale of a point cloud built from a monocular video. By introducing a stereo-video system featuring two spherical cameras and a known baseline distance, the proposed approach eliminates the need for visually coded targets and reduces manual work. We studied the ability of our approach to generate 3D point clouds in a highly structured forest environment represented by three 50m × 50m forest plots. We assessed the quality of the generated videogrammetric point clouds using three baseline distances, i.e., 20, 40, and 60 cm, with Terrestrial Laser Scanning (TLS) data used as a reference. Our results show that the proposed approach allows for feasible 3D reconstruction of complex forest plots. A baseline of 60 cm enabled a correct detection of more than 65% of the trees within the forest plots, producing an average tree position error between 30 cm and 50 cm and clearly outperforming a monocular-video approach. A Multiscale Model-to-Model Cloud (M3C2) analysis provides signed distances between the generated point cloud and the reference TLS data with zero mean and 1 m standard deviation. In this paper, we reveal the high potential of the proposed stereo-video approach and its advantages over monocular videogrammetry, for accurately recovering 3D representations of complex forest plots and retrieving key forest features.


Petia Nikolova: Das Projekt «Gebirgswaldverjüngung» stellt sich vor

Das Projekt «Gebirgswaldverjüngung» hat das langfristige Ziel, innert ca. 20 Jahren fachliche Grund­lagen zu schaffen, damit die Naturverjüngung im Gebirgswald wirksam waldbaulich gefördert werden kann. Zudem sollen praxistaugliche Hilfen zur Beurteilung der Na­turverjüngung bereitgestellt werden. In der Projektphase II (2020-2023) werden 10 Versuchsflä­chen in Fichten- und Fichten-Tannenwäldern fertig eingerichtet, dort waldbauliche Langzeit-Experimente gestartet und aus wiederholten Inventuren von Bestand und Verjüngung Ergebnisse zur Entwicklung der Naturverjüngung gewonnen. Ziel des Vortrags ist, den aktuellen Stand des Projektes vorzustellen und Interessenten für wissenschaftliche Kooperationen zu gewinnen. Es werden Ideen zu möglichen Begleitstudien zu Teilprozessen der Verjüngung vorgestellt.

Daniel Scherrer: Berechnung der Naturnähe der Baumartenmischung auf LFI-Probeflächen

Viele waldbauliche Beurteilungen (z.B. NaiS) und subsequente waldbauliche Massnahmen oder Bewirtschaftungsformen (z.B. Naturnaher Waldbau) basieren auf der Hypothese, dass naturnahe Wälder eine ‘optimale’ Resistenz und Resilienz gegenüber Störungen aufweisen. Ist dies wirklich der Fall? Die repräsentativen Daten des LFI und die neu verfügbaren Angaben zu den Waldstandorttypen liefern erstmals eine optimale Datenbasis dies zu überprüfen. Wie wirkt sich der gegenwärtige und zukünftige Klimawandel auf die Naturnähe von Wäldern aus? Was sind die grössten Herausforderungen, um die Wälder auch in den kommenden 100 Jahren naturnah zu gestalten? Wo besteht der grösste Handlungsbedarf für waldbauliche Eingriffe zur Anpassung der Baumartenzusammensetzung, um eine hohe Naturnähe und kontinuierliche prioritäre Waldfunktionen sicherzustellen?


Moritz Bruggisser, Zuyuan Wang, Lars T. Waser, Christian Ginzler - Assessing forest edge structure on national scale for spatial and temporal edge quality change assessments

