Dr. Nataliia Rehush

Close-range remote sensing technologies show great potential to support national forest inventories (NFIs). Mobile laser scanning (MLS) has performed especially well in recent studies, as it is capable of fast data acquisition while still delivering accurate information on NFI-relevant variables (e.g. tree position, diameter at breast height [DBH]). However, the performance of MLS acquisition under challenging conditions (e.g. steep terrain, dense understorey vegetation) has rarely been addressed. Here, on 29 NFI plots distributed over the entire country of Switzerland and representing various terrain and forest conditions, we evaluated the use and performance of handheld MLS acquisition, using a GeoSLAM ZebHorizon laser scanner, to deliver analysis-ready (with minimal occlusion) point clouds and to estimate primary tree variables such as tree position and DBH. Data acquisition was successful for all tested Swiss NFI plots. The average time needed for point cloud acquisition of a 50 × 50 m sample plot was 13.13 min. Over all tested NFI plots, we achieved a tree detection rate of 82.6%, with a relative root mean square error (RMSE) of 7.25%, a mean absolute error of 2.0 cm and a systematic prediction error (SPE) of 1.0 cm for DBH estimation. The terrain gradient had an especially strong influence on the accuracy of DBH estimation. Occlusion, as a metric for point cloud capacity to hold forest structural information for the entire canopy volume, was minimal when the suggested acquisition pattern was followed, at 9% on average for the entire canopy volume. However, the level of occlusion increased in the upper part of the canopy (> 10 m above ground), to 15% on average.
Overall, handheld MLS effectively produced analysis-ready point clouds within the 15-minute acquisition time limitation for a 50 x 50 m area, even under challenging conditions, showing promise for future use in the Swiss NFI. However, the current accuracy of MLS-estimated forest variables like DBH and tree detection is insufficient to replace conventional measurements within an established NFI. Despite this, estimating NFI-relevant variables, even with lower accuracy, and expanding the area of measured tree variables could improve information on forest functions and services (e.g., protection and biodiversity), currently not possible to derive from NFI data due to the limited plot size. In the future, MLS-acquired point clouds could (1) shift from expert-assessed to more quantitative and reproducible variable estimation, and (2) enhance NFI with variables difficult to measure conventionally. This could open new pathways for assessing and monitoring forest structural diversity nationally and provide valuable calibration and validation data for large-scale forest assessments from airborne or space-borne acquisitions.

1. Proximally sensed laser scanning presents new opportunities for automated forest ecosystem data capture. However, a gap remains in deriving ecologically pertinent information, such as tree species, without additional ground data. Artificial intelligence approaches, particularly deep learning (DL), have shown promise towards automation. Progress has been limited by the lack of large, diverse, and, most importantly, openly available labelled single-tree point cloud datasets. This has hindered both (1) the robustness of the DL models across varying data types (platforms and sensors) and (2) the ability to effectively track progress, thereby slowing the convergence towards best practice for species classification.
2. To address the above limitations, we compiled the FOR-species20K benchmark dataset, consisting of individual tree point clouds captured using proximally sensed laser scanning data from terrestrial (TLS), mobile (MLS) and drone laser scanning (ULS). Compiled collaboratively, the dataset includes data collected in forests mainly across Europe, covering Mediterranean, temperate and boreal biogeographic regions. It includes scattered tree data from other continents, totaling over 20,000 trees of 33 species and covering a wide range of tree sizes and forms. Alongside the release of FOR-species20K, we benchmarked seven leading DL models for individual tree species classification, including both point cloud (PointNet++, MinkNet, MLP-Mixer, DGCNNs) and multi-view 2D-based methods (SimpleView, DetailView, YOLOv5).
3. 2D Image-based models had, on average, higher overall accuracy (0.77) than 3D point cloud-based models (0.72). Notably, the performance was consistently >0.8 across scanning platforms and sensors, offering versatility in deployment. The top-scoring model, DetailView, demonstrated robustness to training data imbalances and effectively generalized across tree sizes.
4. The FOR-species20K dataset represents an important asset for developing and benchmarking DL models for individual tree species classification using proximally sensed laser scanning data. As such, it serves as a crucial foundation for future efforts to classify accurately and map tree species at various scales using laser scanning technology, as it provides the complete code base, dataset, and an initial baseline representative of the current state-of-the-art of point cloud tree species classification methods.

