Conference Agenda

Real-time depth reconstruction from ultra-high-resolution UAV imagery is essential for time-critical geospatial tasks such as disaster response, yet remains challenging due to wide-baseline parallax, large image sizes, low-texture or specular surfaces, occlusions, and strict computational constraints. Recent zero-shot diffusion models offer fast per-image dense predictions without task-specific retraining, and require fewer labelled datasets than transformer-based predictors while avoiding the rigid capture geometry requirement of classical multi-view stereo. However, their probabilistic inference prevents reliable metric accuracy and temporal consistency across sequential frames and overlapping tiles. We present ZeD-MAP, a cluster-level framework that converts a test-time diffusion depth model into a metrically consistent, SLAM-like mapping pipeline by integrating incremental cluster-based bundle adjustment (BA). Streamed UAV frames are grouped into overlapping clusters; periodic BA produces metrically consistent poses and sparse 3D tie-points, which are reprojected into selected frames and used as metric guidance for diffusion-based depth estimation. Validation on ground-marker flights captured at approximately 50 m altitude (GSD ≈ 0.85 cm/px, ~2,650 m² ground coverage per frame) with the DLR Modular Aerial Camera System (MACS) shows that our method achieves sub-meter accuracy, with approximately 0.87 m error in the horizontal (XY) plane and 0.12 m in the vertical (Z) direction, while maintaining per-image runtimes between 1.47 and 4.91 seconds. Results are subject to minor noise from manual point-cloud annotation. These findings show that BA-based metric guidance provides consistency comparable to classical photogrammetric methods while significantly accelerating processing, enabling real-time 3D map generation.

Satellite-based 3D reconstruction has gained prominence with the advancement of Earth Observation techniques. Recent work on Earth Observation Gaussian Splatting (EOGS) demonstrated the potential of adapting 3D Gaussian Splatting to satellite imagery, enabling rapid Digital Surface Model (DSM) generation from multiple images using Rational Polynomial Coefficients (RPCs) as camera models. However, EOGS suffers from critical inefficiencies: it randomly initializes a large number of Gaussians in volumetric space and relies on opacity-based pruning, resulting in unstable memory footprints and premature loss of fine details—particularly problematic for low-resolution satellite data.

This work presents an improved Gaussian Splatting framework for satellite imagery that addresses these limitations through two key contributions. First, we introduce bundle-adjusted initialization, which leverages geometrically precise points from the bundle adjustment process as initialization seeds rather than random placement. This approach ensures Gaussians are anchored to accurate geometric positions from the outset, significantly improving convergence stability. Second, we propose densification-included optimization, which strategically adds Gaussians in regions requiring detailed reconstruction while maintaining computational efficiency. This selective densification preserves fine-scale features without the memory overhead of EOGS's initial over-allocation strategy.

Our method achieves faster processing times and maintains more consistent memory usage while producing higher-quality DSMs, particularly in challenging low-resolution scenarios. By combining geometric priors from bundle adjustment with adaptive densification, we enable more practical and efficient satellite-based 3D reconstruction suitable for large-scale Earth observation applications.

This contribution presents a comparative evaluation of classical and deep learning–based keypoint detectors for Structure-from-Motion (SfM) reconstruction in challenging Arctic UAV imagery. Snow-covered environments pose difficulties for standard feature matching due to low texture, repetitive patterns, and specular surfaces. While deep keypoint pipelines have shown strong performance on indoor and urban benchmarks, their effectiveness in winter aerial domains remains largely unexplored.

Using multi-view UAV datasets collected across several Alaskan sites, we benchmark three feature-extraction front-ends within a uniform pycolmap-based SfM pipeline: (i) classical SIFT with nearest-neighbor matching; (ii) SuperPoint, a self-supervised convolutional detector–descriptor; and (iii) DISK, a reinforcement-learning–based feature extractor. A simple hybrid approach combining SuperPoint and DISK matches is also tested. All methods share identical geometric verification and bundle-adjustment settings to ensure consistency.

Results show that SIFT remains highly robust on moderately textured Arctic scenes, registering all images and producing the most complete point clouds. SuperPoint and DISK achieve similar reprojection accuracy but struggle with image registration and keypoint coverage on some sequences. Conversely, on extremely low-texture scenes where SIFT fails almost entirely, both deep methods still enable partial reconstructions. Persistent failure cases for all techniques include dense canopy and homogeneous snow.

