Metallurgical coal operations are a significant but poorly constrained source of methane (CH4) in Canada. We present a multi-instrument analysis of 63 methane plume detections at Fording River Operations, British Columbia (January 2022–March 2026), using the Airborne Visible/Infrared Imaging Spectrometer—Next Generation (AVIRIS-NG; n = 39), the Earth Surface Mineral Dust Source Investigation (EMIT; n = 4) and Tanager-1 (n = 20). Of these, 41 plumes (65%) were quantified, with retrieved emission rates of 34–3622 kg CH4 h−1; 54% exceeded the 500 kg h−1 super-emitter threshold. Because 73% of detections fall in September and no detections are available for 2023, results characterize the late-summer overpass window and should not be extrapolated seasonally without further coverage. The central finding is a systematic, asymmetric wind-speed disagreement between two numerical weather prediction (NWP) products that maps onto the Integrated Mass Enhancement (IME) quantification outcome. A univariate logistic regression identifies HRRR wind speed as a significant predictor of quantification success (OR = 0.29 per m s−1, 95% CI [0.15, 0.56], p < 0.001; AUC = 0.80; 5-fold cross-validated AUC = 0.79 ± 0.20, fold range 0.45–1.00). Cross-validation against ERA5-Land shows that HRRR exceeds ERA5 by a mean of +0.86 m s−1 (+43%) for unquantified events but shows near-zero disagreement for quantified events (–0.09 m s−1, –6%). A sensitivity analysis restricted to HRRR-forced retrievals (EMIT + Tanager-1, n = 24) confirms the finding is not an artefact of mixed wind data sources (OR = 0.28, AUC = 0.83, p = 0.018).

During the deep learning era, innovations in remote sensing (RS) vision models primarily focused on optimizing network architectures for specific tasks and conducting end-to-end training. However, with the advent of the Vision-Language Models (VLMs) era, constrained by substantial computational resource demands, the mainstream research paradigm has shifted towards adapting general-purpose VLMs using RS vertical domain data for fine-tuning, rather than modifying their core architectures. This transition marks a change in the innovation pathway from model-centric to data-centric. Despite the increasing importance of data, the academic community still lacks a systematic, data-centric survey to guide researchers in constructing high-quality RS vision-language datasets for various application scenarios and in conducting effective model fine-tuning and performance evaluation based on these datasets. To address this gap, this paper presents a systematic review of the development status of RS datasets tailored for VLMs. It further provides a comprehensive practical guide for RS researchers in the large model era, covering the entire pipeline from data construction to fine-tuning and evaluation. This paper aims to encourage the academic community to build more diverse and higher-quality vision-language datasets for specialized RS sub-fields. Ultimately, the goal is to converge these efforts into a comprehensive RS vision-language corpus that offers extensive task coverage and a wide geographical distribution, thereby laying a solid data foundation for advancing towards general artificial intelligence in RS.

Accurate accounting of terrestrial ecosystem carbon storage is essential for understanding how production, decay, and disturbance shape ecological function and resilience. However, few China-wide datasets explicitly resolve spatial heterogeneity in carbon density within terrestrial ecosystem types, resulting in systematic and location-dependent uncertainty. This study produced a 1-km-resolution, annually continuous carbon-storage dataset for China’s terrestrial ecosystems from 2001 to 2020 using an Intergovernmental Panel on Climate Change (IPCC) framework, and projected the total terrestrial ecosystem carbon storage in 2030 under five Shared Socioeconomic Pathways (SSP)–Representative Concentration Pathways (RCP) scenarios (SSP1–RCP2.6, SSP2–RCP4.5, SSP3–RCP7.0, SSP4–RCP3.4, and SSP5–RCP8.5). Multi-source remote sensing, land cover, biomass, and soil data were combined to estimate carbon storage in the four carbon pools (aboveground biomass carbon, belowground biomass carbon, soil organic carbon, and dead organic matter carbon) and across six land-use types (cropland, forestland, grassland, water body, built-up land, and unused land). Validation against national-scale estimates and outputs from the Integrated Valuation of Ecosystem Services and Trade-offs (InVEST) model showed reasonable consistency in magnitude and spatial pattern, although systematic positive biases were observed in the InVEST-based estimates. The total terrestrial ecosystem carbon storage increased from 78.42 Pg C in 2001 to 89.62 Pg C in 2011, then declined to 85.49 Pg C in 2020. Soil organic carbon was the dominant pool (53.8% of the total), while forestland and grassland together contributed 69.2% of the total terrestrial ecosystem carbon storage. Under the SSP–RCP scenarios, China’s total terrestrial ecosystem carbon storage in 2030 was projected lowest under SSP1–RCP2.6 at 90.63 Pg C and highest under SSP2–RCP4.5 at 92.62 Pg C. These findings underscore the influence of policy and land-use changes under low- and medium-emission pathways, while extremely high-emission scenarios may show localized gains from ecological restoration. The dataset produced in this study provides a robust basis for carbon accounting, land-use policy, and ecosystem monitoring in China.