Deep-Learning-basierte Objekterkennung (DL-Objekterkennung) anhand hochauflösender Luftbilder kann wertvolle Informationen und Planungsgrundlagen für die Waldbewirtschaftung, den Naturschutz und die Waldforschung liefern. Durch DL-Objekterkennung können die vorkommenden Baumarten nicht nur identifiziert, sondern auch lokalisiert und gezählt werden. In dieser Arbeit wird eine Methode zur automatischen Baumartenerkennung in Schweizer Mischbeständen vorgestellt, die es ermöglicht, die Hauptbaumarten mit mittlerer bis guter Genauigkeit zu erkennen und zu unterscheiden.

1. In close-range remote sensing data collected in a forest, occlusion often causes incomplete or sparse point cloud representations of individual trees, impeding accurate 3D reconstruction of tree architecture and estimation of tree height and volume. Recent developments in deep learning (DL) for 3D data have produced approaches for point cloud completion, which could potentially be applied to trees.
2. We explored the potential of a DL approach to fill gaps in dense point clouds representing the main structures of deciduous trees by applying an existing transformer-based completion model (PoinTr). Complete point clouds are required as training data, but even dense terrestrial laser scanning (TLS) data sets contain gaps caused by occlusion, making it nearly impossible to acquire such data. We therefore investigated the ability of point cloud completion models trained on a range of synthetic data sets to handle occlusion patterns in real-world point clouds.
3. Despite the limited data set, we successfully fine-tuned a general pre-trained completion model to fill gaps within 1 m³ segments of tree point clouds. Fine-tuning on synthetic tree data improved the model's ability to complete tree objects compared with training on diverse artificial objects. However, the quality of the predictions was influenced by the level of sophistication of the synthetic data. Our results demonstrate that incorporating even limited real-world TLS data during training can considerably improve completion results but may introduce additional noise in the predictions.
4. 3D point cloud completion with DL has the potential to improve and fill gaps in point clouds of individual trees, facilitating further steps in the processing and analysis of 3D forest data.

Understanding and monitoring the surrounding environment increasingly rely on its 3D representations. However, the often high costs of 3D data equipment limit its wide usage, and low-cost solutions are in demand. Here, we propose a novel approach based on spherical stereo videos captured with a known baseline (distance between the cameras) for a low-cost and efficient 3D point cloud reconstruction. In a forest environment, we evaluated 1) the influence of baseline length on point cloud quality and 2) the suitability of the generated point clouds for extracting primary forest attributes (tree position and diameter). Our results show that the proposed approach allows for feasible 3D reconstruction of complex forest plots. The highest point cloud quality was achieved with a baseline of 60 cm. This setup enabled the correct detection of more than 65% of the trees within the forest plots, producing an average tree position error between 30 and 50 cm and clearly outperforming other setups. A Multiscale Model-to-Model Cloud (M3C2) analysis showed signed distances between the generated point cloud and the reference data with zero mean and 1 m standard deviation. We demonstrate that the proposed approach can be a valuable low-cost solution for 3D point cloud reconstruction, facilitating forest assessment and monitoring.

Purpose of review In recent years, the use of 3D point clouds in silviculture and forest ecology has seen a large increase in interest. With the development of novel 3D capture technologies, such as laser scanning, an increasing number of algorithms have been developed in parallel to process 3D point cloud data into more tangible results for forestry applications. From this variety of available algorithms, it can be challenging for users to decide which to apply to fulfil their goals best. Here, we present an extensive overview of point cloud acquisition and processing tools as well as their outputs for precision forestry. We then provide a comprehensive database of 24 algorithms for processing forest point clouds obtained using close-range techniques, specifically ground-based platforms.
Recent findings Of the 24 solutions identified, 20 are open-source, two are free software, and the remaining two are commercial products. The compiled database of solutions, along with the corresponding technical guides on installation and general use, is accessible on a web-based platform as part of the COST Action 3DForEcoTech. The database may serve the community as a single source of information to select a specific software/algorithm that works for their requirements.
Summary We conclude that the development of various algorithms for processing point clouds offers powerful tools that can considerably impact forest inventories in the future, although we note the necessity of creating a standardisation paradigm.

