Conference Agenda

Multi-temporal remote sensing is crucial for crop monitoring, but existing mapping methods struggle with incomplete time series due to data missingness. Current models often assume consistent data, leading to performance degradation when faced with irregular or missing observations. To address this, we propose an enhanced approach combining random temporal masking with neural Ordinary Differential Equation (ODE) optimization, designed to be embedded into existing models.

Our method first employs a random temporal masking strategy during training, forcing the model to learn effective temporal dependencies from sparse, incomplete sequences, thereby boosting its adaptability to diverse missing data scenarios. Second, a time-smoothing regularization term, based on neural ODE, guides the model to learn a continuous, smooth feature trajectory from discrete observations, effectively mitigating temporal inconsistencies and abrupt fluctuations caused by missing data. We also incorporate sine-cosine positional encoding with slight perturbations for precise time representation.

We integrated our approach into the state-of-the-art TSViT model and evaluated it on the PASTIS dataset. Experiments show that while the original TSViT’s accuracy (OA and mIoU) sharply declines with increasing missing frames, our enhanced model maintains significantly better performance. At 80% missing data, our method improves OA by approximately 8% and mIoU by about 12% compared to the baseline. Qualitative results further demonstrate our model’s ability to preserve coherent, smooth spatiotemporal predictions, enhancing robustness and generalization in real-world applications.

Real-world agricultural distributions often suffer from severe class imbalance, typically following a long-tailed distribution. Labeled datasets for crop-type classification are inherently scarce and remain costly to obtain. When working with such limited data, training sets are frequently constructed to be artificially balanced—in particular in the case of few-shot learning—failing to reflect real-world conditions. This mismatch induces a shift between training and test label distributions, degrading real-world generalization. To address this, we propose Dirichlet Prior Augmentation (DirPA), a novel method that simulates an unknown label distribution skew of the target domain proactively during model training. Specifically, we model the real-world distribution as Dirichlet-distributed random variables, effectively performing a prior augmentation during few-shot learning. Our experiments show that DirPA successfully shifts the decision boundary and stabilizes the training process by acting as a dynamic feature regularizer.

Accurate discrimination of cereal crops in heterogeneous agroecosystems requires methods that integrate both spectral and temporal information. This study proposes a compact spectral–temporal framework that combines Optimal Hyperspectral Narrowbands (OHNB) selected from EnMAP imagery using a Spectral Attention Module (SAM) with a Dynamic Time Warping (DTW)-derived phenological distance computed from Sentinel-2 EVI time series. The analysis was conducted in the Saïss region of Morocco, one of the country’s major cereal-producing areas.

SAM identified 29 physiologically meaningful narrowbands spanning the visible, red-edge, near-infrared, and shortwave-infrared regions (429–2438 nm), capturing key pigment, structural, and moisture-related vegetation properties. EVI time series were preprocessed through 10-day median compositing, linear interpolation, and Savitzky–Golay smoothing to generate stable phenological profiles. DTW quantified the temporal similarity of each field’s EVI trajectory to a cereal reference curve, producing a phenology-driven distance feature.

Three classifiers—Random Forest, SVM, and TabPFN—were evaluated under a nested standard and spatial cross-validation strategy. Using only hyperspectral bands, SVM and TabPFN achieved the highest accuracies (ROC-AUC = 0.95–0.93). Incorporating the DTW feature consistently improved performance under spatial CV, especially for RF (ROC-AUC increase: 0.89→0.91), and reduced the performance gap between validation schemes.

Overall, the fusion of SAM-selected hyperspectral bands with DTW-based phenological information enhanced spatial robustness and improved discrimination between cereal and non-cereal fields. The proposed approach offers an efficient and transferable solution for operational crop mapping in semi-arid agricultural landscapes.

