Conference Agenda

General-purpose spatial representations are essential for building transferable geospatial foundation models (GFMs). Among them, the AlphaEarth Foundation (AE) represents a major step toward a global, unified representation of the Earth's surface, learning 10-meter embeddings from multi-source Earth Observation (EO) data that capture rich physical and environmental patterns across diverse landscapes. However, such EO-driven representations remain limited in capturing the functional and socioeconomic dimensions of cities, as they primarily encode physical and spectral patterns rather than human activities or spatial functions. We propose AETHER(AlphaEarth–POI Enriched Representation Learning), a lightweight framework that adapts AlphaEarth to human-centered urban analysis through multimodal alignment guided by Points of Interest (POIs). AETHER aligns AE embeddings with textual representations of POIs, enriching physically grounded EO features with semantic cues about urban functions and socioeconomic contexts. In Greater London, AETHER achieves consistent gains over the AE baseline, with a 7.2% relative improvement in land-use classification F1 and a 23.6% relative reduction in Kullback–Leibler divergence for socioeconomic mapping. Built upon pretrained AE, AETHER leverages a lightweight multimodal alignment to enrich it with human-centered semantics while remaining computationally efficient and scalable for urban applications. By coupling EO with human-centered semantics, it advances geospatial foundation models toward general-purpose urban representations that integrate both physical form and functional meaning.

This work demonstrates the potential of foundation-model-encoded satellite embeddings to bridge Earth's surface and atmosphere. Based on our multimodal foundation model Copernicus-FM, we curate a global embedding dataset at 0.25°x0.25° resolutions (in consistency with ERA5). For each grid, multi-sensor images from Sentinel-1, 2, 3, 5P, and DEM are encoded into an embedding vector. These grid embeddings serve as condensed surface representations for downstream users. We verify their benefits as input for a climate task that predicts the 10-year mean and standard deviation of several climate parameters (e.g., precipitation) from ERA5. Compared to raw coordinates or location encodings, our results suggest that introducing surface embeddings helps produce more accurate prediction maps, reducing RMSEs by an average of up to 45%.

Ground-level horizon imagery is scarce across planetary bodies, making representation-centred approaches attractive for downstream geospatial tasks. We present DESPINA, a geospatial representation system that converts digital elevation models (DEMs) into structured neural embeddings of terrain geometry that condition a diffusion model to produce geometry-preserving, terrain-consistent visual reconstructions for a specified location and view direction.

Our pipeline integrates numeric elevation data (DEMs), structural embeddings (inverse-depth and soft edges), and textual priors, unifying heterogeneous geospatial signals into a shared, metric conditioning space. Using a Stable Diffusion model constrained with ControlNet, we can generate geologically consistent yet texturally diverse horizon datasets. Appearance priors are learned from historical surface photography to capture realistic textures and lighting cues, and geometric validation is performed against DEM-derived skylines and depth structure, independent of photographic training data. Through quantitative evaluation and a pilot qualitative study, DESPINA maintains skyline fidelity and geological boundaries while improving structural similarity relative to an image-conditioned baseline. Although our experiments use lunar DEMs and historical surface photography, the method is domain-agnostic and applicable to Earth, Mars, and other planetary DEMs.

Satellite -based crop classification and maps are important tools for food security and climate change mitigation, but existing approaches are not effective for small field systems. To address this, crop type classification using embeddings generated by a global foundation model, TESSERA, are compared to standard classification approaches in the literature. We find that our embedding -based approach offers a triple win: 1) consistent and statistically significant performance improvement over current methods, 2) greater simplicity due to the elimination of feature engineering, and 3) the reduction of computational cost. Our embedding -based approach achieves significantly higher F1 scores in the classification of 5 of 7 crop types for small fields in Austria (over 10% improvement in one case). Additionally, the TESSERA embedding -based method uses 8% of compute compared to the raw data method. These results indicate that embeddings are an effective approach for crop type classification tasks in small field systems.

This study explores the potential of the AlphaEarth Foundation (AEF) embeddings, a global, annual, analysis-ready satellite embedding dataset, for winter wheat and crop rotation mapping. Firstly, we analyze AEF embeddings for intra-class consistency and inter-class separability, assessing their effectiveness in representing wheat within the semantic embedding space. Subsequently, we compare multiple lightweight classifiers to identify an optimal model and conduct spatiotemporal generalization experiments across Henan Province from 2018 to 2024 using only a limited set of labelled samples from 2020. Based on the resulting wheat distribution maps, crop rotation patterns are further identified.Experimental results demonstrate that AEF embeddings exhibit strong semantic coherence and discriminative capability. Acceptable classification accuracy (OA = 0.85) can already be achieved using simple models such as cosine similarity and linear regression. More advanced lightweight classifiers further improve the performance (OA = 0.86–0.93) while maintaining stable results across different years and regions (spatial consistency = 0.82). In addition, the crop rotation maps show high spatial agreement with existing products, while producing more spatially contiguous field patterns.Overall, this study confirms that AEF embeddings can serve as effective, ready-to-use features for large-scale agricultural remote sensing applications. By substantially reducing the reliance on complex feature engineering and extensive training samples, they provide a practical and scalable solution for mapping winter wheat and its crop rotation patterns.

