The cross-fertilization of the fast-developing AI technology and spatial indexing has given rise to spatial learned indexes. However, these indexes rely on historical data distributions to build models, which limits their ability to anticipate data that has not yet arrived. To address this, we propose a novel Spatio-Temporal Update Method (STUM) that enhances conventional spatial learned indexes by introducing a Spatial Delta Area (SDA) for updates without altering their hierarchical structure. STUM learns spatio-temporal auto-correlation from historical data and integrates predicted future distributions. We apply STUM to the Spatial Learned Block Range INdex (SLBRIN), resulting in the development of the Spatio-Temporal Updatable learned Block Range INdex (STUBRIN), which adopts Revmap to integrate spatio-temporal sequence predictions with the spatial block range. STUBRIN optimizes the retraining process by learning the temporal continuity from spatial distribution and fusing it into the error threshold control mechanism and historical delta learning mechanism. Our results show that STUBRIN achieves 1.9-2.4 & times;, 1.8-13.3 & times;, 3.4-6.7 & times; better build, query and update performance compared to state-of-the-art methods. Additionally, STUBRIN offers superior query and update stability. For concurrent learned indexes, we have also designed parallel scheduling for STUBRIN, which improves build, query and update performance by 6.2-6.8 & times; 0.3-4.2 & times;2.6-5.5 imes$2.6-5.5 & times;, without increasing the index size.

Human mobility serves as an essential proxy for understanding social, economic, and environmental dynamics in urban systems. Geospatial transferability, which measures a model's capability in a new location or unseen region, is a critical dimension for comparing different human mobility generation models. However, few studies have studied the intrinsic characteristics of geospatial transferability. To this end, this study systematically investigates the geospatial transferability of four representative human mobility generation models using a large-scale benchmark dataset of census tract level commuting flows across 2265 counties in the United States. Inspired by the domain adaptation theory in machine learning, we introduce geographic domain shift to describe the intrinsic differences in geographic feature distributions and spatial structures between source and target regions, which may jointly affect model transferability. Moreover, we propose two metrics, mutual information and spatial shift, to quantify the geographic domain shift. To examine their associations with model transferability, we employ linear mixed-effects regression to analyze the associations between geographic domain shifts and transferability. Our results reveal substantial spatial heterogeneity and asymmetry in transfer performance across regions. Both information shift and spatial shift exhibit statistically significant and complementary explanatory power. This indicates that geospatial transferability depends not only on model design but also on intrinsic geographic differences. These findings provide a novel methodological framework for evaluating and improving the geospatial transferability of human mobility generation models and support more robust and fair human mobility data synthesis across diverse regions. It also offers insights on spatial transferability for GeoAI model development.

Spatial learned indexes construct index structures by learning the spatial distribution characteristics of data. Compared to traditional spatial indexes, they significantly reduce storage overhead and query costs. A widely adopted strategy in learned spatial indexes involves employing spatial pivots to model the global distribution of spatial data. Each data point is indexed by its distances to these pivots to enable efficient data partitioning and fast query processing. However, most existing methods rely on overly simplistic assumptions when selecting or modeling pivots. These assumptions limit their effectiveness when dealing with complex or heterogeneous spatial data. To address this, we propose a Distribution-Aware Spatial Pivot Learned Index (DSPI). DSPI leverages a self-attention mechanism to model the spatial distribution of data. It employs a Transformer neural network to jointly optimize pivot placement and pruning efficiency, enabling end-to-end query pruning. Within a given pivot search space, DSPI accurately locates query results using a lightweight function mapping model, optimized via second-order derivatives and Akaike Information Criterion (AIC). Based on DSPI, we further design a range query algorithm with encoding-based pruning and a k-nearest neighbor (kNN) query algorithm with adaptive search radius. Experimental results on real-world and synthetic datasets show that DSPI achieves state-of-the-art efficiency: its index size is about 66.7% of LIMS and 71.4% of TBSI, while average query latency is about 2.5 & times; faster than LIMS and 1.59 & times; faster than TBSI. Under high-dimensional data and large-scale spatial queries, these advantages further widen. Moreover, DSPI maintains stable retrieval performance under dynamic updates.