Modern image models provide strong cues about what should be segmented in each view, but their masks do not by themselves determine where those labels should persist in 3D. We present Cross-Domain Segmentation via Gaussian Splatting (CDSeg), a label-transfer interface that requires no task-specific 3D segmentation training and uses Gaussian primitives as a renderable label carrier. An external mask source supplies the labels, while renderer-derived visibility determines which 3D primitives receive them. The carrier is instantiated either by completing each input point into one Gaussian, preserving its index, or by reusing the native primitives of an optimized Gaussian scene. CDSeg records pixel–primitive associations during rendering and fuses multi-view masks through voting and a local filter. The resulting labels can be returned to the original points, retained on the native Gaussian scene, or rendered into other views. CDSeg covers promptable, automatic instance, semantic, and LiDAR settings and processes scenes with millions of primitives in seconds. It obtains 92.35% mIoU on DesktopObjects-360, 95.89% on NeRDS-360, and 65.77% on the full ScanNet-v2 validation split using the provided 2D semantic annotations. CDSeg thereby provides one interface for reusing 2D masks across point clouds, Gaussian scenes, and image views without a task-specific 3D segmentation network.

Accurately and timely detecting road pavement damage helps monitor road deterioration extent, thereby guiding maintenance projects and ensuring traffic safety. Nevertheless, due to textural similarity and nested distribution between neighboring pavement damages, as well as the damages with the diversity sizes, irregular shapes, multiple categories, current methods have the limitation in the high-quality detection from road street-level images. To tackle these challenges, this paper develops a novel real-time anchor-free network with a one-stage processing architecture, named RPDNet, for precisely and accurately detecting pavement damages from streel-level road images. First, stacked with a layer-by-layer encoding structure boosted by a deformable fully-attentive module as the backbone extractor, the RPDNet can capture more fine-grained information and generate multiscale strong task-aware semantics, favoring significantly the discrimination noteworthy textural and geometric features. Then, by adopting a multi-level efficient aggregation neck, the RPDNet can promote informative spatial details and integrate the different-level damage encoding features, contributing to the light-weight and optimization of the whole architecture. Afterward, designed with a dual-large kernel module, embedded in a decoupled detection head with anchor-free guidance, the RPDNet can project the ranging dependency of salient and task-oriented pavement damage objects by adaptively aggregating information across large kernels in spatial-domain. Qualitative and quantitative evaluations confirmed that the RPDNet provided a promiseful solution for detecting pavement damages in industrial applications under complex street-level road conditions. Furthermore, comparative analysis with the latest anchor-based and anchor-free alternatives also proved the superiority and generalization of the RPDNet in pavement damage detection tasks. The assessment results displayed that the RPDNet obtained an average mAP@0.5, mAP@0.5:0.95, precision, and recall of 69.16%, 44.86%, 72.59%, and 60.41%, respectively, on two dataset. Additionally, we constructed a large-size multi-city road pavement damage image dataset to support urban road health monitoring.

