Lorenzo Beltrame

Mr. Lorenzo Beltrame


Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
21 August 2024 in Drones
Verified
Journal Article
Open access
PDF file
LBLorenzo Beltrame
Lorenzo Beltrame
JSJules Salzinger
Jules Salzinger
LKLukas J. Koppensteiner
Lukas J. Koppensteiner
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
In this study, we propose a scalable deep learning approach to automated phenotyping using UAV multispectral imagery, exemplified by yellow rust detection in winter wheat. We adopt a high-granularity scoring method (1 to 9 scale) to align with international standards and plant breeders’ needs. Using a lower spatial resolution (60 m flight height at 2.5 cm GSD), we reduce the data volume by a factor of 3.4, making large-scale phenotyping faster and more cost-effective while obtaining results comparable to those of the state-of-the-art. Our model incorporates explainability components to optimise spectral bands and flight schedules, achieving top-three accuracies of 0.87 for validation and 0.67 and 0.70 on two separate test sets. We demonstrate that a minimal set of bands (EVI, Red, and GNDVI) can achieve results comparable to more complex setups, highlighting the potential for cost-effective solutions. Additionally, we show that high performance can be maintained with fewer time steps, reducing operational complexity. Our interpretable model components improve performance through regularisation and provide actionable insights for agronomists and plant breeders. This scalable and explainable approach offers an efficient solution for yellow rust phenotyping and can be adapted for other phenotypes and species, with future work focusing on optimising the balance between spatial, spectral, and temporal resolutions.

8 July 2026 in ISPRS Annals of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Verified
Journal Article
Open access
PDF file
LBLorenzo Beltrame
Lorenzo Beltrame
JSJules Salzinger
Jules Salzinger
FSFilip Svoboda
Filip Svoboda
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
JLJasmin Lampert
Jasmin Lampert
MKMarco Körner
Marco Körner
Shadows cast by terrain and tall structures remain a major obstacle for high-resolution satellite image analysis, degrading classification, detection, and 3D reconstruction performance. Public resources offering geometry-consistent paired shadow/shadow-free satellite imagery are essentially missing, and most Earth-observation datasets are designed for shadow detection or 3D modelling rather than removal. Existing deep shadow-removal datasets either target ground-level or aerial scenes or rely on unpaired and weakly supervised formulations rather than explicit satellite pairs. We address this gap with deSEO, a geometry-aware and physics-informed methodology that, to the best of our knowledge, is the first to derive paired supervision for satellite shadow removal from the S-EO shadow detection dataset (Masquil et al., 2025) through a fully replicable pipeline. For each tile, deSEO selects a minimally shadowed acquisition as a weak reference and pairs it with shadowed counterparts using temporal and geometric filtering, Jacobian-based orientation normalisation, and LoFTR–RANSAC registration. A per-pixel validity mask restricts learning to reliably aligned regions, enabling supervision despite residual off-nadir parallax. In addition to this paired dataset, we develop a DSM-aware deshadowing model that combines residual translation, perceptual objectives, and mask-constrained adversarial learning. In contrast, a direct adaptation of a UAV-based SRNet/pix2pix architecture fails to converge under satellite viewpoint variability. Our model consistently reduces the visual impact of cast shadows across diverse illumination and viewing conditions, achieving improved structural and perceptual fidelity on held-out scenes. deSEO therefore provides the first reproducible, geometry-aware paired dataset and baseline for shadow removal in satellite Earth observation.