The goal of the CO3D (Constellation Optique 3D) mission is the full-automatic production of a worldwide accurate DSM (Digital Surface Model). This DSM is generated from stereo acquisitions obtained from a new generation of high-resolution optical satellites, called CO3D. The DSM accuracy is one meter in relative height for moderate slopes and four meters in absolute height with a one-meter grid space. Each of the four satellites of the constellation provides images with 0.50 m resolution in red, green, blue and near-infrared bands. The satellites resource is shared by, on one hand, the French institutions (government, scientists concerned by global Earth monitoring) who have dedicated access and preferred price conditions, and on the other hand, ADS (Airbus Defence and Space) customers interested in 2D and 3D products. The constellation was launched on July 26, 2025

We present the qualification of 3D products as part of the CO3D mission. The CO3D mission is dedicated to creating a digital surface model of the Earth's landmass cover. Massive automatic production is a challenge in itself, as this ground segment produces more advanced data for an optical mission. Firstly, the 3D product, which is generally retouched and checked, which can represent a significant cost. In the case of the CO3D mission, these products will be generated completely automatically. Masks will also be included to describe the processing history and provide precise information on the altitudes restored. All of this data requires detailed qualification with precise reference data and methods to best reflect its quality. The paper will describe all of these methods and data and provide an overview of the performance of these new CO3D products.

The launch of the four CO3D spacecrafts took place on 26th July 2025 aboard Vega-C from Kourou space center with Microcarb microsatellite. After a first week spent calibrating the most critical subsystems of the spacecraft, the instrument was switched on, enabling the 9-months Image Quality commissioning phase to begin. Images are acquired in RGB and NIR spectral bands with a 50 cm Ground Sampling Distance (GSD), thanks to matrix sensors based on a Bayer pattern. This brings new calibration challenges such as demosaicing and 2D line of sight determination. 3D calibration activities take place in a second stage of the commissioning phase once radiometric and geometric calibration are finalized.

This work introduces CO(ast)3D, a predictive pipeline that helps identify when upcoming CO3D satellite overpasses are likely to capture surface wave signals suitable for bathymetry acquisitions. Because CO3D cannot directly image the seafloor, depth must be inferred from the optical expression of surface waves, whose visibility depends strongly on illumination, viewing geometry, and sea state.

The pipeline combines CO3D orbital tracks with forecasted wave parameters from the Copernicus Marine model to construct a directional wave spectrum, generate a time-varying free surface through linear wave theory, and simulate CO3D-like radiance images at native spatial resolution. These synthetic scenes allow the clarity of the wave field to be evaluated a priori for any future time and location.

By predicting whether conditions will yield a sufficiently coherent wave signal, the system supports more efficient tasking, reduces acquisition risk, and improves the likelihood of capturing images suitable for accurate bathymetric inversion.

The challenges, main design elements, and results of the CODIP and ICC components will be presented in this paper, as well as the modalities for accessing and using the CO3D products.

High-resolution Earth observation data are crucial for applications such as agriculture, urban planning, and environmental monitoring. Although commercial satellites provide sub-meter imagery, open-access alternatives like Sentinel-2 are limited to resolutions around 10~m ground sampling distance, which is insufficient for many tasks. In this work, we investigate image super-resolution as a method to bridge this gap, enhancing downstream performance on freely available satellite data. We leverage two 16-bit single-band datasets, consisting of Sentinel-2 (20m --> 10m) and Venus (10m --> 5m) images, to train and benchmark state-of-the-art SR methods, including transformer- and diffusion-based approaches, across multiple dataset mixes. These models are evaluated quantitatively using reference-based metrics (PSNR, SSIM) using ground-truth and no-reference scores (FID, NIQE) for native upscaling from 20m --> 10m and 10m --> 5m. We observe that different SR architectures present trade-offs between standard quantitative metrics and perceptual image quality. We further assess their impact on a practical downstream task: field boundary detection from Sentinel-2 imagery. Our experiments demonstrate that SR pre-processing improves quantitative fidelity and downstream task performance, enabling low-resolution satellites to compete more effectively with commercial imagery

Current diffusion models struggle to achieve fine-grained remote sensing imagery (RSI) generation. This limitation fundamentally stems from their reliance on "flattened" text prompts, which overlook the inherent hierarchical structure of RSI. This paper proposes a fine-grained RSI generation method driven by expert knowledge and hierarchical captions. We first deconstruct RSI into a hierarchical "element-relation-scene" caption and employ an automatic caption optimization mechanism, grounded in spatial relation knowledge, to ensure high fidelity. Critically, we introduce a novel hierarchical caption encoding mechanism that efficiently injects decoupled hierarchical caption segments into the U-Net's cross-attention layers. This design enables the model to exert hierarchical and decoupled attentional control over the global scene, spatial layout, and geographical element details during the denoising process. Experiments demonstrate that, when combined with efficient fine-tuning algorithms such as LoRA, our method significantly outperforms traditional single-level captions across all six evaluation metrics, exemplified by the FID metric decreasing from 228.43 to 205.59 and the GSHPS metric increasing from 0.86 to 0.92. This research provides a new paradigm for controllable remote sensing scene generation, establishing an effective link between hierarchical semantic understanding and the progressive generation process of diffusion models.