High-resolution digital elevation models (DEMs) are critical for applications such as environmental monitoring and urban planning, motivating the development of advanced DEM super-resolution (SR) techniques. While recent methods have shown promising results, effectively exploiting high-resolution remote sensing images (HRSIs) to guide DEM SR remains challenging, and progress has been hindered by the lack of large-scale, open-source benchmark datasets. We propose GSRMTL, a novel and parameter-efficient multi-task learning framework for HRSI-guided DEM SR. Given a low-resolution DEM and a paired HRSI, GSRMTL jointly performs DEM SR and semantic segmentation of the optical imagery, where segmentation acts as an auxiliary task to provide semantic priors for elevation reconstruction. To address the dataset bottleneck, we introduce GDEMSR, the first large-scale benchmark dataset specifically designed for HRSI-guided DEM SR. Extensive experiments on GDEMSR and the RGB-guided depth SR benchmark NYU-v2 demonstrate that GSRMTL consistently outperforms state-of-the-art methods while using significantly fewer parameters, highlighting its effectiveness and practical deployment potential.

Simultaneous Localization and Mapping (SLAM is fundamental to embodied agents, enabling them to estimate their position in previously unseen environments while incrementally constructing a spatial representation of the surroundings. The resulting 3D maps can be formulated through implicit representations, such as neural networks that encode geometry in a continuous manner, or through explicit representations, including meshes, voxel grids, and point clouds that provide direct spatial structure. Each approach carries distinct trade-offs in terms of accuracy, scalability, and computational cost. Beyond geometry, recent works have emphasized the importance of semantic understanding. Semantic maps enrich the raw spatial layout with labels, categories, and contextual cues, enabling agents not only to navigate but also to reason about their environment. Such representations support higher-level tasks, including object retrieval, path planning, and interaction with complex scenes, leading to higher-level task and object-based representations. This paper provides a comprehensive review of the literature on 3D representations for SLAM. We discuss methodological advances, identify limitations of current approaches, and highlight open challenges. Finally, we examine how these developments contribute to the broader goal of building robust Embodied AI (E-AI) systems that integrate perception, reasoning, and action.

Building change detection (CD) from bitemporal remote sensing images aims to identify newly constructed, demolished, or modified buildings by comparing observations acquired at different times. A major limitation of existing supervised approaches is their strong dependence on pixelwise building and change annotations, which are costly and labor-intensive to obtain at scale. To address this challenge, we propose SemiBCD, a semi-supervised framework that integrates visual-language model (VLM) priors with consistency learning for building CD under limited supervision. A VLM pretrained on natural images is first employed to generate building pseudo-labels for remote sensing imagery. To mitigate domain shift and suppress noise in these pseudo-labels, we introduce an uncertainty-aware pseudo-label refinement (UAPLR) module that progressively improves pseudo-label reliability during training. Leveraging a small set of annotated change masks together with the refined pseudo-labels, we further design a dual consistency learning strategy, combining perturbation-based multiview consistency and temporal-hierarchical alignment to better exploit unlabeled data. Experiments on the LEVIR-CD and WHU-CD benchmarks, as well as the WHU Building dataset, demonstrate that SemiBCD consistently outperforms representative semi-supervised baselines under low-label regimes. The results indicate that incorporating VLM priors with uncertainty-aware refinement and consistency regularization provides an effective solution for reducing annotation dependence in building CD.

In March 2020, Neural Radiance Field (NeRF) revolutionized Computer Vision, allowing for implicit, neural network-based scene representation and novel view synthesis. NeRF models have found diverse applications in robotics, urban mapping, autonomous navigation, virtual reality/augmented reality, and more. In August 2023, Gaussian Splatting, a direct competitor to the NeRF-based framework, was proposed, gaining tremendous momentum and overtaking NeRF-based research in terms of interest as the dominant framework for novel view synthesis. We present a comprehensive survey of NeRF papers from the past five years (2020-2025). These include papers from the pre-Gaussian Splatting era, where NeRF dominated the field for novel view synthesis and 3D implicit and hybrid representation neural field learning. We also include works from the post-Gaussian Splatting era where NeRF and implicit/hybrid neural fields found more niche applications. Our survey is organized into architecture and application-based taxonomies in the pre-Gaussian Splatting era, as well as a categorization of active research areas for NeRF, neural field, and implicit/hybrid neural representation methods. We provide an introduction to the theory of NeRF and its training via differentiable volume rendering. We also present a benchmark comparison of the performance and speed of classical NeRF, implicit and hybrid neural representation, and neural field models, and an overview of key datasets.