Forest edges represent the transition zone between open land and the forest interior and fulfil important ecological functions, e.g., for the conservation of biodiversity or the regulation of the microclimate. However, forest edges are under increasing pressure due to intensified agriculture and forest management. In order to support actions for the functioning of forest edges through management activities, assessing the state and changes of forest edges is required.
The aim of our project is to assess horizontal and vertical structure variations of forest edges at the national coverage of Switzerland (41,285 km2).
We use the latest freely available nationwide airborne laser scanning (ALS) based SwissSurface3D model with a point density of 15-20 points/m2, and if available also additional very dense, multitemporal ALS data. The high point densities enable to assess both horizontal and vertical structure at the forest edges. To do so, we compute the most relevant forest edge structure features, including vegetation height distributions, the shrub layer and herbaceous boarders, number of vegetation layers, vegetation height gradient from open land to forest interior, horizontal canopy information such as canopy cover and canopy gaps, vertical density of the canopy layer within the forest edge area, and stem density.
Based on a previous study (Wang et al. 2020), we proposed a method using non-parametric Kernel Density Estimation (KDE) to describe forest edges from vegetation height model data. Since this KDE-based approach is independent from specific site conditions, we adapted this approach by implementing additional information on land-use intensity close to the forest, topographic information and structure of the forest interior. This enables to automatically quantify and assess structural changes of forest edges at national coverage and understand drivers of forest edge structure variability. Moreover, our analysis also includes the quantification of changes of the forest edges in areas where multi-temporal ALS data sets are available.
The proposed approach provides all required information to identify forest edge areas at a national coverage, which are at severe threat of degradation and thus require management intervention. In this way, the quality of forest edges can be preserved or improved for important ecosystem services.


Bronwyn Price - Spatially explicit models of forest attributes for Switzerland: Do we need vegetation height models?

Spatially explicit national-extent maps of forest attributes such as above ground biomass (AGB), basal area and stem density are invaluable information sources for a variety of purposes including carbon accounting, biodiversity assessment and forestry planning. The increasing availability of high spatial and temporal resolution earth observation data at broad spatial extents (regional-national-continental-global), in combination with national forest inventory field data, offers opportunities to develop generalised methods and produce high resolution forest attribute maps at national extents. While high resolution vegetation height models (VHMs), derived from either airborne LiDAR data or digital photogrammetry, have been shown to be very good predictors of forest attributes such as AGB, acquisition of these data can be costly and as such they are not always available at national extents and are often not up to date. Thus, attention has also turned to modelling forest attributes from openly available datasets such as the Copernicus Sentinel-1/-2 (S1, S2) missions.
In this work we compare models of forest attributes based on the high resolution VHM of Switzerland (derived from stereo aerial imagery) to those based on a coarser resolution S2 VHM, and models with no VHM but metrics derived from S2 imagery (growing season metrics of vegetation indices and image texture). NFI field plot data allows us to model the forest attributes AGB, growing stock, number of stems per hectare and basal area, in a spatially explicit manner across Swiss forests at 25m resolution (equivalent to NFI plot size). Additional explanatory variables include the dominant leaf type dataset (also derived from S1/S2 data), soil properties derived from digital soil modelling, climatic variables, and topographical variables.
For each forest attribute, first we pre-select a subset of best predictors, assessing individual predictive power of each explanatory variable in a repeated split sample cross-validation, with pairwise exclusion of collinear variables. A variety of modelling approaches were tested: linear modelling, generalized linear models (GLM), random forest (RF) and boosted regression. Random forest models were the best performing model type across all response variables. The predictive power of the best model for stems per hectare remains relatively low, with approximately 39% variance explained by a RF model, but was reasonable for all other target variables (61% - 68% variance explained).  The explanatory power of the RF models is considerably higher for the NFI VHM dependant models than for those based on the coarser resolution S2 VHM, which in turn were considerably higher than models with no VHM variables. However, the additional non-VHM variables offer substantial improvements in explanatory power as the performance of models without these variables was markedly lower.

Tiziana Koch - Assessing variations in the phenology of tree species in Swiss forests with Sentinel-2

Tree phenology, the study of cyclical biological events of trees, including the timing of leaf unfolding and leaf fall, is closely linked to climatic conditions. Multi-temporal satellite remote sensing allows extensive mapping of phenological information that due to its operating scale relates to land surface phenology representing the phenology of many species. Nowadays with the operating scale of Sentinel-2 species-specific phenology can be derived. We hypothesize that there is great potential to enhance phenological monitoring through a more detailed analysis of the inter- and intra-species specific variation of phenology from remote sensing data.
Our study focuses on forest ecosystems of Switzerland with its complex topography and environmental gradients (elevations between 193 and 4,634 m.a.s.l.), causing significantly varying phenological patterns of individual tree species across six main biogeographical regions in the country with an area of 41,285 km2. We link terrestrial data of tree species from the Swiss National Forest Inventory (NFI) with Sentinel-2 time series for the year 2019 and analyze the variations of these tree species in a spatial, temporal and spectral context. Due to the higher abundance and appearance on 95 % of the NFI plots, we limit our assessment to Beech (Fagus sylvatica), Ash (Fraxinus excelsior), Chestnut (Castanea sativa), Larch (Larix decidua, Larix kaempferi), Norway spruce (Picea abies), Scots pine (Pinus sylvestris) and Silver fir (Abies alba).
Our results can potentially serve as a new baseline for extensive phenological observations of forest tree species and can be used as a benchmark for large-scale tree species classifications. The knowledge about phenological patterns may further be used to adjust forest management strategies to optimize timber availability, biodiversity richness, or carbon storage.