Semantic segmentation of remote sensing images has attracted considerable attentions. For cross domain semantic segmentation, the images captured at different times inevitably exhibit significant domain and feature gaps. Besides, the labels are precious, given that acquiring adequate annotations is time-consuming and laborious. There are numerous methods to cope with these problems, for example, semi-supervised, weakly supervised learning help for the lack of label, while style transfer and domain adaptation are effective for domain gaps. However, the outcomes are still not ideal. Nearly all methods ignore the combination of image-level alignment and feature-level alignment, while few methods consider class-wise constraint to boost the performance. Towards this end, IFSDA, an image-level and feature-level semantic-aware architecture for cross domain semantic segmentation is put forward. In order to acquire sound outcomes, two branches of alignment strategies are realized by self-supervised learning and generative adversarial learning. Besides, a novel semantic discriminator is utilized in image translation process to optimize class-related information, thereby helping to eliminate the intra-class domain gaps between bi-temporal images and optimize the segmentation results effectively. Experiments on ISPRS 2D Semantic Labeling Contest Dataset have shown the superiority of proposed method over other models.

Reliable estimates of seabird colony size are essential for monitoring population dynamics, yet accurate counts are difficult for expansive colonies on remote Arctic cliffs. Northern fulmars (Fulmarus glacialis) breed in large, unevenly distributed aggregations across extensive, towering cliffs in the Canadian Arctic, posing numerous survey challenges. Side-looking helicopter photo surveys generate thousands of photos where birds are small, variably angled, and of inconsistent sharpness against large, complex backgrounds. We used deep learning to automate fulmar counts in this imagery. Our objectives were to (1) develop an object detection model trained on manually annotated imagery sampled from three Arctic colonies, (2) evaluate model performance, and (3) estimate total size of an entire colony. We trained a YOLOX-based model on >16,000 annotated birds, following a two-stage training approach for small objects interspersed across expansive and heterogenous backgrounds. Compared to ~20,000 additional manual annotations in a sample of the Cape Liddon colony in the territory of Nunavut, the model detected 90% of birds with a 9% false-positive rate (i.e. 90% recall, 91% precision). The model's detection sensitivity was calibrated to achieve a ~1:1 ratio between total model detections (true positives + false positives) and the 'true' count, which required manual annotation of ~15% of the colony imagery. Overall, the model detected 38,723 fulmars across the entire colony, providing a robust estimate of its full population. These results highlight deep learning’s potential to greatly streamline and scale up seabird monitoring in remote polar environments where conventional surveys are constrained.

Air pollutants such as nitrogen dioxide (NO₂) have detrimental effects on human health and ecosystems. It is therefore very crucial to pinpoint the location of high pollutant concentrations over large areas. Ground-based stations, while offering continuous temporal measurements, cannot provide broader spatial coverage for regions like cities. This study uses Tropospheric Emissions: Monitoring Pollution (TEMPO) satellite observations and a machine learning model to estimate high-resolution surface-level NO₂ concentrations over the Greater Toronto Area (GTA), Ontario, Canada. The random forest regression model was trained with input parameters such as hourly tropospheric NO₂ vertical column density (VCD) values and boundary layer height (BLH), which are the two most effective parameters in feature importance. The model achieved a coefficient of determination (R²) of 0.84, a root mean square error (RMSE) of 1.703 µg/m³, and a mean absolute error (MAE) of 0.939 µg/m³, indicating strong and reliable predictive performance. The findings of this research can support air quality forecasting, public health studies, and urban planning decisions, especially in regions with scarce ground-based pollutant data.

Automatic updating of topographic maps remains a significant challenge, as current workflows still rely heavily on manual interpretation of airborne data. This study proposes a method for identifying topographic changes by learning object representations from existing maps and using them as reference data for change detection. Map-derived labels are used to train independent 2D and 3D segmentation networks that generate semantic predictions from orthoimages and point clouds. Unlike conventional change-detection approaches that require temporally aligned datasets of the same modality, the proposed method directly compares newly acquired airborne data with existing map vectors. Semantic predictions from both modalities are vectorized and selectively fused into polygon geometries, which are subsequently compared with reference map vectors to identify object-level "from–to" changes. The workflow highlights potential change regions and their predicted semantic classes, allowing operators to focus inspection on relevant areas rather than the entire dataset. Detected changes include both real-world developments, such as new construction and demolitions, and inconsistencies in the reference map caused by outdated or inaccurate delineations. To assess the effect of multimodal integration, the workflow is compared with a 2D-only baseline. The results indicate that integrating 3D geometric information can reduce noisy detections and improve the spatial consistency of candidate change objects, particularly for water and bridge classes.

High-resolution (HR) satellite imagery is essential for urban monitoring and disaster management, but its use is constrained by high cost and limited accessibility. Super-resolution (SR) offers an efficient alternative by reconstructing high-quality images from low-resolution (LR) inputs, making large-scale geospatial analysis more feasible. We propose the Multi-Attention Pyramid Super-Resolution Network (MAPSRNet), which delivers two main innovations:

1. A multi-attention model that integrates a Pyramid Vision Transformer for long-range spatial dependencies with a cross-channel Involution+ module to enhance feature interactions, generating SR images with superior structural preservation and sharper boundaries.