Mapping banana crops in heterogeneous tropical landscapes remains challenging due to spectral similarity with surrounding vegetation, fragmented smallholder systems, and complex land-use mosaics. This study applies a deep learning approach, using a U-Net model, on high spatial resolution CBERS-4A imagery to map banana crops in Brazil’s Ribeira Valley, a subtropical region with high rainfall and heterogeneous land cover. Reference data were created through manual interpretation of satellite imagery supported by field knowledge. Representative image tiles were selected and divided into smaller patches for model training, validation, and testing. The U-Net model was trained with standard optimization techniques and evaluated using common semantic segmentation metrics. On the validation set, it achieved strong performance (accuracy 0.91, F1-score 0.84, AUC-ROC 0.96, AUC-PR 0.92). Performance was maintained or improved on the independent test set (accuracy 0.91, F1-score 0.86, AUC-ROC 0.97, AUC-PR 0.93), indicating good generalization. with high agreement between predicted and reference data. Most errors occurred at boundaries between crops and natural vegetation. Additional validation using official agricultural statistics confirmed consistency at the municipal scale. The approach demonstrates that high-resolution imagery combined with deep learning can effectively map banana crops in the region and offers a promising tool for agricultural monitoring and land-use planning in complex environments. The code, trained models, and data are publicly available at https://github.com/hnbendini/banana-unet-mapping.

This study is based on the best crop classification result generated by the proposed unsupervised Machine Learning (ML) method in Li et al., 2025a, using the spectral indices calculated by the formula with spectral bands from Sentinel-2 image products, Normalized Difference Vegetation Index (NDVI), Soil Adjusted Vegetation Index (SAVI), Enhanced Vegetation Index (EVI) and Normalized Difference Moisture Index (NDMI). The patterns and characteristics of these spectral indices, across arable fields with different crop types following the winter growing seasons, have not yet been analyzed in detail. This research aims to provide a comprehensive study of each input spectral index and its impact on the crop classification model. Each spectral index is analyzed across a series of crop fields, using Sentinel-2 images, carefully selected to follow the patterns of winter crop phenology, and the results of unsupervised classification for each crop type in Norfolk, UK are successfully generated and analyzed. The different growing rates between winter barley and wheat have been classified found on a monthly basis using Sentinel-2 RGB images and thus the images during the harvest time, May and June, can support crop classifications. Wild grasses or other plants on the fields led to some crop misclassification from November to March in the Sentinel-2 RGB images. Similarity between winter barley and wheat and the different sowing time among the same type of crop also led to misclassification. In future these misclassifications could be avoided through better understanding of the relation between spectral indices and crop planting cycles.

Current pest monitoring relies on labor-intensive manual scouting, often leading to preventive insecticide use, highlighting the need for automated surveillance. This study presents low-cost RGB camera sensors integrated with edge artificial intelligence (AI) for real-time aphid detection, enabling timely and targeted interventions. Using field images, we trained the YOLO11-n model and evaluated its performance under commercial farming conditions, achieving an average precision of 85 % for apterous aphids. The complex structure of lettuce, with overlapping leaves and shaded areas, limits detection accuracy, particularly for nymphal stages. Nevertheless, these results pave the way for affordable precision agriculture solutions to sustainably improve pest management.

Accurate land cover classification requires both spatial details and temporal information of remote sensing data. While publicly available satellite image time series (SITS) offer short revisit times, they suffer from limited spatial resolution. In contrast, aerial imagery provides fine-grained spatial details, but its temporal coverage is limited. Thus, combining data from those sensors is of interest as their properties are complementary w.r.t. the problem domain. However, the large gap in spatial resolution between these two sensors makes their integration challenging. Generating super-resolution-SITS (SR-SITS) before fusion can help to reduce this gap. In this work, we propose a new approach that integrates diffusion models for generating SR-SITS into a method for the joint pixel-wise classification of aerial and SITS data. Specifically, we employ a diffusion model to generate SR-SITS at an intermediate resolution from the raw SITS and aerial imagery of the same observed area. The SR-SITS are temporally encoded and fused with the aerial features using a cross attention module to produce pixel-wise classification at the geometrical resolution of aerial image. Experimental results on the existing FLAIR benchmark dataset indicate that our approach achieves state-of-the-art results, with a mean Intersection over Union score of 64.0% and an overall accuracy of 76.6%.

We present Sky-NeRF, a novel method for cloud topography estimation based on Dynamic Neural Radiance Fields. Similar to NeRF, we propose to model the 3D structure of clouds as a radiance field, encoded in the parameters of a neural representation. Our goal is to reconstruct the 3D geometry, appearance, and motion of the cloud using a stereo-video of high-resolution top of the atmosphere radiance images. In this paper, we evaluate a novel way of modeling the dynamic behavior of clouds, with the goal of extracting added-value physical information regarding the cloud such as advection speed and direction, velocity field and cloud trajectories. We investigate how to include a simple physical prior, advection, into the learning system and evaluate its impact. Our results show that Sky-NeRF is able to provide a more complete 4D reconstruction than traditional stereo-matching-based algorithms. Moreover, thanks to a physics-based interpolation, Sky-NeRF is able to generate coherent new images from unseen viewing angles, and at any time between the observed frames.