Ross Shackleton - SwissAIM: An update on progress

Due to current rapid social-ecological changes, there is a need to update forest monitoring and inventorying initiatives. To address this need, the SwissAIM (an Advanced Inventorying and Monitoring System for Swiss Forests) initiative was established in 2020 through an initial launch workshop. The SwissAIM initiative aims to develop an “integrated terrestrial and remote sensing observation system based on a permanent panel of enhanced NFI plots. It will provide high-quality periodical (intra-annual, annual, multi-annual) results with known statistical errors for the status, change and response of forests to biotic and abiotic drivers”.

This presentation overviews the progress in the development of SwissAIM to date. In particular, some of the 92 questions and 52 variables of interest suggested for SwissAIM will be discussed. Broadly speaking we highlight that SwissAIM would better contribute to assessing the status and change, processes and dynamics of swiss forests and advancing forest monitoring tools and techniques.

A preliminary plot design for SwissAIM will also be presented. This design is based largely on the current Swiss NFI plot design, however, with many novel additions for SwissAIM. In particular, these novel additions relate to initiating ground vegetation surveys on the central circular plot of the Swiss NFI, adding new plots for gap regeneration assessments, the addition of terrestrial laser scanning and drone-based measures with thin the Swiss NFI interpretation area, and the inclusion of five new sub-plots in the interpretation area for minimally destructive and invasive sampling. In these five sub-plots, new measures relate tree and soil samples for nutrient and eDNA analysis, the installation of conspicuous and underground sensors and litter sampling.

Furthermore, we present the results of the power analyses we conducted to identify required sample sizes (n) to assess forest change. We used three variables to do this. To assess change at a power level of 80% and an alpha value of 0.05, 157 plots would be needed to assess multi-annual changes in species richness, 245 for annual changes in the basal area, and 403 plots for assessing annual changes in defoliation. Fewer plots would be needed to assess the stats of forests and for modelling. Based on this SwissAIM should be conducted on an 8x8 km grid or denser.

We conclude by discussing future directions for the SwissAIM initiative, in particular, starting engagements with relevant stakeholders across Switzerland and prototyping the plot design and methods.

Roman Flury - Spatial Feature Identification to Evaluate Biodiversity Indices on Different Scales

Dominant features of spatial data are connected structures or patterns that emerge from location-based variation and manifest at specific scales. A sequential application of scale-space multiresolution decomposition and covariance function estimation can be used to identify such dominant features. The decomposition separates data into additive components and enables the recognition of their dominant features. A general scale-space multiresolution decomposition method is developed for different spatial data types, where the underlying model includes a precision and spatial-weight matrix to capture spatial correlation. The data are separated into their components by smoothing on different scales, such that larger scales have longer spatial correlation ranges. Covariance function estimation can be used to describe attributes in spatial data. Therefore, such functions are estimated for each component to determine, e.g., its effective range, which assesses the width-extent of the dominant feature. Finally, Bayesian analysis enables the inference of identified dominant features and to judge whether these are credibly different. The efficient implementation of the method relies mainly on a sparse-matrix data structure and algorithms. This method can lead to new insights in disciplines that use spatial data, as exemplified by identifying the dominant features in a forest dataset. In that application, the width-extents of the dominant features have an ecological interpretation, namely the species interaction range, and their estimates support the derivation of ecosystem properties such as biodiversity indices.

For more information see:

Zuyuan Wang - From historical aerial images to informative woody plant changes in long term mountain treeline ecotones research using CNNs

Assessing changes of woody plants is a prerequisite for monitoring environmental resources as well as characterizing treeline dynamic.