2. The first SR network to surpass the performance of original HR images in downstream tasks, demonstrated through building detection with a ConvNeXtV2 backbone and U-Net decoder. MAPSRNet reduces false positives and negatives and, across multiple datasets, exceeds HR performance in IoU, F1-score, and overall accuracy.

Extensive experiments on the Massachusetts building dataset, the Wuhan University building dataset, and the Waterloo building datasets confirm that MAPSRNet consistently outperforms representative SR methods in both image fidelity (PSNR, SSIM) and task-level metrics. Its ability to preserve fine structural details, suppress background noise, and learn resolution-invariant features through multi-resolution training makes the reconstructed images more task-aware than raw HR data. Beyond buildings, this flexibility suggests strong potential for generalization to other land-cover classes such as roads, vegetation, and water bodies.

These results establish MAPSRNet as a cost-effective alternative to HR acquisitions and a milestone in task-driven SR research, advancing both image reconstruction and downstream geospatial analysis.

High revisit-rate SAR constellations generate large volumes of imagery that require consistent geolocation accuracy to support applications such as change detection and interferometry. However, variations in orbit determination, attitude knowledge, and external factors such as Global Navigation Satellite System (GNSS) interference can introduce geolocation errors that vary across acquisitions, making large-scale validation challenging. This study presents an automated approach to detect and quantify geolocation offsets in ICEYE SAR imagery by aligning orthorectified scenes with reference images using feature-based matching and correlation-based refinement. The method is validated against independently derived absolute geolocation measurements from corner reflector calibration sites in the United States, Canada, Australia, and Poland. Evaluation across 726 acquisitions demonstrates strong agreement with reference measurements, achieving an overall root-mean-square error (RMSE) of 1.39 m, with RMSE values of 1.18 m for Spotlight mode and 1.93 m for Stripmap mode. Operational applicability is demonstrated through large-scale acquisition campaigns, including nationwide Stripmap coverage over Japan and coherent image stack analysis. The results show that the proposed method can reliably estimate geolocation offsets, detect anomalies, and monitor geometric consistency across large SAR archives, providing a practical and scalable solution for automated geolocation quality control in micro-satellite SAR constellations.

Urbanisation and land-use change increase the demand for temporally consistent urban maps from high-resolution Earth observation imagery. A key obstacle is label ageing: benchmark annotations are often years older than current true orthophotos (TOP), causing semantic and geometric mismatches (e.g., demolished/new buildings, shifted vegetation boundaries) that degrade supervised learning, calibration, and transfer. This paper presents a probabilistic, quality-aware segmentation framework based on a compact U-Net. Legacy annotations are converted into edge-adaptive soft labels to encode boundary uncertainty. A HELIX-derived per-pixel supervision quality score Q is computed and integrated as a weight in a Q-weighted Kullback--Leibler objective with an agreement-focal component, reducing the influence of unreliable or outdated regions. Global temperature scaling is then applied to obtain calibrated per-class probability fields with comparable confidence magnitudes. Experiments on ISPRS Potsdam and Vaihingen combined with recent (2024) TOPs evaluate temporal transfer (archival supervision vs. updated imagery of the same area) and spatial transfer (cross-city application). Finally, calibrated probability fields are used to derive probabilistic semantic transitions and temporal reliability scores, supporting uncertainty-aware mapping of urban change such as construction, sealing, and vegetation loss.

Survivorship bias refers to the fact that conclusions are drawn from a non-representative sample limited to cases that have survived a selection process. This article shows that this bias affects scientific literature, which tends to select successful experiments and hide failures. Remote sensing, like other data-driven sciences, is affected by survivorship bias, making it difficult to have a clear idea of the data's and methods' actual potential and limitations. A typology of failure causes is proposed to encourage critical reading of the bibliography, and perspectives are outlined to overcome survivorship bias by improving practices within the academic and industrial remote sensing communities.

Satellite remote sensing is inherently constrained by a fundamental spatio-temporal trade-off by physical sensor limitations. Super-Resolution (SR) techniques are required to overcome these constraints and obtain high-resolution time-series data. However, Single Image Super-Resolution (SISR) provides insufficient information for robust restoration. To address this, Reference-Based Super-Resolution (Ref-SR), which utilizes a high-resolution (HR) reference (Ref) image, has been investigated. Nonetheless, Ref-SR introduces the challenge of reference misuse, stemming from the temporal mismatch (or inconsistency) between the target low-resolution (LR) image (e.g., clouds, seasonal changes) and the Ref image (often a long-term median composite). To address this reference misuse problem, this study proposes an adaptive Ref-SR framework that incorporates a similarity weight map derived from the LR and Ref information. This weight map is computed solely from the pixel-wise similarity between the LR and Ref inputs, requiring no ground truth HR, and functions as a gating mechanism. This allows the network to dynamically control Ref reliability, guiding it to suppress Ref influence in mismatched regions and leverage its textures in similar ones. Validation experiments using Sentinel-2 data (LR 240m, Ref/HR 60m) demonstrate that the proposed method achieves significant performance improvements over SISR in both spatial (Peak Signal-to-Noise Ratio, Structural Similarity Index) and spectral (Spectral Angle Mapper, Error Relative Global Dimensionless Synthesis) metrics. Furthermore, qualitative analysis confirms that the framework effectively suppresses artifacts caused by the blind injection of Ref textures in inconsistent areas. This framework could contribute to the future fusion and quality enhancement of heterogeneous LR sensor data, such as GOCI-II.