The study highlights a domain gap between existing deep keypoint models and Arctic aerial imagery, suggesting that domain-specific training and improved spatial keypoint diversity could substantially enhance deep SfM performance in polar regions.

Accurate 3D reconstruction is a key enabler for construction progress monitoring and digital-twin maintenance. However, in tower-crane imagery, persistent dynamic occluders such as hooks and slings violate the static-scene assumption of conventional Structure-from-Motion (SfM), leading to feature mismatches and degraded reconstruction consistency. In this paper, we present a geometry-guided occlusion-handling pipeline for crane-mounted construction-site SfM. Our approach leverages geometric cues from reprojection errors and depth inconsistencies to identify outlier observations, clusters them into spatially coherent prompts, and uses these to guide a foundation segmentation model (SAM2). The resulting per-frame masks are integrated into mask-constrained SfM optimization, ensuring that only static background contributes to reconstruction. Experiments on three real-world crane-mounted sequences (30m, 45m, and 120m) show consistent reductions in mean reprojection error relative to the unmasked baseline. In the most challenging case, the error decreases from 0.962 to 0.872 pixels (9.4%). Compared with a fixed rectangular masking strategy, the proposed masks yield similar reprojection errors while better preserving valid observations and sparse-point completeness. These results indicate that the proposed framework provides a practical geometry-guided strategy for improving internal reconstruction consistency in crane-mounted construction environments.

Video panoptic segmentation (VPS) unifies panoptic segmentation and object tracking by assigning each pixel a semantic class label, or for thing classes, an instance identifier that is consistent across frames. Addressing this task, we propose a novel online VPS method for processing stereoscopic image sequences, which is based on depth-aware kernel-based panoptic segmentation. Specifically, we introduce a geometrical constraint based on predicted bounding boxes into the segmentation of thing instances to overcome the fundamental limitation of kernel-based panoptic segmentation that only appearance information is considered in this step; this regularly leads to panoptic segmentation results in which distinct instances are erroneously merged into one mask. To link detected instances across frames, we propose to extend the commonly employed appearance-based association with a motion-related constraint based on optical flow; this resolves ambiguities in case of instances of similar appearance and, thus, reduces the number of incorrect associations. We experimentally evaluate our method on the publicly available Cityscapes-VPS dataset and compare our results to those of several related methods from the literature. The results demonstrate that our method improves the panoptic quality for a single frame and enhances the instance association across frames, leading to an overall improvement of 3.5% in Video Panoptic Quality on thing classes compared to the employed baseline.

This study addresses the morphological evolution of the Belvedere Glacier (Monte Rosa, Macugnaga – Italy) over the period 2009–2023, using a photogrammetric methodology based on Pleiades (2017) and Pleiades Neo (2023) Very-High Resolution (VHR) satellite imagery, integrated with historical aerial data from 2009. The main objective was to quantify altimetric and volumetric variations of the glacier, assess the intensity of ice mass loss, and analyze the geomorphological effects of the flood event that occurred on August 27, 2023, which generated a major debris flow.

Raster differencing between Digital Elevation Models (DEMs) revealed a significant lowering of the glacier surface. Between 2009 and 2017, the glacier lost approximately 19.3 × 10⁶ m³ of ice (about 2.4 × 10⁶ m³/year), while in the following period (2017–2023) the loss reached 16.9 × 10⁶ m³, with an increased average annual rate of 2.8 × 10⁶ m³/year. These values confirm an acceleration in the ablation process, consistent with other studies (De Gaetani 2021; Ioli 2023) and with the general retreat trend observed in Alpine glaciers due to climate warming.

Tropospheric nitrogen dioxide (NO2) is a crucial trace gas for air quality assessment, yet satellite observations often suffer from spatial gaps (e.g., cloud cover) and temporal limitations. While the geostationary satellite GEMS provides hourly data over Asia, its short historical record and missing data restrict long-term studies. Therefore, a physics-guided adapted normalization network (PhysNorm-Net) is designed to reconstruct a gapless, hourly, and high-resolution (0.05°) tropospheric NO2 dataset over Asia from 2019 to 2024.