Urban agglomeration, a product of advanced urbanization and industrial transformation, plays a critical role in national spatial and economic systems. Urban agglomeration resilience remains underexplored in existing literature, particularly regarding how multi-scale interactions between macro-level network structures and micro-level public perceptions jointly shape regional adaptive capacities during crises. To address this gap, this study proposes a novel dual-perspective framework integrating population mobility network analysis (macro) and public sentiment analytics (micro) to evaluate the resilience of eight major Chinese urban agglomerations during the COVID-19 epidemic. The results indicate the agglomerations robust intercity connectivity and stable structural networks, such as polycentricity structures or breaking the inter-provincial effect, exhibit higher resilience. Based on the experimental results, we propose multi-scale development strategies. It is argued that the development of urban agglomerations should focus on breaking down urban barriers, enhancing network hierarchies and complexity, and mitigate long-term urban vulnerabilities of unexpected events.

Recent and upcoming carbon satellites, such as the Orbiting Carbon Observatory-3 (OCO-3) and the Copernicus Anthropogenic Carbon Dioxide Monitoring Mission (CO2M), offer unprecedented opportunities for top-down estimation of urban CO2 emissions. Their observations, i.e., 80 x 80 km2 Snapshot Area Map (SAM) for OCO-3 and 250 km wide swath for CO2M, enable the detection of urban emissions in a single pass. However, accurately identifying urban plumes remains challenging due to their broad spatial extent, low signal-to-noise ratio, and substantial data gaps in quality-filtered XCO2 snapshots. To address these challenges, we propose a Transformer-based deep learning (DL) model for XCO2 interpolation and plume detection. Our approach uses masked pre-training on synthetic CO2M data to learn spatial dependencies and emission-related structures of XCO2 values before fine-tuning for plume detection tasks. Experimental results on synthetic datasets show that the model reconstructs XCO2 with mean absolute errors below the instrumental noise and achieves stable plume detection performance across noise levels. It improves XCO2 gap-filling accuracy especially under regional and swath-missing conditions and significantly outperforms test-and wind-based methods in plume region segmentation accuracy. We further validated the model using 110 SAMs from 39 cities observed by OCO-3, integrating it into a lightweight inversion workflow. The resulting top-down emission estimates show improved consistency with bottom-up inventories compared to baselines (R2 = 0.61, total relative deviation = -0.10), and the city-level aggregation reproduces the bottom-up emission rankings with a Pearson's r of 0.90. These results confirm the transferability and practical utility of our approach across global cities. This study presents a promising approach for reconstructing and detecting urban emission signals from XCO2 snapshots, demonstrating clear potential to support the next-generation carbon monitoring satellites.

For tropical cyclone (TC) forecast, meteorological satellites and observatory sites are deployed all around the world, generating large amounts of multi-source data constantly covering every region of the Earth. Given such considerable data, current TC forecast mainly relies on Numerical Weather Prediction (NWP) combined with the provisional analysis and the unanimous decision of meteorologists. Furthermore, the ”training-before-testing” paradigm prevents the prevalent TC forecast DL models from utilizing the real-time data from multiple sources. It would upgrade the intelligence of the AI-powered TC forecast if the multi-source and multi-modal data could be timely exploited along with the domain knowledge, imitating the human forecast process. Therefore, we propose an Online Interactive and Self-adaptive Network (ISAnet). It is a smart framework that automatically predicts TC intensity with real-time remote sensing multispectral images (MSIs), NWP data and TC structural descriptors computed from NWP or Best Track data via efficient tensor-attention networks while consulting with human experts in scheduled timing. Through this multi-modal interactive learning, it ensures that the model evolves in the correct direction. To predict TC intensity with freshly sampled data without annotations in real-time, it incorporates a self-adaptive module which selects confident samples from massive amounts of MSIs. This mechanism also allows the model to keep learning from new data and improve itself over time. Extensive experiments on long-term datasets (2018–2023) from China’s FY-4A and Japan’s Himawari-8 satellites demonstrate that ISAnet outperforms state-of-the-art methods. Specifically, the 6-hour Maximum Sustained Wind (MSW) prediction errors are reduced to 4.16 m/s and 3.05 m/s, respectively. These results validate the framework’s effectiveness and robust generalization across diverse observation satellites.