The Copernicus Land Monitoring Service (https://land.copernicus.eu) has been providing geophysical data derived from Earth Observation (EO) at a global scale for several decades. This global dataset includes temperature and reflectance, vegetation, soil moisture, snow and water bodies variables. To ensure the quality of these dataset, yearly validation assessment is performed. The collection and processing of ground data for the purpose of validating Copernicus products represents in itself a huge task. In 2018, the European Commission (EC) has established a new service to ensure the independent production of these data: Ground-Based Observations for Validation (GBOV) https://gbov.land.copernicus.eu). The prime objective of GBOV has been for the last 8 years, to provide high-quality validation data for seven Copernicus Land Monitoring Service core products:

• Top Of Canopy Reflectance (TOC-R),

• Albedo (ALB),

• Leaf Area Index (LAI),

• Fraction of Absorbed Photosynthetically Available Radiation (FAPAR),

• Fraction of Vegetation Cover (FCOVER)

• Surface Soil Moisture (SSM) and

• Land Surface Temperature (LST).

In its third phase, new product have been included to support the growing Copernicus land products portfolio, namely:

•GPP and NPP

•Phenology

•Evapotranspiration

GBOV includes three components in the service:

•Component 1: consists of using data from existing in situ networks to generate EO validation datasets. Multi-year ground-based observations of high relevance for EO are collected from these global networks.

•Component 2: consists of upgrading existing monitoring sites with new instrumentation or establishing entirely new monitoring sites to close thematic or geographic gaps.

•Component 3: deals with data distribution of the validation dataset to the user community.

Light Detection and Ranging (lidar) has become an essential tool to quantify forest structure in three dimensions, allowing extraction of tree-level metrics such as height, crown volume, diameter at breast height (DBH), and biomass. Accurate forest structure quantification supports applications such as wildfire management, biodiversity assessment, forest health monitoring, and timber management. This is particularly urgent in countries such as Canada, where wildfires pose a significant challenge to forest management due to their increasing frequency and severity; advanced fire behavior models aid wildfire preparedness by predicting fire behaviour at fine-scale in 3D but require detailed 3D fuel information including canopy and ladder fuels.

Terrestrial Laser Scanning (TLS) and Uncrewed Aerial Vehicle (UAV) lidar provide dense point clouds that allow highly accurate characterization of individual trees, critical for assessing forest attributes and wildfire fuel characteristics. However, their limited spatial coverage makes them neither time- nor cost-effective for mapping extensive forested regions. Airborne Laser Scanning (ALS), in contrast, covers broad areas efficiently by collecting data from higher altitudes, but at the cost of lower point densities (typically 1–100 points/m²), insufficient for precise individual tree characterization.

To address this challenge, this study reconstructs densified tree point clouds from low-resolution ALS data using an upsampling framework based on a deep generative network trained on real and synthetic datasets. This approach bridges the gap between ALS’s extensive coverage and the detailed structural information provided by TLS and UAV lidar, enabling accurate, large-scale quantification of forest structure for applications such as wildfire management and monitoring.

Accurate tree positions play a vital role in precision forestry and environmental sciences. In this study, we propose an accurate, efficient, and adaptable method for tree localization by integrating heading, diameter at breast height (DBH), and ultra-wideband technology. The proposed method is simple to implement in different forest environments and can determine the position of each tree within a few seconds. Compared with traditional field measures, such as laser rangefinders and inclinometers, the proposed approach is more efficient. In comparison with commonly used measures, such as terrestrial laser scanning (TLS) and mobile laser scanning (MLS), the proposed method is more cost-effective and easier to implement, making it particularly suitable for natural forests that are remote from roads yet require accurate measurements. Field experiments were conducted in a managed boreal forest in southern Finland, characterized by minimal understory vegetation and good visibility, where a total of 50 trees were mapped. Experimental results indicate that the proposed method can accurately determine tree positions with an RMSE of 0.12 m and an MAE of 0.11 m.

Plant phenotyping has become a fundamental tool in modern agronomic research, enabling quantitative analysis of morphological characteristics that can be collected in three dimensions using photogrammetric techniques or point clouds obtained by LiDAR systems. However, automatic segmentation of plants, especially the main stem and its branches, still poses a challenge for certain crops. This work proposes a non-destructive, geometry-based methodology for morphological phenotyping of tomato plants (Solanum lycopersicum) using photogrammetric point clouds. The proposed methodology consists of the following steps: stratification of the plant into horizontal sections; clustering of each stratum using the DBSCAN algorithm; selection of clusters based on the linearity tensor derived from eigenvalue analysis; and the fitting of a 3D cylinder to the linear clusters to approximate the main stem. The method was validated using manually labeled point clouds from nine tomato cultivars, achieving accuracy between 88% and 97%, with average F1-scores of 63.6% for the stem and 96% for the branches

