256869 research outputs found
Sort by
Multispectral airborne laser scanning for tree species classification: A benchmark of machine learning and deep learning algorithms
No full textClimate-smart and biodiversity-preserving forestry demands precise information on forest resources, extending to the individual tree level. Multispectral airborne laser scanning (ALS) has shown promise in automated point cloud processing, but challenges remain in leveraging deep learning techniques and identifying rare tree species in class-imbalanced datasets. This study addresses these gaps by conducting a comprehensive benchmark of deep learning and traditional shallow machine learning methods for tree species classification. For the study, we collected high-density multispectral ALS data (>1000 pts/m2) at three wavelengths using the FGI-developed HeliALS system, complemented by existing Optech Titan data (35 pts/m2), to evaluate the species classification accuracy of various algorithms in a peri-urban study area located in southern Finland. We established a field reference dataset of 6326 segments across nine species using a newly developed browser-based crowdsourcing tool, which facilitated efficient data annotation. The ALS data, including a training dataset of 1065 segments, was shared with the scientific community to foster collaborative research and diverse algorithmic contributions. Based on 5261 test segments, our findings demonstrate that point-based deep learning methods, particularly a point transformer model, outperformed traditional machine learning and image-based deep learning approaches on high-density multispectral point clouds. For the high-density ALS dataset, a point transformer model provided the best performance reaching an overall (macro-average) accuracy of 87.9% (74.5%) with a training set of 1065 segments and 92.0% (85.1%) with a larger training set of 5000 segments. With 1065 training segments, the best image-based deep learning method, DetailView, reached an overall (macro-average) accuracy of 84.3% (63.9%), whereas a shallow random forest (RF) classifier achieved an overall (macro-average) accuracy of 83.2% (61.3%). For the sparser ALS dataset, an RF model topped the list with an overall (macro-average) accuracy of 79.9% (57.6%), closely followed by the point transformer at 79.6% (56.0%). Importantly, the overall classification accuracy of the point transformer model on the HeliALS data increased from 73.0% with no spectral information to 84.7% with single-channel reflectance, and to 87.9% with spectral information of all the three channels. Furthermore, we studied the scaling of the classification accuracy as a function of point density and training set size using 5-fold cross-validation of our dataset. Based on our findings, multispectral information is especially beneficial for sparse point clouds with 1–50 pts/m2. Furthermore, we observed that the classification error follows a power law ɛ(m)≈m−α as a function of the training set size m, and the classification error of the point transformer reduced significantly faster with increasing training set size compared to RF
Machine Learning Supporting Science: Representative Examples in Evolving Systems
No full textThis presentation examines how machine learning can support scientific discovery in evolving, dynamic systems. Five common ML use cases in experimental science are introduced — dimensionality reduction, structure discovery, behavioral prediction, temporal change detection, and knowledge extraction under limited data — providing a conceptual framework that guides the subsequent applied examples.
Three domains illustrate these ideas in practice. In cybersecurity, ML is used for network traffic analysis and anomaly detection, highlighting the critical role of data representation and the often-overlooked statistical pitfalls of real-world deployment. In environmental science, digital twin architectures and time series forecasting methods are applied to the monitoring and prediction of natural resources such as groundwater levels. In medicine, persistent homology is leveraged to uncover hidden structure in the longitudinal microbiome data of cystic fibrosis patients, revealing cyclic dynamics that conventional ML approaches had failed to detect.
