Understanding when trees initiate growth each season is fundamental to linking forest productivity with changing environmental conditions. Yet, capturing this timing at high temporal resolution remains a challenge in forest ecology. Here, we use near-daily close-range laser scanning observations from a permanent laser scanning (PLS) system installed on a flux tower to detect the onset timing and magnitude of height growth in 106 Scots pine individuals within the same forest stand in Southern Finland. We were able to detect the onset of tree height growth at individual tree scale with daily precision by using a break detection method applied to height time series from the PLS observations. We found up to 27 days of variability in the timing of growth onset (132-159 Day of the Year) and up to 33 cm total absolute tree height growth variability (7-40 cm) among the Scots pine during the growing period of 2021. By studying various biotic and abiotic factors that could explain differences in growth dynamics, we found that light availability and microclimate, indicated by competition, local species richness, topography, and elevation, were the most important drivers of height growth onset timing (adj. R2 = 0.10 in linear regression model and adj. R2 = 0.14 in generalized additive model - GAM) while the relative height growth during the season was most impacted by the onset timing and microclimate, which were able to better explain the magnitude of growth in both the linear mixed-effects model (marginal R2 = 0.31) and in the GAM (adj. R2 = 0.50). Consistent with previous findings, our results suggest that variations within forest structure influence tree height growth dynamics. In this study, high spatiotemporal resolution PLS enabled the quantification of conventionally challenging tree functioning variables, such as the height growth onset, thereby offering quantitative evidence to better understand the tree functioning variabilities and the drivers of such variabilities in the context of intelligent forest management under changing climatic conditions.
Forests provide essential ecosystem services beyond timber, yet locating high-yield areas for non-wood forest products such as bilberries (Vaccinium myrtillus) remains a challenge for both recreational and commercial pickers. By integrating Airborne Laser Scanning (ALS), Geographical Information System (GIS) data, and crowdsourced terrestrial imagery analyzed via deep learning (YOLO), we developed a predictive system optimized for identifying high-yield hotspots. We demonstrate that YOLO detection remains highly accurate, but plant height significantly contributes to berry omission. However, this limitation can be mitigated by selecting the maximum berry count from multi-angle terrestrial images. Using a Random Forest classifier across a 36-km² study area in Nuuksio, Finland, we achieved a precision of 58% for the highest yield category. This represents a 20-fold increase in the probability of encountering a high-yield area compared to random searching. Extensive user testing over two years validated the practical utility of the system, showing a 22.5% increase in harvested yield and a 36.5% reduction in time required to locate hotspots. Furthermore, 97% of users reported that the platform provided an accurate big picture of bilberry yield. These results highlight the potential of combining crowdsourced citizen science with advanced LiDAR metrics to create digital twins of forest ecosystems that enhance human interaction with nature and optimize the sustainable harvest of wild food resources.
Mapping individual tree crowns is essential for tasks such as maintaining urban tree inventories and monitoring forest health, which help us understand and care for our environment. However, automatically separating the crowns from each other in aerial imagery is challenging due to factors such as the texture and partial tree crown overlaps. In this study, we present a method to train deep learning models that segment and separate individual trees from RGB and multispectral images, using pseudo-labels derived from aerial laser scanning (ALS) data. Our study shows that the ALS-derived pseudo-labels can be enhanced using a zero-shot instance segmentation model, Segment Anything Model 2 (SAM 2). Our method offers a way to obtain domain-specific training annotations for optical image-based models without any manual annotation cost, leading to segmentation models which outperform any available models which have been targeted for general domain deployment on the same task.
High-resolution water mapping supports environmental monitoring and related applications, but accurate pixel-level labels are difficult and costly to produce. Official hydrographic vectors provide scalable weak supervision, but they contain artifacts like boundary noise, temporal mismatch, and omissions of small water structures. We propose a two-stage framework for weakly supervised water segmentation in high resolution multispectral imagery. Stage 1 learns initial masks from rasterized vector pseudo-labels, and Stage 2 converts these masks into structured component-wise prompts for localized refinement. On a manually corrected validation set, refinement improves SegFormer-B0 from 0.9509 to 0.9535 IoU and U-Net from 0.9408 to 0.9486 IoU, with corresponding F1 gains from 0.9749 to 0.9762 and 0.9695 to 0.9736. It leads to sharper shorelines, reduced boundary spillover, and better thin-structure delineation. The results indicate that prompt-guided refinement can improve pseudo-label-based water segmentation by targeting local errors that are poorly captured by global training supervision.
Machine learning (ML) methods are increasingly explored for Global Navigation Satellite Systems (GNSS) spoofing detection, as they can potentially identify transmitter-specific hardware imperfections (RF fingerprints) that differentiate genuine from spoofed signals. However, robust cross-testing, meaning the training on one dataset and testing on another, remains a major challenge. Differences in environments, receiver hardware, signal processing chains, and attack implementations lead to distribution shifts that significantly limit model generalization. In particular, commonly used features such as carrier-to-noise ratios and Doppler are highly sensitive to both hardware and processing pipelines, while the lack of standardized datasets further exacerbates inconsistencies. As a result, ML models often overfit to dataset-specific artifacts rather than learning general spoofing indicators, leading to poor cross-testing performance. This paper presents a detailed cross-testing analysis using three open-access GNSS spoofing repositories (TEXBAT, OAKBAT, and FGI-SpoofRepo), each containing multiple datasets. We evaluate two ML approaches: a conventional Convolutional Neural Network (CNN) and a Model-Agnostic Meta-Learning (MAML) approach recently introduced by our team for GNSS applications. Results show that the best cross-testing performance is achieved by the MAML approach using post-correlation Doppler features. Moreover, training on more complex spoofing scenarios (e.g., TEXBAT DS8) consistently improves generalization across datasets compared to training on simpler attacks. Finally, the analysis provides insights into combining pre-and post-correlation features, highlighting directions for improved feature extraction at different stages of a GNSS receiver.