Linking remotely sensed forest backscatter with fine-scale tree crown structural dynamics provides insights into tree growth strategies under varying conditions. In this study, we investigate whether branch-scale tree growth allocation dynamics, derived from multi-temporal TreeQSM models, are reflected in SAR and ALS observations. We analyzed branch organization dynamics of silver birch (Betula pendula) using terrestrial laser scanning data from 2021, 2023 and 2025 at a boreal forest site in southern Finland. Branch allocation metrics, including volume-weighted mean diameter (VWMD), small branch fraction (SBF), distal volume fraction, relative branch height, and top canopy volume, were quantified to capture shifts between structural reinforcement and exploratory growth. These metrics were compared with Sentinel-1 SAR features (α, entropy, C11, C22) and ALS-derived canopy metrics (plant area index, vertical complexity index) alongside local structural variables. Results show a consistent trade-off between coarse and fine branching, with strong negative correlations between ΔVWMD and ΔSBF across both periods (ρ = –0.92). SAR-derived α exhibits strong associations with these allocation shifts during 2021–2023 (ρ = –0.81 with ΔVWMD; ρ = 0.75 with ΔSBF), indicating sensitivity to internal redistribution of branch material. ALS metrics from 2021 reflect initial canopy structure and are associated with subsequent allocation shifts. Despite the small magnitude of observed changes, consistent monotonic relationships across datasets suggest that subtle within-crown branch allocation is detectable from satellite and aerial observations, reflecting the surrounding canopy context. However, weakened correlations in 2023–2025 highlight the influence of external factors on SAR signals.

Tree instance segmentation of airborne laser scanning (ALS) data is of utmost importance for forest monitoring but remains challenging due to variations in the data caused by factors such as sensor resolution, vegetation state at acquisition time, terrain characteristics, etc. Moreover, obtaining a sufficient amount of precisely labeled data to train fully supervised instance segmentation methods is expensive. To address these challenges, we propose a weakly supervised approach where labels of an initial segmentation result obtained either by a non-finetuned model or a closed-form algorithm are rated by a human operator. The labels produced during the quality assessment are then used to train a rating model, whose task is to classify a segmentation output into the same classes as specified by the human operator. Finally, the segmentation model is finetuned using feedback from the rating model. This in turn improves the original segmentation model by 34% in terms of correctly identified tree instances while considerably reducing the number of non-tree instances predicted.

Trees play a key role in our planet. They regulate climate, preserve biodiversity, and contribute to human well-being. Each species has different contributions to our globe and a specific carbon storage potential. Identify tree species enable better measurement of global carbone, help authorities for better manage forests and green spaces. Unmanned Aerial System (UAS) LiDAR has become a powerful source of 3D point cloud for vegetation analysis, given its ability to captured large area in a short time and its capacity to penetrate canopy layers. Deep learning methods extract discriminative features directly from raw point clouds and generalize well to unseen datasets. This study optimises PointNet++ deep learning architecture for tree species classification by analysing the influence of sampling configurations on the performance of model detection, by using an open-source dataset “FOR-species20K”.Three-point cloud sampling configurations (4 096, 8 192, and 16 384 points per tree) were tested with three random seeds (0,42 and 123) to assess their impact on classification accuracy and ensure stability of prediction. Results on a separate test set of 508 trees show a consistent improvement in performance of PointNet++ with a sampling configuration of 8 192 points per tree, reaching a macro-average F1-score of 89.65%, surpassing the 74.9 % reported by (Puliti et al., 2025) for evaluating the same architecture. Dominant species such as Fagus sylvatica, Picea abies, and Pinus sylvestris achieve F1-scores exceeding 90%, indicating high model robustness.

Terrestrial laser scanning technology can efficiently acquire high-precision three-dimensional spatial information in complex forest environments, making it an important technical means for detailed analysis of forest structure and resource monitoring. However, traditional terrestrial laser scanners planning methods are prone to coverage gaps and data redundancy due to factors such as tree obstructions, terrain undulations, and canopy overlap, making it difficult to simultaneously balance observation completeness and scanner station deployment cost. To address this, this paper proposes an intelligent survey station planning for terrestrial laser scanners in complex forest environments. The method first uses airborne LiDAR data to build a prior forest model, which is then used to quantitatively evaluate forest visibility features by calculating the cumulative visible central angle through visibility analysis. Finally, an integer linear programming model is further introduced to achieve global optimization of the station set based on an initial feasible coverage solution obtained using a greedy algorithm. To test the performance of the proposed method, this paper applies the proposed method to the forest plot located in Lushan city, Jiangxi province, China. Experimental results indicate that the proposed method achieves an overall coverage rate of 94.55% with only seven stations, reducing the number of stations by approximately 30% and 22% compared with the greedy algorithm and genetic algorithm, respectively. The results demonstrate the effectiveness and superiority of this method for station planning in complex forest areas and provide efficient and precise technical support for forest structure monitoring and spatial information acquisition.