The complex geometries, environmental variability, and inconsistent imaging conditions in shield tunnel linings pose substantial challenges to water leakage detection. Existing models heavily rely on extensive annotated data from diverse environments to ensure reliable performance across varying scenarios, which incurs significant time and labor costs in data annotation. To alleviate the annotation burden, we propose Co-MixPL, a novel semi-supervised learning approach that integrates labeled data with pseudo-labels generated by the Mixed Pseudo Label (MixPL) strategy to iteratively update the teacher-student models. Specifically, Co-MixPL integrates an additional head into the MixPL framework to enhance the encoder’s discriminative capability and introduces a Soft Regression method to mitigate the inherent localization bias in pseudo-labeling, refining the regression loss of pseudo-labels through adaptive reliability scores. Remarkably, experiments on the public “water leakage” dataset, Mendeley Data V1, demonstrate that Co-MixPL approaches state-of-the-art (SOTA) performance using only one-seventh of the training data and outperforms the SOTA by 2.8 AP with merely one-third of the annotations. These findings highlight the effectiveness of Co-MixPL in delivering superior detection performance with significantly reduced annotations, thus better meeting the practical demands of engineering inspection and maintenance. Codes are available at https://github.com/LXJ010/Co-MixPL.

This study examines the spatiotemporal variability of Terrestrial Water Storage (TWS) and Water Availability (WA) across the Middle East (ME) from 2002 to 2024 using exclusively open-access datasets, including GRACE/ GRACE-FO mascon solutions, GLDAS-Noah simulations, CHIRPS precipitation records, and global aridity indices. The contributions of six hydroclimatic variables, such as snow water equivalent, canopy water storage, soil moisture storage, groundwater storage, precipitation, and evapotranspiration, to TWS and WA were quantified through component contribution ratio analysis and Least-Squares Cross Wavelet Analysis (LSCWA). The harmonized and reconstructed datasets provided here are openly accessible, enabling reproducibility and further regional water studies. Results reveal a critical decline in ME water storage, with an average depletion of -45 km3 annually, and widespread WA deficits affecting about half the region. Groundwater storage emerged as the dominant contributor to TWS variability, particularly under arid and hyper-arid conditions, whereas soil moisture and snow water played stronger roles in humid zones. The coherency analysis indicates that annual cycles of TWS and WA were strongly linked with hydroclimatic drivers before 2020 but weakened in subsequent years. These findings, underpinned by openly shared datasets, provide essential resources and insights for water management strategies and sustainable policy development in one of the world's most water-stressed regions.

Cross-platform Light Detection and Ranging (LiDAR) point clouds of tree species classification (TSC) from remains challenging due to substantial variations in point density, geometry, and noise across different LiDAR systems. Existing TSC studies are typically designed for a single data source and a limited number of species, which restricts their generalization to new acquisition platforms and previously unseen species. To address these challenges, we propose a learning framework that explicitly targets geometry only inputs, heterogeneous point densities, and limited labeled data in target domains. Specifically, we develop a compact preprocessing pipeline that augments geometric coordinates with local surface orientation and multi scale density descriptors, together with a sampling and normalization strategy that preserves informative local and global structure after subsampling. Furthermore, a multi-phase transfer learning strategy is introduced and verified to enable efficient adaptation from multi source pretraining to new sensors and species with minimal supervision. Experiments using the FOR-species20K dataset for pretraining and an unseen six-species ULS dataset of demonstrate that the proposed framework achieves an overall accuracy of Image 1001 and a mean F1-score of Image 1002 , exceeding models trained from scratch while converging up to 13-fold faster. These findings highlight the potential of transfer learning to enable rapid and accurate adaptation to new tree species, reducing data collection costs, and offering a scalable solution for cross-platform point cloud analysis for forest monitoring.