The model features an asymmetric U-Net architecture. It handles irregular data gaps using Partial Convolution with a dynamic mask and extracts spatiotemporal representations from meteorological and chemical priors. A novel Physics-Aware Normalization (PhysNorm) module bridges the modality gap by dynamically modulating satellite feature maps using physical backgrounds, ensuring adherence to atmospheric diffusion laws.

Extensive evaluations show that PhysNorm-Net achieves high prediction accuracy (R2 = 0.886). It robustly recovers spatial morphologies and pollution plumes even under extreme missing data scenarios. The generated 2019-2024 dataset accurately captures complex diurnal variations and localized hotspots, providing valuable insights into human activities and pollution policies in Asia.

Accurate estimation of nitrogen oxide (NOx) emissions is essential for understanding their role in atmospheric chemistry and managing air pollution. This study presents a novel approach using Physics-Informed Neural Networks (PINNs) to invert NOx emissions from TROPOspheric Monitoring Instrument (TROPOMI) satellite data. By coupling the physical laws of atmospheric processes, effectively bridging traditional data assimilation techniques with the computational efficiency of deep learning. Unlike purely data-driven models, it directly integrates physical constraints from atmospheric mass continuity equation into the model training process, eliminating the need for inputs or outputs from computationally intensive chemical transport models. Application to the Yangtze River Delta region of China (2018–2023) revealed detailed spatiotemporal NOx emission trends, including the impacts of the COVID-19 pandemic and subsequent recovery. Uncertainty quantification through Monte Carlo dropout provides robust error estimates. This physics-informed approach demonstrates strong potential for efficient NOx emission inversion and offers a versatile foundation for broader quantitative remote sensing applications.

With improvements in the spatial resolution of satellite spectrometers such as TROPOMI, Sentinel-4 and Sentinel-5, more homogeneous cloudy scenes can be resolved at the pixel scale. Therefore, it is worthwhile to use a scattering cloud model in cloud retrieval algorithms. DISAMAR (Determining Instrument Specifications and Analysing Methods for Atmospheric Retrieval) is a computer model developed to simulate the retrieval of atmospheric trace gases, aerosols, clouds, and land-surface properties from passive remote-sensing observations in the 270–2400 nm wavelength range. As a line-by-line radiative transfer model, DISAMAR provides accurate simulations but is computationally expensive. Machine learning techniques can improve the speed of cloud retrieval, because a neural network trained with detailed radiative transfer calculations for scattering clouds can replace the most time-consuming part of the retrieval algorithm.

In this study, we plan to build a cloud retrieval algorithm based on DISAMAR and accelerate it using neural network methods. The algorithm uses TROPOMI observations in the O₂-A band and supports the joint retrieval of cloud optical thickness (COT) and cloud-top pressure (CTP). The neural network models are trained offline using a large, high-resolution spectral data set in the O₂-A band generated by the DISAMAR forward model. All neural networks share the same set of input features but predict different targets, including reflectance and the derivatives of reflectance with respect to cloud pressure and cloud optical thickness. These predictions are then used within an optimal estimation framework to retrieve the cloud parameters.

This study involves remote sensing and artificial intelligence technologies. The study proposed a deep learning-based algorithm to generate the nighttime visible bands for Advanced Himawari Imager geostationary satellite.

Atmospheric aerosols play a vital role in regulating air quality, ecosystems, and climate. Owing to their short atmospheric lifetime, aerosols exhibit strong spatial and temporal variability. Accurate global and regional monitoring of aerosol properties is essential for ecological processes, and radiative forcing. Satellite remote sensing has become a key tool for monitoring aerosol optical thickness (AOT) because of its broad spatial coverage. Traditional physical approaches rely on radiative transfer models (RTMs) to simulate top-of-atmosphere radiances. However, RTMs simplify the real atmosphere, and their accuracy depends strongly on assumed aerosol optical properties and surface reflectance, leading to major uncertainties and inter-algorithm discrepancies. In recent years, data-driven methods have rapidly advanced, driven by developments in machine learning and the increasing availability of collocated satellite and ground-based AOT datasets. The data-driven methods exclusively rely on the data pairs of satellite observations and ground-measured aerosol properties. It learns empirical relationships between satellite observations and measured aerosol properties, and it is more flexible to incorporate more diverse information. However, the AERONET ground stations, commonly used for training, are unevenly distributed and concentrated in urban regions, leaving other surface types such as forests and barren lands underrepresented. Besides, extreme pollution events (e.g., dust storms) are often misclassified as clouds and masked out in AERONET records, introducing bias into training datasets. To mitigate these limitations, this study proposes integrating simulated RTM data into the inversion framework to enhance the robustness and generalization of data-driven AOT retrieval models.