A recurring theme across all examples is that methodological rigor, domain understanding, and thoughtful data representation are as decisive as algorithm choice when applying ML to real scientific problems
Fock state probability changes in open quantum systems
No full textOpen quantum systems are powerful effective descriptions of quantum systems interacting with their environments. Studying changes of Fock state probabilities can be intricate in this context since the prevailing description of open quantum dynamics is by master equations of the systems’ reduced density matrices, which usually requires finding solutions for a set of complicated coupled differential equations. In this article, we show that such problems can be circumvented by employing a recently developed path integral-based method for directly computing reduced density matrices in scalar quantum field theory. For this purpose, we consider a real scalar field φ as an open system interacting via a λχ²φ²-term with an environment comprising another real scalar field χ that has a finite temperature. In particular, we investigate how the probabilities for observing the vacuum or two-particle states change over time if there were initial correlations of these Fock states. Subsequently, we apply our resulting expressions to a neutrino toy model. We show that, within our model, lighter neutrino masses would lead to a stronger distortion of the observable number of particles due to the interaction with the environment after the initial production process
Automated Recognition and Valuation of Reusable Building Components Using State-of-the-Art Object Detection: A Case Study on Multi-Family Redevelopment Properties
No full textDer Bau- und Immobiliensektor erzeugt ca. 39% der globalen Treibhausgasemissionen und 36% des festen Abfalls. EU-Vorgaben fordern mittelfristig eine 70%-ige Wiederverwendung oder das Recycling von Bau- und Abbruchabfällen, doch Vorab-Auditierungen sind langsam, teuer und von wenigen Experten abhängig. Diese Arbeit untersucht, ob moderne Objekterkennung und Regression die Inventarisierung wiederverwendbarer Gebäudekomponenten automatisieren und deren Wert schätzen können. Videos von elf klassischen Wiener Zinshäusern mit Sanierungsbedarf wurden von Fachexperten für sieben Objektklassen annotiert. YOLOv11 und Mask R-CNN wurden mit identischem Labelraum trainiert und auf einem Test-Split evaluiert. Ein Regressionsmodell sagte den Komponentenwert basierend auf Klassenzugehörigkeit und Vorhersagekonfidenz voraus. Bewertet wurden Objekterkennung ([email protected], [email protected]:0.95, F1-Kurven, Konfusionsmatrizen), Kalibrierung (ECE) sowie Wertermittlung (ME, MAE, RMSE, Bland–Altman). Eine Fallstudie verglich Aufwand und Kosten von Mensch und KI. Mask R-CNN erreichte AP50 = 0.046 und YOLOv11 AP50 = 0.019; die maximalen F1-Werte lagen bei ca. 0.104 bzw. ca. 0.144. Die Kalibrierung verringerte den ECE von 0.779 auf 0.609. Bei 44 abgeglichenen Gegenständen betrug der mittlere Fehler –14,77€, MAE=19,32€ und RMSE=32,42€, wobei 90.9% der Residuen innerhalb der Limits of Agreement lagen. Menschliche Experten bewerteten im Durchschnitt ca. 1,925€ pro Immobilie, die KI ca. 1,275€ (33% Bias). Die KI-Pipeline erzielte einen Zeitgewinn von ca. 22.5x (ca. 20 min. vs 450 min. pro Immobilie). Unsere Forschungsarbeit zeigt somit, dass Computer-Vision-Modelle wiederverwendbare Gebäudekomponenten trotz niedriger mAP zuverlässig erkennen und bewerten können. Die KI neigt zur Unterbewertung, doch lassen sich Bias mittels Kalibrierung reduzieren. Die Pipeline reduziert den Aufwand drastisch und ermöglicht die Erfassung durch Laien, was die Circular Economy fördert. Zukünftige Arbeiten sollten temporale Verfolgung, eine größere Datensammlung und reichere Annotationen untersuchen, um Genauigkeit und Generalisierbarkeit zu verbessern.The construction and real estate sector produces approx. 39% of global greenhouse-gas emissions and 36% of solid waste. EU regulations mandate in the medium-term 70% reuse or recycling of construction and demolition waste, but pre-demolition audits are slow, costly and reliant on scarce domain experts. This thesis investigates whether state-of-the-art object detection and regression models can automate the inventorying of reusable building components and estimate their recoverable value. Videos of eleven Viennese multi-family buildings in need of renovation were annotated by domain experts for seven object classes. YOLOv11 and Mask R-CNN were trained with identical label spaces and evaluated on a held-out test split. A regression model predicted component value based on detection confidence and class. We assessed detection performance ([email protected], [email protected]:0.95, F1 sweeps, confusion matrices), calibration (ECE) and value estimation (ME, MAE, RMSE, Bland–Altman). A human–vs-AI case study compared time and cost efficiency. Mask R-CNN reached AP50 = 0.046 and YOLOv11 AP50 = 0.019; F1 maxima were ca. 0.104 and ca. 0.144. Calibration reduced ECE from 0.779 to 0.609. On 44 matched items the mean error was –14.77€, MAE=19.32€ and RMSE=32.42€, with 90.9% of residuals within the limits of agreement. Human experts recovered ca. 1,925€ per property whereas AI recovered ca. 1,275€, a 33% bias. The AI pipeline achieved a ca. 22.5x speed-up (ca. 20 min. vs 450 min. per property). Our research demonstrates that computer-vision models can reliably detect and value reusable building components despite low frame-level mAP. While the AI tends to undervalue assets, calibration and post-processing mitigate bias. The pipeline dramatically reduces time and cost, enabling layperson captures and democratizing circular-economy audits. Future work should explore temporal tracking, dataset expansion and richer annotations to improve accuracy and generalization