Accurate real-time positioning of ground moving targets remains challenging in UAV-based applications, especially for consumer-grade platforms equipped only with a monocular camera. The lack of direct depth information, low-precision inertial measurement units (IMUs), and limited computational resources make reliable target positioning difficult. To address these issues, this paper proposes a real-time positioning method for moving targets using a monocular camera UAV. First, a single-strip bundle adjustment (BA) model with yaw error pre-correction is designed to improve the reliability of Position and Orientation System (POS) estimation and acceleration convergence. Second, a local surface refinement is introduced to better approximate the ground around the targets, thereby improving positioning accuracy. Extensive experiments confirm the feasibility of the proposed method. Compared with conventional BA assisted by a POS, our approach reduces the total Root Mean Square Error (RMSE) by 8.7%. and achieves 3D positioning accuracy ranging from 0.068 m to 0.123 m while maintaining real-time performance. Furthermore, compared with DSM-based ray intersection and deep learning-based depth estimation methods, the proposed framework achieves higher accuracy without requiring external terrain data or computationally expensive models, providing an efficient solution for real-time, high-precision moving target positioning using a consumer-grade monocular UAV.

With the growing demand for high-quality 3D urban scene understanding in applications such as building information modeling (BIM) and digital twins, large-scale and well-annotated 3D datasets have become essential for advancing scientific research and algorithm development. However, existing building facade datasets are predominantly image-based, suffering from drawbacks such as a lack of spatial information and sensitivity to lighting and weather conditions. Moreover, publicly available large-scale labeled datasets of building point clouds still remain scarce and have a relatively small coverage area. To this end, we introduce a city-level building facade point cloud dataset named City-Facade for semantic-level and instance-level segmentation. Firstly, the paper conducts a comprehensive review and analysis of existing urban & building point cloud datasets and point cloud segmentation algorithms. Secondly, we present a large-scale building facade dataset with approximately 200 millions of labeled 3D point clouds (over 60 km roads) belonging to urban scenarios, realized to facilitate the development and evaluation of semantic and instance level algorithms in the urban understanding. Finally, baseline experiments for semantic and instance segmentation are conducted to encourage further research. The proposed dataset is accessible at https://github.com/gorgeouseping/City-Facade, comprising the dataset and segmentation baselines for better comparison and presentation of strengths and weaknesses of different methods. Additionally, the data will undergo continuous improvement and updates based on feedback from the community.

Agentic systems are increasingly integrated with geographic information systems (GIS), where multi-agent coordination enables complex conversational and spatial analysis but introduces security risks. This work presents a security-oriented framework for risk identification, evaluation, and mitigation in a multi-agent GIS system while maintaining adaptability to broader agentic architectures. We test the agentic system of a commercial geospatial partner while developing a modular state-machine-based orchestration framework that abstracts agent behavior into reusable components. We evaluate robustness using a red-teaming framework with an adaptive attacker LLM and a deterministic judge that produces binary outcomes with supporting rationales across multi-turn attacks. We further improve resilience with a prompt optimization framework that treats prompts as structured signatures and injects adversarial demonstrations, enabling systematic security improvements without degrading task performance.

Timely and accurately extracting and assessing pavement cracks is crucial for intelligent transportation systems (ITS) to improve road maintenance and safety. In this paper, we present an automated framework for crack semantic segmentation and quantification using optical images. First, a unique boundary-guided real-time high-resolution network is proposed, termed as BulletNet, for crack semantic segmentation. BulletNet is a bullet-head structure that can retain crack details while ensuring real-time inference speed, in which a Cross-Scale Global Attention (CSGA) module is designed to enhance global feature representation and pixel-level relations, as well as a Boundary-Guided Fusion (BGF) module proposed to utilize boundary features to guide the fusion of crack details and contextual information. Second, a Pixel-level Crack Quantification (PCQ) algorithm is proposed for complex cracks, incorporating an Improved Discrete Skeleton Evolution (IDSE) method to optimize skeleton pruning for accurate crack length and a normal vector correction method to adjust propagation direction for precise crack width. Comprehensive experiments on three datasets showed that the proposed BulletNet surpassed the comparative models in terms of efficiency and performance, with average F1-score, mIoU, and Frames per second (FPS) of 87.20%, 88.70%, and 125.53, respectively. In addition, tested on 200 images, the PCQ calculated the crack maximum widths and lengths with an average relative error of 6.96% and 4.62%, respectively. Finally, BulletNet was deployed on edge devices for field testing, and a system based on the PCQ algorithm was developed to validate the effectiveness of the entire framework.