Reflectance is a widely used feature for all types laser scanning data. Thus, the accuracy and improvement of the reflectance parameter is a persistent topic of research. For short laser pulses with medium-sized footprints, previous work has investigated the effects of inclined targets on the recorded waveform of full-waveform LiDAR systems. In this work, a new methods to extract incidence angle from only a single waveform can be leveraged to improve reflectance values through recalculation based on the laser-radar equation and correcting for angle of incidence artifacts. The results of the proposed method are evaluated with two datasets based on two different topo-bathymtric laser scanners. For both systems, we calculated the relative biconical reflectance and relative averaged bidirectional reflectance distribution function (rBRDF) and evaluated them on homogeneous roof faces. The two reflectance measures are then compared to the initial reflectance values of the laser scanners used in the study. Both measures showed improvements compared to the standard values. The biconical reflectance shows the best overall mean score for all surveyed roofs with an MAD improvement of 0.80 dB to 62 dB for Sensor I and 0.61 dB to 0.56 dB for Sensor II, in addition the rBRDF also displays an improvement with varying results depending on the deployed system. These results highlight the advantages of the proposed reflectance measures and the potential improvement of the widely used LiDAR attribute.

Accurate classification of topo-bathymetric LiDAR data remains challenging due to the heterogeneous nature of land-water transitional environments, where terrestrial, water surface, and submerged features must be distinguished simultaneously. This study presents a multi-branch deep learning architecture for classifying bathymetric LiDAR data into different classes: soil ground, trees and vegetation, water surface, seabed, aquatic plants and other underwater objects (dead wood, coral reef). The proposed framework employs three parallel feature extraction branches, while the first branch captures spatial structure by focusing on three-dimensional geometric coordinates (XYZ), the other two branches use two independent 1D U-Net architectures to extract signal-based features from RGB spectral reflectance and waveform-derived attributes (intensity, return number, number of returns). The discrete LiDAR attributes, though represented as point-wise numerical values, preserve signal characteristics derived from full-waveform analysis. The encoder-decoder of 1D U-Net architecture with skip connections effectively captures sequential patterns and multi-return patterns in different classes especially in vegetation canopies. The three feature streams are fused through fully-connected layers before final classification. Evaluation using different metrics demonstrates the capability of the framework to simultaneously classify diverse coastal zone and inland waters contexts spanning terrestrial and submerged domains within a unified processing pipeline, eliminating the need for separate terrestrial and bathymetric classification workflows.

Permanent laser scanning (PLS) is an effective tool for near-continuous monitoring of topographical changes in beach-dune systems. While PLS systems were traditionally costly, the emergence of affordable LiDAR sensors enables larger-scale setups with multiple scanners or sites. However, the different characteristics compared to high-end devices, create challenges for one-on-one replacement. To assess how low-cost sensors can replace high-end sensors, we compare the performance of a setup with several low-cost Livox AVIA sensors to a single high-end RIEGL VZ-2000i sensor in its ability to capture an embryonic dune field with large variation in topography. This is evaluated using HELIOS++ virtual laser scanning (VLS). To also assess the representativeness of the simulations, we further compare the VLS to real-world measurements with the Livox AVIA. Based on a VLS setup with six AVIAs mounted on tripods at 2 m above ground, a coverage of 52% can be obtained, which is similar to the coverage of a single RIEGL VZ-2000i on a tower 8 m high. The real-world experiments confirm the VLS results with a slightly lower point cloud coverage of 42%. Furthermore, the effective range of the Livox AVIA in a beach-dune system lies around 100-150 m. At larger ranges, only pulses at high incidence angles (angle between surface and incoming beam, >20°) are registered at the scanner. The variations in coverage between the VLS and real-world scans highlight the need for careful consideration of the occlusion potential of different representations of the topography, beam divergence shapes, and the moisture conditions.

The aim of this abstract is to assess the accuracy of the trajectory of the handheld laser scanner CHCNAV RS10. The trajectory data of this PLS device is compared with a simultaneously measured total station measurement.

Green-wavelength LiDAR systems enable high-resolution 3D sensing in underwater environments, but the geometric evaluation of measurements across the waterline remains difficult. A main challenge is that traceable reference instruments usually operate only in air, while refraction at the air-water interface systematically affects both the reconstructed 3D point cloud and the geometry of partially submerged objects. To address this problem, this study presents a controlled experimental framework for evaluating waterline-induced effects in an Underwater LiDAR (ULi) system, using the Z+F IMAGER 5016A as an in-air reference. A rigid reference frame (RRF) spanning the waterline was deployed in a swimming pool. The RRF was first scanned by the IMAGER in air to establish the reference geometry and was then measured by the ULi system under waterline conditions. The analysis considered the above-water, cross-waterline, and underwater parts of the RRF. The evaluation was based not only on overall geometric deviations but also on rigid-body-invariant internal quantities, especially pairwise distances that are independent of the pose of the RRF. In addition, the sensitivity of the reconstructed geometry to the refractive-index setting used in processing was assessed by perturbing the refractive index and quantifying the resulting changes. The proposed workflow provides a practical and traceable basis for isolating and evaluating waterline-related refraction effects in controlled ULi experiments.