Short annotation如下

Reliable satellite-based estimation of near-surface air pollutants increasingly relies on data-driven models, yet their credibility is hindered by biased generalization assessment and unverified uncertainty estimates. Spatially sparse and unevenly distributed monitoring networks together with strong spatial autocorrelation cause conventional cross-validation approaches to substantially overestimate predictive skill, especially in regions lacking in situ observations. At the same time, although many models produce pixel-level uncertainty estimates, the degree to which these uncertainties reflect true prediction error remains largely unexplored.

This study introduces a controlled, model-agnostic evaluation framework to rigorously examine both spatial generalization and uncertainty reliability in air-quality remote sensing models. A chemical transport model provides a continuous, full-coverage nitrogen dioxide field that serves as an idealized truth. Sampling this field at actual monitoring locations reproduces real observational sparsity while preserving an unbiased reference for domain-wide evaluation. Multiple machine learning models are assessed using sample-based, site-based, and spatially optimized cross-validation to quantify evaluation bias and its dependence on spatial structure.

A dual-path uncertainty strategy is implemented to separately characterize aleatoric and epistemic components, complemented by diagnostic metrics assessing calibration, interval coverage, and sharpness. The framework provides a rigorous pathway for diagnosing reliability in data-driven atmospheric estimation models and supports the development of robust, trustworthy applications in quantitative remote sensing.

This work presents a multi-camera mobile mapping solution designed to deliver accurate and efficient 3D reconstructions across a wide variety of challenging environments, ranging from confined indoor spaces to complex urban outdoor settings. Traditional photogrammetric and terrestrial laser scanning approaches, while capable of high accuracy, often suffer from limitations related to acquisition speed, logistical complexity, and significant post-processing effort—especially in large, occluded, or hard-to-access sites. Mobile Mapping Systems (MMS) based on Visual SLAM (V-SLAM) offer a compelling alternative, thanks to their ability to acquire high-frequency imagery in continuous motion and estimate sensor trajectories in real-time. However, MMS outputs frequently face issues such as reduced geometric accuracy, scale drift in monocular sequences, and the need for extensive optimisation to reach survey-grade results.

To address these limitations, the study extends an existing multi-camera V-SLAM pipeline by tightly integrating monocular estimates with multi-stereo trajectories within the ATOM-ANT3D fisheye multi-camera system. A novel monocular scale-recovery strategy is introduced, based on path-length ratios derived from concurrently recorded stereo tracks. This metrized monocular trajectory is then fused with stereo estimates through a robust pose graph optimisation, followed by a multi-view, feature-based refinement leveraging pre-calibrated camera geometry.

The proposed method is evaluated across four real-world scenarios—spiral tower staircases, dark underground caves, narrow urban corridors, and constrained industrial pipelines. Accuracy is assessed against reference 3D point clouds, while efficiency is compared to a standard multi-view stereo photogrammetric pipeline. Results demonstrate that the integrated approach significantly improves reconstruction consistency, robustness, and end-to-end throughput.

Traditional image matching methods rely heavily on gradient or intensity information. However, the severe nonlinear radiometric distortion (NRD) between infrared and visible images hinders the extraction of repeatable feature points, leading to poor matching performance. To address this, we propose Shape2Match, a novel framework that replaces point features with more consistent, modality-invariant shape features. Specifically, the method utilizes EfficientSAM to extract shape contours and employs elliptic fourier descriptors (EFD) to parameterize and normalize them, creating shape descriptor that is invariant to translation, rotation, and scale. Shape2Match adopts a coarse-to-fine hierarchical strategy: it first performs robust global shape matching using a weighted EFD distance, followed by precise keypoint matching—using Shape Context—within the coarsely aligned shape pairs. We validated Shape2Match on 153 image pairs from 6 datasets, comparing it against methods like SIFT, RIFT, and MS-HLMO. Experimental results demonstrate that Shape2Match achieves a 100\% success rate (SR) across all datasets and significantly outperforms other methods in the number of correct matches (NCM), proving its effectiveness and robustness against NRD, rotation, and scale variations.