Automatic identification and mapping of tree species is an essential task in forestry and conservation. However, applications that can geolocate individual trees and identify their species in heterogeneous forests on a large scale are lacking. Here, we assessed the potential of the Convolutional Neural Network algorithm, Faster R-CNN, which is an efficient end-to-end object detection approach, combined with open-source aerial RGB imagery for the identification and geolocation of tree species in the upper canopy layer of heterogeneous temperate forests. We studied four tree species, i.e., Norway spruce (Picea abies (L.) H. Karst.), silver fir (Abies alba Mill.), Scots pine (Pinus sylvestris L.), and European beech (Fagus sylvatica L.), growing in heterogeneous temperate forests. To fully explore the potential of the approach for tree species identification, we trained single-species and multi-species models. For the single-species models, the average detection accuracy (F1 score) was 0.76. Picea abies was detected with the highest accuracy, with an average F1 of 0.86, followed by A. alba (F1 = 0.84), F. sylvatica (F1 = 0.75), and Pinus sylvestris (F1 = 0.59). Detection accuracy increased in multi-species models for Pinus sylvestris (F1 = 0.92), while it remained the same or decreased slightly for the other species. Model performance was more influenced by site conditions, such as forest stand structure, and less by illumination. Moreover, the misidentification of tree species decreased as the number of species included in the models increased. In conclusion, the presented method can accurately map the location of four individual tree species in heterogeneous forests and may serve as a basis for future inventories and targeted management actions to support more resilient forests.

Accurate estimates of above-ground tree biomass within forest inventories are essential for calibration and validation of biomass mapping products based on Earth observation data. Terrestrial laser scanning (TLS) enables detailed and non-destructive volume estimation of individual trees, which can be converted to biomass with wood basic density. Existing TLS-based approaches range from simple geometrical features to virtual 3D reconstruction of entire trees. Validating such approaches with weight measurements is a key step before the integration of TLS or other close-range technologies into operational applications such as forest inventories. In this study, we firstly evaluate individual tree volume estimation approaches based on 3D reconstruction through quantitative structure models (QSM) against destructive reference data of 60 trees and compare them to operational allometric scaling models (ASM). Secondly, we determine the explanatory power of TLS-derived geometric parameters regarding total wood, stem, coarse wood and fine branch volume. We observe similar accuracy in merchantable (>7 cm) wood compartments for ASMs (NRMSE = 25 %) and QSMs (NRMSE = 29 %), with QSMs showing better results for broadleaves than conifers and generally overestimating fine branch volume. Feature selection shows that a combination of stem diameters and volume of convex hulls around tree crowns has the most potential to model the entire tree volume including branches, especially for conifers. In cases where the quality of available point clouds is insufficient for QSMs, 3D information can thus still be utilised by deriving geometric parameters. The integration of crown dimension parameters into new allometric models could substantially improve the estimation of branch wood volume.

Monitoring biodiversity in forests is crucial for their management and preservation, especially in light of increasing climatic disturbances. However, traditional methods of surveying forest biodiversity, such as the inventory of tree-related microhabitats (TreMs), are costly and time-consuming. For many years, terrestrial laser scanning (TLS) was the main method for producing highly accurate 3D models of forests. However, with recent advancements in 3D scanning technologies, there are now numerous alternatives available on the market. The aim of this study was to evaluate the performance of four different 3D data acquisition methods, i.e. close-range photogrammetry (CRP), fish-eye photogrammetry (FEP), mobile laser scanning (MLS), and mixed reality depth camera (MRDC), in terms of accuracy and ability to measure biodiversity (TreMs) at tree-stem level, in comparison to TLS. Analysis was performed based on geometric accuracy and point neighbourhood relevance. CRP was the most accurate alternative to TLS for TreM measurement with a median error of 1.5 cm, while FEP provided a good balance between accuracy (median error 1.4 cm) and speed of data collection. Although MLS showed promising results (median error 1.6 cm), noise in the point cloud limited its ability to identify TreMs. MRDC, on the other hand, had lower quality (median error 3.6 cm) and lower point density, making it unsuitable for TreM segmentation. Nevertheless, the study demonstrated the feasibility of augmenting the real world with virtual content at single-tree-stem level using mixed reality technology. Overall, the 3D scanning technologies presented hold great promise for recording the evolution of biodiversity at stem level.

Depth estimation from a single image is a challenging task, especially inside the highly structured forest environment. In this paper, we propose a supervised deep learning model for monocular depth estimation based on forest imagery. We train our model on a new data set of forest RGB-D images that we collected using a terrestrial laser scanner. Alongside the input RGB image, our model uses a sparse depth channel as input to recover the dense depth information. The prediction accuracy of our model is significantly higher than that of state-of-the-art methods when applied in the context of forest depth estimation. Our model brings the RMSE down to 2.1 m, compared to 4 m and above for reference methods.
 

Mit dem Ziel, mehr auf Messungen statt auf Schät-zungen beruhende Daten auf grösserer Fläche zu erhalten, werden an der Eidgenössischen Forschungsanstalt WSL und im Schweizerischen Landesforstinventar (LFI) derzeit verschiedene Methoden der nahen Fernerkundung getestet. Mit diesen Methoden können hochaufgelöste 3-D-Punktdaten auf Inventurplots und Dauerbeobachtungsflächen effizient erhoben und Merkmale zu Bäumen und Waldstrukturen abgeleitet werden.