The intensity, frequency, and duration of wildland fires are growing in Canada and around the world. Timely fire intelligence products from remote sensing platforms can assist fire managers and lead to fewer impacts. New onboard processing techniques using machine learning allow greater levels of analysis and refinement on edge devices like aircraft and satellites, reducing bandwidth and latencies. Our Fire Band Analysis Network approach brings together wildfire science and management experts and academics, an Indigenous owned business that specializes in satellite communication and community outreach, and a Canadian space company with expertise in deploying machine learning models to spacecraft. We show initial results with onboard segmentation models and present a path to prototype this onboard processing model on a cubesat currently in orbit and on drones equipped with infrared sensors, ultimately bringing the derived data products to user communities on the ground.

This presentation will provide an overview and update on the WildFireSat mission and its data product algorithm development. Specifically, we will summarize the 2025 Science and Applications Plan and share updates from the Tier 2 stage of products and algorithms. The Tier 2 products that will be shown include the multi-source fire events, time of arrival outputs, and satellite-derived fire behaviour products (e.g., satellite-observed rate and direction of spread, fireline intensity). Ongoing science-development activities include algorithmic validation, uncertainty characterization, and completion of algorithmic theoretical basis documents.

Built through Canadian and international partnerships, WildFireSat will support fire monitoring and management while enabling major scientific advancements for the global fire monitoring community. The scientific applications of WildFireSat are broad, covering all stages of a fire event’s life cycle. By prioritizing the needs of wildfire managers and a broad range of end-users, the WildFireSat mission is a strong model for future satellite missions to integrate user engagement throughout all phases of the mission timeline.

Normalized Hotspot Indices (NHI)—originally developed for volcanic hotspot detection—has emerged as a powerful, flexible tool for the identification and characterization of high-temperature sources using Sentinel-2 MSI and Landsat-8/9 OLI/OLI-2 observations. By exploiting the combined radiance information from the Near Infrared (NIR) and Short-Wave Infrared (SWIR) spectral bands, the NHI algorithm leverages the multispectral capabilities to identify and characterize hotspots of various origins.

A specific configuration of the NHI algorithm has recently been developed for wildfire mapping. This improved version demonstrated strong performance in complex environments such as California, Hawaii, Canada, Greece, Spain, and Australia, significantly improving the delineation of flame fronts and substantially reducing omission and commission errors.

In this work, we present the results of applying NHI-F to various wildfire events, including the wildfires in Canada in May 2025. Our analysis focuses on two main dimensions essential for modern fire science: (i) the spatial characterization of active flaming fronts and burned-area dynamics at 20–30 m scale and (ii) the quantification of fire intensity through Fire Radiative Power (FRP) and SWIR-based radiance metrics.

In the initial response to wildfires, securing rapid and accurate geographic information is essential. However, helicopter imagery acquired on-site often lacks precise sensor metadata, such as camera pose and internal parameters, making the application of georeferencing difficult. In particular, obliquely captured wildfire imagery presents additional registration challenges due to severe viewpoint changes, scale variations, and low-texture environments. This study proposes an automated georeferencing pipeline capable of operating under these constraints. The proposed method consists of five stages: preprocessing, image retrieval, feature extraction and matching, Exterior Orientation Parameters (EOP) estimation, and orthomosaic generation. An initial Area of Interest (AOI) is defined using inaccurate initial position data, and the Region of Interest (ROI) within the reference map is obtained through a ResNet50-based image retrieval approach. Subsequently, virtual Ground Control Points (GCPs) are generated through deep learning-based feature matching. Elevation data is then assigned using a Digital Elevation Model (DEM), and EOP are estimated via Perspective-n-Point (PnP) and RANSAC algorithms. Intermediate frames are initialized via interpolation and refined through bundle adjustment to produce the final orthomosaic. Experimental results demonstrated that utilizing SuperGlue and LightGlue complementarily increased the number of successfully georeferenced intervals from 5 to 9. Furthermore, a minimum RMSE of 28.30 m was achieved in the most accurate interval. This method proves that by automating the feature-based georeferencing process, practical geographic information can be rapidly provided for initial disaster response, even in sensor-limited environments.

California experiences globally intense wildfire activity with accelerating human casualties and economic losses. Existing research quantifies anthropogenic and climatic contributions to wildland-urban interface (WUI) fires at aggregate levels, yet overlooks heterogeneity arising from differences in ignition locations and dominant spread areas. Using multi-source data from California (2002–2023), we classified WUI fires into four behavioral modes based on ignition site and primary spread zone: I-I (WUI ignition, WUI spread), I-W (WUI ignition, wildland spread), W-I (wildland ignition, WUI spread), and W-W (wildland ignition, wildland spread). We systematically analyzed size characteristics, inter-annual trends, fuel composition, and climate sensitivity across modes. Key findings include: (1) WUI fires accounted for 95.6% of total burned area from large fires, with only 12.2% of burned area within the WUI; both total and mean burned area increased significantly over two decades. (2) Lightning-caused WUI fires showed significantly delayed ignition dates, whereas human-caused fires occurred significantly earlier, with elevated fire frequency observed during Independence Day, Labor Day, and Thanksgiving. (3) I-I fires were predominantly driven by anthropogenic factors with the highest proportion of shrub/grass fuel and the smallest mean size; W-W and I-W fires exhibited significant climate sensitivity, with I-W showing a higher rate of increase than W-W over the study period. These findings reveal differentiated driving mechanisms across WUI fire behavioral types, providing scientific evidence for targeted fire management strategies.