Historical film-based images, acquired by aerial sensors since the 1930s and by satellite platforms since the 1960s, provide a unique opportunity to document changes in the Earth surface over the 20th century. Yet, they present significant and specific challenges, including complex distortion in the scanned image pixel grid and poorly known camera exterior and interior orientation. In recent years, semi- or fully-automated approaches, based on photogrammetric and computer vision methods, have emerged, but the performance and limitations of these methods have yet to be directly compared.

The objectives of the Historical Images for Surface Topography Reconstruction Intercomparison eXperiment (Historix) project are to compare existing methods for processing stereoscopic historical images and harmonize processing tools. Here we present the study site and dataset selected for this comparison, the design of the intercomparison and evaluation metrics, as well as preliminary results. Full evaluation will be presented at the conference.

The location accuracy of an image acquired with a space borne camera relies on the knowledge of the orbit of the spacecraft and the orientation of the camera. The a posteriori estimation of a satellite orbit has been a well mastered technique for a long time. Sub-meter accuracy is achievable with a reasonable effort. The geolocation, with a similar accuracy, of the line of sight of an optical instrument flying at 500km or above is a much more challenging task..

On the other hand, the geolocation of a synthetic aperture radar (SAR) image depends only on the orbit of the spacecraft. It is, therefore, easy to acquire space borne SAR images with a sub-metric native geolocation. The Airbus SAR constellation (TerraSAR-X, TanDEM-X and PAZ) provides, on a commercial basis, images with a (better than) 0.2m geolocation accuracy.

The ability to find, through image matching, homologous points in SAR and optical images would transfer the native accuracy of SAR to optical observations, using classical photogrammetric bundle adjustment.

This paper describes an operational way to perform this SAR/Optic images matching and a validation of the absolute location accuracy achieved.

High-precision mapping of planetary surfaces, such as Mars, relies on matched control points derived from existing georeferenced data, as ground control points (GCPs) cannot be obtained through field measurement. However, the handcrafted image matchers like SIFT limit the robustness of this approach, particularly on texture-scarce and self-similar Martian terrain. While deep learning-based matchers offer a new paradigm, their performance gain for bundle adjustment remains inadequately quantified. This paper systematically evaluates four matchers (hand-crafted SIFT and deep learning-based DOG+HardNet+LightGlue, DISK+LightGlue, and LoFTR), assessing their impact on georeferencing tasks using Tianwen-1 high-resolution imagery. Deep learning methods, such as LoFTR, generate more correspondence points with a more uniform spatial distribution, halving the outlier rate and improving bundle adjustment accuracy by 10%. Our study provides a benchmark for planetary mapping and shows that powerful, learning-based image matchers are pivotal for next-generation automated mapping systems.

This work investigates the solutions developed at Natural Resources Canada to produce analysis-ready mapping products from Canada's national air photo library including two main workflows: 1) The photogrammetric processing of historical photos with an emphasis on the more challenging automated georeferencing component; 2) Enhancing interpretability through generative artificial intelligence models for super-resolution and deep colorization.

The accuracy of the aerial triangulation (AT) performed in the map projection for a GNSS-INS-supported image block consisting of 4342 vertical images, GSD 20 cm, with 22 main strips and 5 cross strips is investigated. Using 169 check points the obtained results are compared with the accuracy achieved by running the AT in an undistorted tangential system. It turns out, that in both systems the same accuracies can be achieved, with RMSE in (X, Y, Z) of (7, 10, 11) cm, if Earth curvature and scale distortion are correctly modelled in the map projection. If the scale distortion is not considered, then the RMSE in Z increases by 100% to 300% (depending on the height distribution of the GCPs). In AT software packages, that do not consider the scale distortion, a partial compensation is possible by either adapting the height of the projection centres or the principal distance leading to RMSE of around (10, 11, 15) cm.