14 March 2026
Verified
Conference Paper
Open access
ABAlbin Bjärhall
Albin Bjärhall
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
LBLorenzo Beltrame
Lorenzo Beltrame
JSJules Salzinger
Jules Salzinger
JLJasmin Lampert
Jasmin Lampert
BSBenjamin Schumacher
Benjamin Schumacher
The Austrian Research Centre for Forests (BFW) uses a range of remote sensing (RS) data—including aerial imagery, airborne laser scanning, and high‑resolution Copernicus Sentinel data—to support the National Forest Inventory and produce national forestry products. These include tree species maps, timber stock maps, and a Sentinel‑2‑based anomaly detection map for identifying forest disturbances across Austria. The success of these products is grounded in three factors: (1) strong collaboration with researchers to ensure that recent scientific advances are translated into operational applications; (2) the integration of high‑resolution aerial data with extensive local expertise, enabling high‑quality training datasets for machine‑learning approaches; and (3) the Austrian National Forest Inventory, which provides a robust ground‑truth basis for validating RS‑derived results.At the same time, the demand for precise, timely, and spatially detailed national forestry products is steadily increasing. This demand is driven by growing monitoring and reporting requirements, as well as by the increasing impacts of land-use and climate change on forest ecosystems. These developments highlight existing limitations of RS-based forestry products, particularly in complex alpine terrain, where terrain shadows and cloud cover can delay or obscure the detection of natural disturbances such as windthrows. Within the SAFIR project, we investigate how AI–based tools can be used to address these challenges and enhance the performance of existing forest disturbance monitoring tools.As presented in our poster, the SAFIR project combines BFW’s anomaly detection map with ground-truth training data on windthrow events provided by the Österreichische Bundesforste (ÖBF). This fusion allows us to: (1) distinguish windthrow events from other disturbance types within the anomaly detection map; (2) assess the spatial and temporal agreement between modelled disturbances and inventoried windthrow events across Austria; and (3) quantify the detection rate of windthrow events in the existing product. Building on this assessment, identified windthrow sites can be specifically targeted with machine-learning approaches for cloud and terrain-shadow removal, thereby improving both the timeliness and accuracy of windthrow detection and providing validated inputs for developing new datasets and AI models specialized in de‑clouding and de‑shadowing windthrow areas.By integrating established Copernicus-based forest disturbance products with an extensive, independently collected ground-truth dataset on windthrows, SAFIR enables a systematic evaluation of current windthrow detection capabilities and provides a pathway for targeted methodological improvement. By leveraging AI-based techniques to overcome known limitations, the project contributes to the development of more robust national forestry products for monitoring windthrow damage.

20 August 2024 in Preprints.org
Verified
Preprint
Open access
PDF file
LBLorenzo Beltrame
Lorenzo Beltrame
JSJules Salzinger
Jules Salzinger
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
LKLukas Koppensteiner
Lukas Koppensteiner
In this study, we propose a scalable deep learning approach to automated phenotyping using UAV multispectral imagery, exemplified by yellow rust detection in winter wheat. We adopt a high granularity scoring method (1 to 9 scale) to align with international standards and plant breeders’ needs. Using a lower spatial resolution (60m flight height at 2.5cm GSD), we reduce the data volume by a factor of 3.4, making large-scale phenotyping faster and more cost-effective while obtaining results comparable to those of the state-of-the-art. Our model incorporates explainability components to optimise spectral bands and flight schedules, achieving top-3 accuracies of 0.87 for validation and 0.67 and 0.70 on two separate test sets. We demonstrate that a minimal set of bands (EVI, Red, and GNDVI) can achieve results comparable to more complex setups, highlighting the potential for cost-effective solutions. Additionally, we show that high performance can be maintained with fewer time steps, reducing operational complexity. Our interpretable model components improve performance through regularisation and provide actionable insights for agronomists and plant breeders. This scalable and explainable approach offers an efficient solution for yellow rust phenotyping and can be adapted for other phenotypes and species, with future work focusing on optimising the balance between spatial, spectral, and temporal resolutions.