3D topographic change estimation is a fundamental task for understanding Earth surface dynamics in fields of photogrammetry and laser scanning. However, at the current state of research, it is still challenging to accurately separate and quantify various components of topographic surface changes (i.e., rigid spatial movement and non-rigid morphological deformation). In this paper, we conceptualize a surface change tensor to describe 3D surface change based on the displacement field, considering contribution of neighboring points to their center point on the surface. With this concept, we design a new method that is able to quantitatively separate rigid and non-rigid topographic change components from the mixed topographic change. Experiments on synthetic datasets demonstrate that our method is accurate and robust to quantify rigid and non-rigid surface changes, with superiority to the baseline method (M3C2). Additionally, real-world experiments on 3D point clouds collected at four epochs show the effectiveness of the proposed method for monitoring topographic dynamics and identifying geomorphological processes in complex large-scale mountain environments.

Monitoring surface change in dynamic environments is essential to preserve the integrity of human infrastructure and livelihoods from natural hazard consequences. With the advent of 4D remote sensing, near-continuous monitoring of dynamic scenes is unlocked. However, the unordered and irregular nature of point clouds, compounded by temporally variable occlusions and diverse acquisition conditions, hinders the accurate analysis of highly information-rich 4D data. This work addresses the challenge of irregular spatiotemporal sampling in time series of 3D point clouds for the case study of a dynamic sandy beach at different time scales. We explore the use of implicit neural representations (INRs) to model 4D data as continuous spatiotemporal functions that are optimised to estimate the beach topography continuously through space and time. By comparing four model variants and assessing their performance to reconstruct spatially and temporally subsampled data, we evaluate the applicability of INRs to high-frequency topographic monitoring, especially in the context of 4D change analysis. Our results show the ability to reconstruct missing epochs from time series of 3D point clouds with centimetric to decimetric accuracy at time scales ranging from seasonal to daily observations. Our findings highlight the importance of hyperparameter tuning to enable the capture of local details in complex spatiotemporal datasets. Through this, our work lays the foundation for continuous spatiotemporal representation of dynamic scenes, supporting a potentially broad range of change analysis applications.

Spatiotemporal analysis of geospatial time series data has gained increasing attention with the emergence of 4D point clouds and automatic acquisition technologies such as permanent laser scanning (PLS), time-lapse photogrammetry, and uncrewed aerial vehicle (UAV) platforms, enabling near-continuous monitoring of Earth surface dynamics for change detection and process characterization. However, facing massive data volumes through the temporal domain, current topographic data curation practices often rely on empirically determined data processing and management, which may significantly affect reusability, interoperability, and hence processing efficiency due to the absence or heterogeneous nature of metadata. The need for standardized approaches to manage time-dependent metadata has become critical as the demands for sharing data and reproducing analysis across tools and application domains increase. We propose a topographic 4D extension (topo4d) to the SpatioTemporal Asset Catalog (STAC) framework, which provides an open and extensible specification for automatic metadata curation and FAIR data management practices. This paper demonstrates how the topo4d extension facilitates the interoperability and reusability of 4D datasets and presents the corresponding metadata curation workflows applied to two real-world environmental monitoring applications.

This paper proposes the development of a multi-agent system (MAS) for Building Information Modeling (BIM) environments, where users interact with a 3D model and a chat-bot to query, validate, and analyze building elements. By leveraging conversational AI and modular agents capable of semantic understanding and geometric computation, this system allows users to retrieve data, perform quality checks, and visualize computed results directly using the BIM information. The approach supports diverse tasks, from attribute completion and filtering to volumetric calculations, thus enabling a more intelligent and accessible BIM experience for analytical purposes.

The adoption of innovative digital heritage workflows in the Architecture, Engineering, and Construction (AEC) sector faces significant challenges, particularly in integrating digital survey data with Building Information Modeling (BIM) into a unified model. This paper begins with a literature review that outlines the geometric and software-environment constraints complicating such integration and examines various proposed solutions, with particular attention to open-source tools and standard formats.

