Conference Agenda

Urban rooftop solar planning relies on solar radiation inputs, yet estimates vary across models and measurement methods. This study compares radiation estimates from ArcGIS Solar Analyst, NASA solar radiation values, in-situ observations from research-grade and personal weather stations, and SDT-based Solar Radiation Modelling. We derive hourly global horizontal irradiance (GHI) values from these solar radiation data centres, model building-level estimates, harmonise all sources through temporal alignment, and then evaluate the values. The comparison reveals the hourly modelling of solar radiation models and common solar radiation centres, highlighting where an urban-adjusted local sensor provides lower solar radiation values because of the limited representation of the built and urban environment. Results show that utilising gridded or terrain-based models over urban-adjusted solar radiation values overrepresent due to the uncaptured localised shadings, roof placement effects, and increasing systemic errors for downstream rooftop PV terrain-based assessments. The cross-validated workflow of sensor-based city-scale solar radiation modelling is reproducible and scalable, offering local governments a more nuanced understanding of their solar capacity, and paves the way for carbon emission budget management.

Deep learning, in particular, has driven hundreds of new studies in remote sensing each year. However, ensuring the reliability of these models requires robust uncertainty quantification, an aspect that remains insufficiently explored. Current remote sensing deep learning models typically yield single, deterministic predictions, such as a class label for each pixel or a single biomass value for a given location or region. While commonly used metrics such as RMSE or classification accuracy summarize overall model performance, they fail to convey the reliability of individual predictions, leaving users without guidance on how much confidence to place in each output. Uncertainty estimation addresses this critical gap by quantifying the variability or confidence associated with model predictions. This enables practitioners to interpret not only what the model predicts but also how confident it is in those predictions, providing a more nuanced understanding that is essential for informed decision-making.

We address aleatoric uncertainty using Sentinel-1 and Sentinel-2 time series, proposing two approaches: (i) Gaussian UC, which predicts mean and standard deviation, and (ii) Quantile UC, which estimates the 10th, 50th, and 90th quantiles to capture asymmetric errors. We evaluate these approaches on three representative EO tasks: building height, canopy height, and aboveground biomass estimation. Our results (ID and OOD) show that both models achieve accuracy comparable to deterministic benchmarks while providing well-calibrated, interpretable, and operationally useful confidence estimates. Notably, both models outperform existing global canopy height products on evaluated sites, including the recent 1 m canopy height maps produced by vision transformers.

Recent advances in remote sensing have shifted the focus from the analysis of individual image scenes to the understanding of complex earth systems. This means that the analysis of dynamic evolutions replaces previous static examinations for fixed time points. Furthermore, interdisciplinary research and the integration of heterogeneous data sources are characterizing this transformation process. Digital Elevation Models (DEMs) are predestined for supporting this process by supplementing orthophotos and map data. Promising applications include city planning, landslide analysis, and flood risk assessment where spatio-temporal change detection is a central concept to be applied. Concerning map data, the OpenStreetMap project, based on the idea of Volunteered Geographic Information, has revolutionized the effective production and update of digital maps. However, OSM data does not include elevation information and often contains incorrect geometric information. In this paper, we introduce a self-training framework for validating OSM building footprints with the aid of high-resolution DEMs. The framework supports building segmentation with a self-supervised approach to improve the representation of OSM building footprints. The availability of Digital Elevation Models is used to check the quality of OSM data. The applicability of the approach is demonstrated by a case study conducted in Karlsruhe, Germany. The promising results are described in detail. With our approach, change detection of OSM data can also be carried out using different temporal versions of DEM and OSM data.

Accurate and reliable road extraction from LiDAR data remains a major challenge when spectral cues are limited or spatial heterogeneity increases model uncertainty. This study introduces a comparative, entropy-driven framework for evaluating the performance and reliability of road asset detection using three supervised machine learning algorithms—XGBoost, Random Forest (RF), and Support Vector Machine (SVM). Using a high-density aerial point cloud, a reproducible computational pipeline was implemented, to help practitioners in real-world scenarios for selecting the most robust and reliable machine learning methods for large-scale road assets mapping. Beyond traditional accuracy metrics (Overall Accuracy, F1-score, and Kappa coefficient), uncertainty-based evaluation of the outputs has been conducted using KPIs of entropy and sensitivity to training sets to quantify model reliability and spatial instability. Results reveal that the inclusion of RGB significantly reduces entropy across all models. XGBoost achieved the lowest mean entropy (0.084–0.143) and the most consistent probabilistic behaviour, reflecting confident and well-calibrated model. SVM, while statistically the most accurate (OA and Kappa > 0.97), exhibited higher local entropy (≈ 0.23–0.26), implying precise yet less certain classification. RF demonstrated the highest entropy (≈ 0.65–0.70) and the greatest variability, underscoring its sensitivity to feature noise. Under the WOR configuration, mean entropy rose markedly—most for RF_WOR (≈ 0.93) and moderately for SVM_WOR (≈ 0.39)—while XGBoost retained low uncertainty.