This paper is an attempt to respond to the growing need and demand of 3D data in forestry, especially for 3D mapping. The use of terrestrial laser scanners (TLS) dominates contemporary literature for under-storey vegetation mapping as this technique provides precise and easy-to-use solutions for users. However, TLS requires substantial investments in terms of device acquisition and user training. The search for and development of low-cost alternatives is therefore an interesting field of inquiry. Here, we use low-cost 360° cameras combined with spherical photogrammetric principles for under-storey vegetation mapping. While we fully assume that this low-cost approach will not generate results on par with either TLS or classical close-range photogrammetry, its main aim is to investigate whether this alternative is sufficient to meet the requirements of forest mapping. In this regard, geometric analyses were conducted using both TLS and close-range photogrammetry as comparison points. The diameter at breast height (DBH), a parameter commonly used in forestry, was then computed from the 360° point cloud using three different methods to determine if a similar order of precision to the two reference datasets can be obtained. The results show that 360° cameras were able to generate point clouds with a similar geometric quality as the references despite their low density, albeit with a significantly higher amount of noise. The effect of the noise is also evident in the DBH computation, where it yielded an average error of 3.5 cm compared to both the TLS and close-range photogrammetry.

Historical black-and-white (B&W) aerial images have been recognized as an important source of information for assessing vegetation dynamics. However, the use of these images is limited by the lack of multispectral information, as well as by their varying quality. It is therefore important to study and develop methods that are capable of automatic and accurate classification of these B&W images while reducing the need for tedious manual work. The goal of this study was to assess changes over 30 years in woody vegetation cover along alpine treeline ecotones using B&W aerial images from two time points. A convolutional neural networks model was firstly set up based on three structure classes calculated from Airborne Laser Scanning data using the B&W aerial images from 2010. Then, the model was improved by active addition of training samples of those that were wrongly predicted from historical B&W aerial images from 1980. 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 study illustrates that vegetation changes at the treeline ecotone can be detected in order to assess areawide long-term vegetation dynamics at a fine spatial resolution.

Künstliche Intelligenz, Deep learning, hochaufgelöste Satellitenbilder, automatisierte Auswertungen, landesweite Datensätze etc. sind alles Begriffe, mit denen viele im Bereich Wald- und Forstwissenschaften schon einmal konfrontiert worden sind. Was steckt aber nun genau dahinter bzw. wird bereits angewandt? Im folgenden Artikel wird anhand des neuen «Waldmischungsgrades LFI» aufgezeigt, was fernerkundungs-technisch alles möglich ist und wo eben auch die Grenzen liegen. Dabei wird auch deutlich, dass ein Austausch zwischen der Praxis und der Forschung wichtiger denn je ist.

Countrywide winter and summer Sentinel-1 (S1) backscatter data, cloud-free summer Sentinel-2 (S2) images, an Airborne Laser Scanning (ALS)-based Digital Terrain Model (DTM) and a forest mask were used to model and subsequently map Dominant Leaf Type (DLT) with the thematic classes broadleaved and coniferous trees for the whole of Switzerland. A novel workflow was developed that is robust, cost-efficient and highly automated using reference data from aerial image interpretation. Two machine learning approaches based on Random Forest (RF) and deep learning (UNET) for the whole country with three sets of predictor variables were applied. 24 subareas based on aspect and slope categories were applied to explore effects of the complex mountainous topography on model performances. The reference data split into training, validation and test data sets was spatially stratified using a 25 km regular grid. Model accuracies of both RF and UNET were generally highest with Kappa (K) around 0.95 when predictors were included from both S1/S2 and the topographic variables aspect, elevation and slope from the DTM. While only slightly lower accuracies were obtained when using S2 and DTM data, lowest accuracies were obtained when only predictors from S1 and DTM were included, with RF performing worse than UNET. While on countrywide level RF and UNET performed overall similarly, substantial differences in model performances, i.e. higher variances and lower accuracies, were found in subareas with northwest to northeast orientations. The combined use of S1/S2 and DTM predictors mitigated these problems related to topography and shadows and was therefore superior to the single use of S1 and DTM or S2 and DTM data. The comparison with independent National Forest Inventory (NFI) plot data demonstrated precisions of K around 0.6 in the predictions of DLT and indicated a trend of increasing deviations in mixed forests. A comparison with the Copernicus High Resolution Layer (HRL) DLT 2018 revealed overall higher map accuracies with the exception of pure broadleaved forest. Although, spatial patterns of DTL were overall similar, UNET performed better than RF in areas with a distinct DLT on forest stand level, with the largest differences occurring when only S1 and DTM data was used. In contrast, predictions obtained from RF were more accurate in mixed stands. This study goes beyond the case study level and meets the requirements of countrywide data sets, in particular regarding repeatability, updating, costs and characteristics of training data sets. The 10 m countrywide DLT maps add complementary and spatially explicit information to the existing NFI estimates and are thus highly relevant for forestry practice and other related fields.

