Dr. Hristina Hristova

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.

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.

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.