14 March 2026
Verified
Conference Paper
Open access
JSJules Salzinger
Jules Salzinger
LBLorenzo Beltrame
Lorenzo Beltrame
LSLukas-Till Schawerda
Lukas-Till Schawerda
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
Reliable, field-scale indicators of crop water status and plant condition are needed to support plant breeding, precision irrigation and climate adaptation, yet UAV-based monitoring must balance predictive accuracy with deployability. We present an explainable Deep Learning workflow (TriNet) for scalable UAV phenotyping from multispectral time series, aligned with agronomic and breeding practice through high-granularity in situ scoring in accordance with established standards (such as those of the AGES - Österreichische Agentur für Gesundheit und Ernährungssicherheit). TriNet disentangles spatial, temporal, and spectral information and incorporates attention-based interpretability to identify influential inputs and guide efficient acquisition strategies. The framework supports handling multispectral data acquired from comparatively high altitudes with respect to the state of the art (e.g., 60 m with 2.5 cm Ground Sampling Distance (GSD)), and allows the exploration of the trade-off between model performance and GSD. This supports a reduction of flight times and data volumes (e.g., 1.74 GB at 60 m vs. 5.96 GB at 20 m in our reference setup) while maintaining controlled predictive accuracy.We study the case of winter wheat breeding, and extend this approach with new results for the traits drought stress and plant health and a comprehensive analysis of flight height as an operational design variable, systematically simulating and evaluating acquisitions from 20 to 120 m. Results indicate that predictive accuracy is largely insensitive to flight height across this range, supporting higher-altitude, high-coverage monitoring until the release of larger datasets provide a clear justification for lower-altitude, higher-resolution acquisitions. Finally, we translate these findings into practitioner-oriented operational insights for drone-based High-Throughput Phenotyping.

14 March 2026
Verified
Conference Paper
Open access
JLJasmin Lampert
Jasmin Lampert
PFPhillipp Fanta-Jende
Phillipp Fanta-Jende
PTPascal Thiele
Pascal Thiele
LBLorenzo Beltrame
Lorenzo Beltrame
JSJules Salzinger
Jules Salzinger
DDDiogo Duarte
Diogo Duarte
The EMERALD project addresses critical challenges in enhancing forest resilience to climate-driven natural hazards, with a particular focus on the timely detection and monitoring of forest disturbances such as e.g. windthrows. These disturbances are increasingly amplified by climate extremes and pose substantial ecological and economic risks, including biodiversity loss, carbon stock degradation, and cascading impacts on ecosystem services. Despite the growing availability of Earth observation (EO) data, operational forest monitoring remains constrained by cloud cover, terrain-induced shadows, and limited spatial resolution, reducing the reliability of hazard assessment and early response.To overcome these limitations, EMERALD extends SAFIR’s de-clouding and de-shadowing core capabilities and introduces super-resolution methods to enhance the spatial resolution of Sentinel-2 data. More specifically, EMERALD introduces a latent super-resolution approach, in which high-resolution representations are not generated as an end product but as intermediate feature states optimized for downstream hazard-relevant tasks, such as forest disturbance detection, tree species discrimination, and health assessment. The super-resolution component is therefore task-supervised, coupling image reconstruction objectives with performance metrics from downstream applications to ensure that enhanced spatial detail directly translates into improved hazard assessment capability rather than purely visual fidelity.A third core component of EMERALD is the rigorous validation of AI-derived products using high-quality image pairs combining Sentinel-2 observations with very high-resolution Uncrewed Aerial Vehicles (UAV) data for validation purposes. These paired datasets enable quantitative assessment of reconstruction fidelity, uncertainty, and disturbance detectability across spatial scales, strengthening confidence in AI outputs for decision makers. By leveraging datasets from diverse European forest landscapes ranging from Austria to Portugal, EMERALD explicitly addresses geographic transferability and bias, a critical requirement for continental-scale hazard and resilience monitoring.By improving the accuracy, timeliness, and transparency of forest disturbance detection, EMERALD supports AI-enabled decision-making for forest managers and policymakers, demonstrating how advanced digital technologies can enhance resilience to climate-driven natural hazards.