Building on this foundation, the paper introduces an innovative two-stage method: (1) segmenting, classifying, and enriching digital survey data into a BIM model; and (2) developing a web viewer that hybridizes this BIM model with the original survey data. The proposed workflow relies exclusively on open-source tools and open standards, with Industry Foundation Classes (IFC) used as the native editing format. A seamless continuity is established between the Bonsai add-on for Blender, used as a BIM authoring environment, and the web library That Open Engine, which serves as a dissemination tool enabling interactive querying of BIM data within a web browser. This library shares a common dependency on Three.js with 3DTilesRendererJS, allowing the overlay of a tiled photomesh of the asset. This integration enables the combination of an accurate geometric and visual representation with structured metadata interaction within a unified web environment. Overall, the proposed approach provides a robust and flexible framework for supporting practical applications such as dissemination, documentation, and diagnostic studies of heritage assets.

This study presents a comprehensive and semi-automated framework for generating multi-Level of Detail (LoD) 3D building models using multi-source LiDAR point clouds to support digital twin development. By integrating airborne, drone-based, mobile, and terrestrial LiDAR platforms, the framework addresses limitations of single-source datasets and enables scalable reconstruction across urban and building scales. A robust preprocessing workflow—encompassing subsampling, denoising, colorization, and two-stage registration—significantly enhances point-cloud quality and achieves seamless fusion of heterogeneous datasets with millimetre-level accuracy.

The framework supports outputs ranging from city-scale footprints (LoD0) to detailed parametric building models (LoD4), enabling applications in smart city planning, facility management, and heritage documentation. A knowledge-based segmentation layer further enables the creation of “Smart Point Clouds,” facilitating component-level querying and efficient generation of floor plans, elevations, and façade models. Real-world evaluations in downtown Toronto demonstrate high accuracy and strong computational performance, with LoD0–LoD2 models produced in minutes on a standard workstation.

By ensuring compatibility with CityGML and IFC standards, the framework enhances interoperability within digital twin ecosystems and supports integration with simulation and decision-support systems. While detailed LoD3–LoD4 modeling still requires manual refinement, the workflow establishes a foundation for future automation through AI-driven segmentation and cloud-based parallel processing. Overall, this research advances scalable 3D modeling practices and provides a practical pathway toward comprehensive, data-rich digital twins for smart cities.

With the Czech Republic's impending "BIM Act" driving the creation of a basic built environment model, the study proposes a compliant workflow for incorporating 3D models of two key vegetation feature types from the fundamental geographic vector database: "Forest ground with trees" and "Significant or lonely tree, grove." Modelling relies on nationwide datasets, the digital terrain model, the digital surface model based on image matching of aerial imagery, and supplementary aerial laser scanning data. For the forest features, the process comprised optical point cloud filtration and constrained triangulation, resulting in height-extruded forest base polygons with canopy cover tops. The 3D representation uses MultiSurface geometry, recorded as a PlantCover object in CityGML/3DCityDB, and is in line with the LoD2 standard for buildings. For solitary trees, predefined prototypes were scaled and positioned based on individual tree detection and parameters extracted from point clouds. Features were mapped to the CityGML/3DCityDB SolitaryVegetationObjects class, utilizing Implicit geometry to optimize for data volume and visualization speed. While the digital surface model, which can be easily generated from periodically acquired optical imagery, was sufficient for the forest features, aerial laser scanning data was superior in individual tree modelling. The number of extractable parameters increases with point density and is dependent on the platform used. However, the availability of such higher-density laser scanning data in Europe is limited and varies across countries and regions. The results demonstrate the generation of LoD2 compliant 3D models from nationwide datasets for both vegetation features, visually enriching the basic built environment model.

Despite the rapidly evolving and widely adopted tools in the Architecture, Engineering, Construction, and Operations (AECO) industry, Construction Supply Chain Management (CSCM) remains a fragmented practice with poor integration and interoperability between Building Information Modelling (BIM), Geographical Information Systems (GIS), and logistics systems. This research aims to bridge the gap between BIM, GIS, and logistics information by developing a unified, data-informed Digital Twins (DT) framework necessary to support multi-criteria decision-making (MCDM) in CSCM. They key characteristics of this work include: (1) a repeatable integration for heterogenous BIM-GIS environments powered by IoT networks; (2) a short-horizon predictive module optimized for construction logistics and Just-in-Time (JIT) delivery; and (3) a democratized analytics interface.