Entwicklung und Validierung eines Dosimetrie-Setups für LiF:Mg,Ti-Thermolumineszenzdetektoren bei Röntgen- und Ionenstrahlbestrahlung
No full textThermoluminescent dosimeters (TLDs) are of great interest in radiation therapy and radiation protection due to their compact size and tissue equivalence. The aim of this MSc thesis was to investigate the application possibilities of these detectors in ion beam therapy and reference irradiation (i.e., kV x-rays). LiF:Mg,Ti TLDs were used for this purpose, which were available as pellets with a diameter of 4.5 mm and a thickness of 0.9 mm (type MTS-N, manufactured by Radcard s. c., Cracow, Poland). First, individual correction factors (CFs) were calculated for each TLD using beta irradiation. CFs were applied to correct for the varying responses between individual TLDs, due to the variation in their crystal structure. The respective measurements, as well as all the further TLD-readouts, were performed using a Risø TLD reader with an integrated beta irradiation source. The CFs were then applied to the reference beam qualities (kV x-rays, proton and carbon ion beams) used in the calibration of the TLDs. Furthermore, a TLD holder in the form of a dedicated ionization chamber was designed with Autodesk Fusion 2.0.21487 (Autodesk, Inc., San Rafael, CA, United States) and 3D printed from acrylonitrile butadiene styrene (ABS) during this thesis. This prototype in combination with the ionization chamber (PTW 30013 Farmer) was used for the dose-calibration of TLDs in a water-equivalent phantom under x-ray and ion beam irradiation. For calibration in x-rays, dose values between 0.1 Gy and 10 Gy with 70 kV and 200 kV tube voltage were used, while the experimental setup aligned with the recommendations of the IAEA TRS-398 (Rev. 1) guidelines [1], as accurately as possible. The calibration factor obtained from the 200 kV calibration was then used in an in-vivo kV x-ray irradiation setting for a spinal cord transplantation experiment conducted with mice at the AKH (General Hospital, Vienna). Finally, linear energy transfer (LET) dependencies were analyzed in mono-energetic and spread-out Bragg peak (SOBP) proton and carbon ion beams. The TLD holder in combination with the ionization chamber was previously used for the dose-calibration of TLDs in the water-equivalent phantom with the experimental setup adapted as closely as possible to the IAEA TRS-398 (Rev. 1) guidelines [1]. For this calibration, dose values of 1 Gy, 2 Gy and 5 Gy with energies of 179.2 MeV for proton beams and 346.6 MeV per nucleon for carbon ion beams were used. The calibration factors obtained during this calibration with mono-energetic beams were then used in two different modulations per particle type at varying positions within SOBPs with a physical dose (D50) of 2 Gy, to analyze the LET dependencies of TLDs along SOBPs with a constant dose. The TLDs exhibited linear behavior in response within the dose range considered in x-ray and ion beam calibrations. Combined with further assessments of the repeatability of the TLD response and validations of the calibration factor obtained in 200 kV x-ray irradiation indicated that the calibration method used provides potentially reliable results. Dose measurements in kV x-rays with calibrated TLDs in combination with the use of CFs were comparable to the measurements with the reference detector (PTW 30013 Farmer), with the exception of a higher uncertainty in the order of one magnitude. With this result and assuming the TLD measurements during the in-vivo experiment were correct, the experimental setup for small animal whole body irradiation was identified as requiring recalibration by means of TLD verification. However, dose determinations using calibrated TLDs in ion beams were less reliable because the TLD response strongly decreased with increasing LET values. The finding in those LET dependencies when using calibration factors obtained with the calibration method developed during this thesis was an approximately linear behavior of the decrease in the TLD response between 4 keV/μm and 100 keV/μm, which could potentially be used to determine specific LET values in SOBP ion beams, provided the dose is known. Therefore, further research on the LET dependencies of the TLD response is strongly recommended
Ingrid Kretschmer – eine prägende Gestalt der österreichischen Kartografie
No full textAls Ingrid Kretschmer am 22. Jänner 2011 in Linz verstarb, verlor die österreichische und internationale wissenschaftliche Gemeinschaft eine Persönlichkeit, die über Jahrzehnte hinweg zentrale Impulse für Forschung, Lehre und Institutionenentwicklung im Bereich der Kartografie setzte. Ihr Wirken umfasste nicht nur bedeutende wissenschaftliche Veröffentlichungen, sondern auch grundlegende Leistungen für die Etablierung und Organisation kartografischer Forschung in Österreich
Style Brush: Guided Style Transfer for 3D Objects
No full textWe introduce Style Brush, a novel style transfer method for textured meshes designed to empower artists with fine-grained control over the stylization process. Our approach extends traditional 3D style transfer methods by introducing a novel loss function that captures style directionality, supports multiple style images or portions thereof and enables smooth transitions between styles in the synthesized texture. The use of easily generated guiding textures streamlines user interaction, making our approach accessible to a broad audience. Extensive evaluations with various meshes, style images and contour shapes demonstrate the flexibility of our method and showcase the visual appeal of the generated textures. Finally, the results of a user study indicate that our approach generates visually appealing mesh textures that adhere to user-defined guidance and enable users to retain creative control during stylization. Our implementation is available on: https://github.com/AronKovacs/style-brush