In this context, a broad range of methods have been developed over the last decades that include dendro-ecological techniques, aerial photograph assessments, field surveys, remote sensing data integrated historical maps, etc.

Long-term dynamics of up to 30 years at coarse spatial resolution have been assessed using time series obtained from MODIS and Landsat data.

Meanwhile, historical black-and-white (B&W) aerial images have been recognized as an important source of information for assessing vegetation dynamics with high spatial resolution – although their use is still considered to be relatively limited.

In Switzerland, the archive of historical B&W aerial images goes back to the 1920ies. The Federal Office of Topography (swisstopo) scanned and oriented the data sets from nationwide flights starting from  1979. With these data sets a continuous assessment of vegetation cover change over an increasing time span becomes feasible.

In our previous study (Wang et al. 2022), we used a deep learning approach to detect changes in woody vegetation along alpine treeline ecotones between 1980 and 2010 for two biogeographic areas in the Swiss Alps using the three vegetation structure classes “dense forest”, “group of trees” and “others”. A comparison with visual image interpretation revealed generally high agreement for the class “dense forest” and lower agreement for the class “group of trees”. The use of single-band B&W images was challenging for images from all dates due to the heterogeneity of the image patches, i.e. spectral distortion, brightness and contrast. Another potential challenge was related to the fuzzy boundary for the definition of the three classes.

The present study builds up on these findings and aims to assess woody plant changes in treeline ecotones for 170 sample plot areas (1km×1km) distributed over the whole Switzerland. The main achievements are: 1) adopting suitable data augmentation as an effective method of supplementing the training samples, 2) modification of the CNN models to include short-range, mid-range and long-range semantic information at different stages of the network in order to generate a robust feature descriptor, which preserves the spatial details. Moreover, heterogeneous instead of homogeneous transfer leaning is applied in order to incorporating human knowledge by correlating different feature spaces.

With this new method negative performance is quantitatively reduced, including a discriminator that assigns different weights and uses an iterative processing, in order to detect the negative transfer.

The study reveals that archived historical B&W images in combination with CNNs have great potential to assess changes in woody vegetation cover along alpine treeline ecotones at the national coverage of Switzerland over the last decades.

For more information see:

16.6.2022 (FOREMA day at Waldlabor)

Hristina Hristova - Estimating distance to trees in NFI 360-degree images

Knowledge about forest stand characteristics (i.e., spatial distributions of trees, tree size distributions, etc.) is crucial for forest management and monitoring, evaluating the protective function of forests, etc. These features are traditionally derived by manual field sampling performed by the NFI. Except for being time-consuming, field sampling can also be expensive when carried out in a large forest area.
Deriving forest stand features for large areas is essential to properly monitor the forest’s current state, allowing for knowledge-based management decisions. The collection of forest data could be aided by close-range remote sensing techniques, such as Terrestrial Laser Scanning (TLS) and imaging. With the use of 360° imagery, efforts can be made to estimate tree features, such as distance, diameter, and height. Recently the interest in 3D reconstruction from monocular images has increased thanks to the introduction of deep networks. Monocular depth recovery benefits from the newest deep learning (DL) architectures and provides an alternative to traditional stereo approaches.
In our work, we aim to recover the absolute distance from the camera to detectable trees in the NFI 360° forest images. We achieve it by training our own DL model on this set of 360° forest images. We first show that related work fails to produce plausible results for forest imagery. Reference methods are trained on large data sets and achieve outstanding precision on the corresponding test sets. Despite that, their training sets contain a limited number of forest images which may hinder their performance.
We consider the specifics of the forest environment and propose a novel framework for depth prediction in forest images. Our method recovers the forest 3D structure from an input RGB image and a sparse depth channel. We utilize high-resolution TLS to collect the training data for our model and compute ground-truth depths. To capture the initial high-quality 360° RGB-D images in various NFI plots, we collected data using a FARO Focus 3D 120S terrestrial laser scanner with an integrated camera. We captured the RGB data using the integrated camera and derived the depth channel D from the point cloud data resulting from the scanning. After the data collection, we de-noised the point clouds by removing unstable objects, such as leaves and thin branches. We then merged the depth information with the RGB images from the integrated camera to create the initial 360° RGB-D images comprising our forest data set.
Our proposed model takes forest (360°) RGB-D images as input, where the depth channel is a sparse sampling of the ground-truth depth. For feature extraction, we trained ResNet-18 with a loss function that depends on the L1 depth distance, gradient loss, and normal loss. The model was trained for 120 epochs with an adaptive learning rate with an initial value of 0.1. Our analysis shows that our depth maps are visually the closest to the ground truth and our predictions are highly correlated to the true depth, with an absolute error of 0.89 m for 1000 input samples.