Background Achieving carbon neutrality requires balancing ecosystem carbon storage and economic development to ensure environmental and social equity. However, few national-scale studies have quantified the spatial relationships between ecosystem carbon storage and economic development, or examined how policy interventions could mitigate these disparities. This study addresses this gap by evaluating eco-economic equity dynamics in China under multiple policy scenarios. Results We integrated multi-source datasets to construct a high-resolution, multi-pool account of China’s terrestrial ecosystem carbon storage, then applied a Carbon-improved GDP (C-GDP) metric to assess equity. Results reveal a severe spatial mismatch: western carbon-rich regions hold 74.6% of national carbon stocks but contribute less than 5% of C-GDP, while economically developed eastern regions dominate C-GDP with lower carbon endowments. The monetized carbon value in C-GDP declined to 2.67% by 2020, and the Dagum Gini coefficient remained above 0.51, indicating persistent regional inequality. Policy simulations show that combining regionally differentiated carbon pricing (302.4 CNY/ton) with a six-tier progressive carbon tax most effectively reduces inequality, lowering the Gini coefficient to 0.49. Conclusions Our findings demonstrate that market mechanisms alone cannot resolve spatial eco-economic inequities. Targeted policy measures are essential to align ecological and economic values, ensuring more equitable and sustainable pathways toward carbon neutrality in China.

High-resolution mapping of land disturbance and reclamation is important for assessing the cumulative environmental effects of oil/ gas production. The growing availability of high-resolution satellite imagery, combined with recent advances in deep learning, offers a desirable solution for detecting land surface changes on disturbed land. In this study, we constructed the Alberta oil/gas wells semantic change detection (SCD) data set in Alberta, Canada, based on high-resolution satellite imagery from WorldView-2 and SPOT-6. The data set consists of 328 pairs of bitemporal images (512 & times; 512 pixels at 1.5-m resolution), along with corresponding semantic change maps, binary change maps, and land cover maps. In addition, we proposed a constrained dual-head convolutional neural network (CNN) framework that jointly learns semantic change and binary change tasks. Specifically, two segmentation heads are designed-one for semantic changes and one for binary changes-and are explicitly connected through a cosine similarity loss that enforces consistency between the two tasks. Taking High-Resolution Net (HRNet)-v2 as the backbone, our model was pretrained on the large-scale SEmantic Change detectiONData Set (SECOND) and fine-tuned on our developed data set. Comparative experiments with BiSRNet, HGINet, and SCanNet demonstrate that our approach achieves superior performance, with the highest mean intersection over union (mIoU) (79.47%) and separated Kappa (SeK) (28.40%) afterfine-tuning. Incorporating land cover maps as additional supervision further boosts results, with our approach reaching an mIoU of 80.05% and a SeK of 29.71%. These findings highlight the effectiveness of the proposed constrained dual-head CNN architecture and the benefit of leveraging land cover information for advancing SCD in remote sensing.

Accurate quantification of forest coverage and combustible biomass (fuel load) is critical for wildfire risk assessment and ecosystem management. However, traditional methods relying on airborne LiDAR or field surveys are cost-prohibitive and time-intensive, while satellite imagery often lacks the vertical resolution required for canopy volume analysis. This paper proposes a novel, automated pipeline for rapid forest inventory using virtual remote sensing data derived from Google Earth Studio (GES). Our approach first generates low-altitude orbital imagery and camera poses for a target region. For dense 3D reconstruction, we employ Pi-Long, developed within the VGGT-Long framework. This model serves as a scalable extension of the Pi-3 feed-forward Transformer architecture. To address the inherent scale ambiguity in monocular reconstruction, we introduce a metric recovery module that aligns the reconstructed trajectory with GES ground truth poses via Sim(3) Umeyama optimization. The metric-scale point cloud is then orthogonally projected into Bird's-Eye-View (BEV) height and density maps. Finally, we employ a watershed-based segmentation algorithm combined with height variance analysis to classify tree species (conifer vs. broadleaf), calculate Leaf Area Index (LAI), and estimate total fuel load. Experimental results demonstrate that this pipeline offers a scalable, cost-effective alternative to physical scanning, enabling near-real-time estimation of forest biomass with high geometric consistency.