Methods and workflows for creating area-wide products using remote sensing methods have been developed for various reasons. For the sample-based estimates in the Swiss National Forest Inventory (NFI), area-wide data sets are used for the two-phase estimations, which have substantially lower estimation errors than one-phase inventories. Using area-wide data sets, it is possible to skip dense manual interpretation of the stereo-images and thus save resources. Further, many functions of the forest, such as biodiversity and protection against natural hazards, can be described and quantified better with area-wide spatial data than with field plot data.

 

Tree-related microhabitats (TreMs) play an important role in maintaining forest biodiversity and have recently received more attention in ecosystem conservation, forest management and research. However, TreMs have until now only been assessed by experts during field surveys, which are time-consuming and difficult to reproduce. In this study, we evaluate the potential of close-range terrestrial laser scanning (TLS) for semi-automated identification of different TreMs (bark, bark pockets, cavities, fungi, ivy and mosses) in dense TLS point clouds using machine learning algorithms, including deep learning. To classify the TreMs, we applied: (1) the Random Forest (RF) classifier, incorporating frequently used local geometric features and two additional self-developed orientation features, and (2) a deep Convolutional Neural Network (CNN) trained using rasterized multiview orthographic projections (MVOPs) containing top view, front view and side view of the point’s local 3D neighborhood. The results confirmed that using local geometric features is beneficial for identifying the six groups of TreMs in dense tree-stem point clouds, but the rasterized MVOPs are even more suitable. Whereas the overall accuracy of the RF was 70%, that of the deep CNN was substantially higher (83%). This study reveals that close-range TLS is promising for the semi-automated identification of TreMs for forest monitoring purposes, in particular when applying deep learning techniques.

Forest structure reflects the forest disturbance regime and can provide important information about the rate of human impact. A better understanding of the structural variability and large-scale dynamics of natural forests is crucial for "close to nature" forest management planning. In this study, we developed a partly automated approach to assess the structure of potential primeval and managed beech forests in the Ukrainian Carpathians using WorldView-2 imagery. We analyzed the local (50 m × 50 m scale) canopy closure of these forests by extracting the canopy gaps and determined four forest structure types ranging from very closed to low density. The occurrence and frequencies of these structure types were significantly different in the primeval and managed beech forests. The four forest structure types were predicted and mapped using multinomial logistic regression based on the textural features derived from the original image bands and two vegetation indices. A 10-fold cross-validation resulted in an overall accuracy of 83% and a kappa coefficient of 75%, with the highest agreement for the very closed structure type (87%) and the lowest agreement for the medium density and low density structure types (79%). The forest structure type maps can be helpful for planning management activities in beech forests.

Although wall-to-wall, accurate, and up-to-date forest composition maps at the stand level are a fundamental input for many applications, ranging from global environmental issues to local forest management planning, countrywide mapping approaches on the tree type level remain rare. This paper presents and validates an innovative remote sensing based approach for a countrywide mapping of broadleaved and coniferous trees in Switzerland with a spatial resolution of 3 m. The classification approach incorporates a random forest classifier, explanatory variables from multispectral aerial imagery and a Digital Terrain Model (DTM) from Airborne Laser Scanning (ALS) data, digitized training polygons and independent validation data from the National Forest Inventory (NFI). The methodological workflow was optimized for an area of 41,285 km² that is characterized by temperate forests within a complex topography. Whereas high model overall accuracies (0.99) and kappa (0.98) were achieved, the comparison of the tree type map with independent NFI data revealed significant deviations that are related to underestimations of broadleaved trees (median of -3.17%). Constraints of the tree type mapping approach are mostly related to the acquisition date and time of the imagery and the topographic (negative) effects on the prediction. A comparison with the most recent High Resolution Layers (HRL) forest 2012 from the European Environmental Agency revealed that the tree type map is superior regarding spatial resolution, level of detail and accuracy. The high-quality map achieved with the approach presented here is of great value for optimizing forest management and planning activities and is also an important information source for applications outside the forestry sector.