Spatial entropy maps further highlighted that uncertainty concentrates along road edges with RGB data but expands diffusely under WOR conditions, emphasizing the critical role of spectral–spatial synergy in constraining ambiguity. entropy-based evaluation provided insights beyond conventional accuracy metrics, revealing paradoxes between correctness and confidence.

Efficient processing of large-scale 3D point clouds acquired from Terrestrial or Airborne Laser Scanning (TLS/ALS), presents a significant computational challenge. While transformer-based architectures excel at modeling the global context crucial for interpreting these complex scenes, their quadratic computational complexity makes them infeasible for direct application on massive point sets. To address this scalability bottleneck, we propose the Spatio-Sequential Point Transformer (S2PT), a novel hierarchical architecture for efficient and effective large-scale point cloud processing. Our approach begins by serializing the point cloud into an ordered sequence, which enables the use of attention with linear complexity. This not only circumvents the quadratic bottleneck of standard transformers but also establishes a global receptive field at every layer. To compensate for potential information loss during serialization, we further introduce a novel Spatio-sequential Positional Encoding (S2PE) that synergistically combines 3D local geometric features with 1D sequential order information, enhancing the model’s spatial awareness. Experiments on multiple benchmarks demonstrate that S2PT achieves performance comparable to state-of-the-art methods while being significantly more efficient during training and inference, offering a promising path towards scalable representation learning for large-scale 3D scenes.

Accurate semantic segmentation of 3D point clouds is a fundamental task in photogrammetry, robotics, and large-scale scene understanding. Despite recent advances in point-based architectures such as PointNeXt, segmentation performance remains limited near semantic boundaries, where local neighborhoods often contain points from multiple classes, leading to feature ambiguity and oversmoothing.

In this paper, we propose a lightweight boundary-aware learning framework that explicitly models boundary regions during training. Boundary supervision is automatically derived from local semantic label disagreement, eliminating the need for additional annotations. An auxiliary boundary prediction head is introduced to learn boundary-sensitive features, which are subsequently integrated into the segmentation process through a late-stage feature fusion mechanism. In addition, a boundary-aware loss formulation emphasizes boundary regions during optimization, encouraging improved feature discrimination at class transitions.

Experimental results on the S3DIS dataset using the standard 6-fold cross-validation protocol demonstrate consistent improvements over the PointNeXt baseline. The proposed method achieves gains of 3.22% in mean Intersection over Union (mIoU) and 2.85% in mean class accuracy (mACC), with notably improved segmentation quality at object boundaries. Importantly, these improvements are obtained without modifying the backbone architecture or increasing inference complexity.

The results indicate that incorporating boundary-aware supervision provides an effective and efficient strategy for improving segmentation performance in challenging regions.

China is actively pursuing synergistic governance to address air pollution and carbon mitigation issues. This study, focusing on concentration as a key feature, assesses the synergy performance in the Yangtze River Delta Urban Agglomeration (YRDUA), revealing fluctuating trends with only seven cities showing improvement. To further understand the influences from local drivers, we employed an explainable spatial machine learning approach, integrating Geographical Weighted Regression (GWR), Random Forest (RF), and Shapley Additive Explanation (SHAP) to capture nonlinear, threshold, and interaction effects among explanatory variables. The analysis identifies longitude, SO2 emissions from industrial sources, wind speed, latitude, and the proportion of GDP from tertiary sector as the top five influencing factors, emphasizing the importance of geographical position, local air pollution emission, and meteorological condition. Most drivers exhibit nonlinear impacts and interactions with clear thresholds. Such as, wind speeds, exceeding 9.3 m/s negatively impact synergy. Furthermore, spatial heterogeneity of drivers' influence is evident across cities and regions. Specifically, cities along the eastern coast benefit from geographical advantages that enhance synergy in air quality improvement and carbon mitigation. Meteorological conditions, especially wind speed, significantly influence synergy, with notable differences between northern and southern coastal cities. These findings underscore the need for locally tailored governance strategies that leverage each city's unique geographical and socioeconomic attributes to enhance synergistic governance effectiveness. This research contributes to understanding the complex interplay of local drivers influencing synergistic governance in the YRDUA, providing valuable insights for policymakers aiming to improve air quality and promote sustainable development in rapidly urbanizing regions.