Esther Thürig and Aline Bornand - How much biomass does a tree contain? - Truth versus model

How much biomass does a tree contain? - Truth versus model

In times of climate and energy change, the sequestration of carbon (C) and the provision of biomass are important ecosystem services (ES) of the forest. At WSL, much data is available to estimate stemwood volume, but no ground-truth data is available to directly model total aboveground biomass on individual trees. Moreover, new measurement methods such as laser technology offer previously unattainable possibilities for modelling above-ground biomass.
In the "Swiss Biomass" project, both existing and new, innovative measurement methods are being evaluated in order to fill these data gaps. The aim is to be able to determine the biomass of individual trees more precisely using non-destructive measurement methods. These biomass estimates are central for the greenhouse gas inventory, Kyoto reporting and for the estimation of wood resources and bioenergy potentials.
In an internally WSL-funded pilot project, different measurement methods were tested and evaluated using two stands. The old-growth stand was completely surveyed using the non-destructive measurement methods Lidar (air and ground-based) and traditional forest surveys. During felling, about 36 trees in each stand were measured lying section by section, the weight of the whole biomass was determined with a crane scale, and 5 stem slices were taken from each. The stem slices were weighed (fresh weight), brought to the WSL and further processed there to determine the fresh volume, dry volume and dry weight.
A PhD study evaluates several approaches for estimating above-ground tree volume, using destructive measurements as a ground-truth reference and ground-based lidar point cloud data as an explanatory basis. One approach derives geometric tree attributes directly from the point cloud, e.g. crown diameter and length or convex hull volume. Another approach extracts stem diameters at regular height intervals from the point cloud and constructs stem regeneration curves for each tree. A more complex approach is called quantitative structural model (QSM) and involves reconstructing the entire morphological structure of a given tree with cylindrical shapes. Given a complete point cloud, this allows a direct and detailed estimate of the total volume.
Since the end of 2021, the project has been in the second phase. Six more forest stands are being sampled, financed by the FOEN. The first results are expected towards the end of 2023.

Peter Brang, Kathrin Streit - Experimental plantations of tree species adapted to future climates

A warmer climate with drier summers will affect the climatic suitability of tree species in their current habitat and, subsequently, future ecosystem services. Forest managers have started to respond with assisted migration and assisted gene flow, but important fundamentals are lacking for their implementation. Therefore, the Experimental plantation project was initiated in 2018 to find out which factors determine mortality, health and growth of tree species and provenances along large environmental gradients. In its frame a network of 59 experimental plantations is established throughout Switzerland as a joint effort of WSL, FOEN, cantonal forest services, forest managers and forest owners. 18 tree species are tested, including the most important tree species like silver fir, Norway spruce, European beech, oaks and maples. For each species, 7 provenances are tested, with seed from Swiss populations, but with some seed sources originating from the dry-warm edge of their range. At present, 40 plantations have been established.


Beate Kittl - What's new(s)? How your research makes the headlines

What's new(s)? WSL media officer Beate Kittl explained what a WSL news is and how your research finds its way to the greater public. What are the criteria? What should you do? Find all the information on the intranet (for WSL members only).

Florian Zellweger - Remote sensing advances forest microclimate mapping across Switzerland

Forest microclimates contrast strongly with the climate outside forests. To fully understand and better predict how forests' functions and biodiversity relate to climate and climate change, microclimates need to be integrated into ecological research. We have set up a forest temperature network across Switzerland and combine our measurements with latest remote sensing data to predict forest microclimates for the past, present and future. Preliminary results show that the structure of the canopy and the related radiation regimes play a key role in driving small scale variation in forest temperatures close to the ground and in the topsoil. Our approach can now be used to estimate the effects of different forest management and disturbance regimes on microclimates, which helps to better understand how our forests will be affected by the combined effect

Dataset Responsibles