Human mobility flow captures the intensity of spatial interactions within a city and offers insights into both the spatial structure of urban systems and human dynamics. However, the acquisition of high-quality human mobility data remains challenging due to high collection costs, limited accessibility, and sparse spatial coverage, motivating the adoption of generative AI to produce representative and scalable human mobility data. In this research, we propose a generative GeoAI-based multimodal flow matching method for large-scale origin-destination (OD) mobility flow generation by fusing remote sensing and social sensing data. This method demonstrates superior performance in both intra-city and cross-city scenarios in the three largest metropolitan areas in the United States, evaluated across metrics regarding mobility flow volume estimation error and graph structure distribution. We further develop an interpretability framework tailored to this model, revealing the role of attention in heterogeneous modality alignment and interpreting the stepwise inference process of the generative model. Through visual and quantitative analysis of the attention scores before and after multimodal alignment, we identify explainable fusion patterns between remote sensing and social sensing data. The deterministic and continuous OD flow generation process defined by flow matching provides transparent evolutionary paths. This study not only accurately models human mobility flows and urban functional coupling between regions but also offers insights on the use of generative AI for understanding urban dynamics.

With the rapid development and adoption of 3D Gaussian Splatting (3DGS), the need for effective copyright protection has become increasingly critical. Existing watermarking techniques for 3DGS mainly focus on protecting rendered images via pre-trained decoders, leaving the underlying 3D Gaussian primitives vulnerable to misuse. In particular, they are ineffective against **Partial Infringement**, where an adversary extracts and reuses only a subset of Gaussians. In this paper, we propose **NGS-Marker**, a novel native watermarking framework for 3DGS. It integrates a jointly trained watermark injector and message decoder, and employs a gradient-based progressive injection strategy to ensure full-scene coverage. This enables robust ownership decoding from any local region. We further extend NGS-Marker with hybrid protection (combining native and indirect watermarks) and support for multimodal watermarking. Extensive experiments demonstrate that NGS-Marker effectively defends against partial infringement while offering practical flexibility for real-world deployment.

In this letter, we explore text-guided image editing in the remote sensing (RS) domain using generative modeling. We propose RSEdit, a collection of models from U-Net to diffusion transformer (DiT) with various configurations. In particular, we present the first comprehensive study of conditioning strategies for building image-editing models from off-the-shelf text-to-image ones. Our experiments show that RSEdit achieves the best instruction-faithful edits while preserving geospatial structure. We release the code at https://github.com/Bili-Sakura/RSEdit-Preview and the checkpoints at https://huggingface.co/collections/BiliSakura/rsedit

Recent advances in 3D Gaussian splatting (3DGS) have demonstrated remarkable progress in high-fidelity scene reconstruction. However, when applied to large-scale scenes from multiview uncrewed aerial vehicle (UAV) imagery with complex structures and diverse categories, existing methods often suffer from geometric instability, floating artifacts, and limited semantic interpretability. To address these challenges, we propose a semantic-geometric collaborative 3DGS reconstruction framework. Specifically, we introduce geometric constraints to enhance structural stability and geometric coherence during global coarse reconstruction. Category-aware semantic constraints are further introduced in blockwise refinement to regulate Gaussian primitives and assign semantic labels. In addition, we develop a semantic uncertainty-guided scheme for adaptively controlling Gaussian densification. Extensive experiments on multiple public large-scale UAV datasets demonstrate that our method achieves competitive rendering quality while significantly improving geometric accuracy, yielding a semantically enriched 3D representation for downstream tasks, such as semantic understanding and urban analysis.

Abstract. Dense greenhouse-gas (GHG) observations, particularly from satellites, require fast atmospheric transport models that link surface emissions to these observations. We introduce GenGHG, a generative machine learning (ML) model that emulates footprints from the Stochastic Time-Inverted Lagrangian Transport (STILT) model. These footprints estimate how a unit of emissions would alter a downwind atmospheric measurement. Unlike existing deterministic ML emulators, which give a single prediction, GenGHG predicts an ensemble of plausible footprints under given meteorological forcing conditions to represent stochastic atmospheric transport. GenGHG retains ML-level computational efficiency, with each member generated in less than 2 s on a single GPU, and we evaluate its generative advantage over deterministic ML across a benchmark of 60 urban areas worldwide. Results show that GenGHG preserves footprint total mass substantially better than deterministic ML, with a mean relative bias of +2.67 % compared with −22.38 %, and also better preserves the footprint-value distribution. Grid-level accuracy also shows that GenGHG better captures spatial footprint patterns, with larger gains in more dispersive transport regimes. We further test GenGHG’s advantage in a synthetic inverse modeling experiment, where transport from GenGHG yields methane emission estimates that closely follow the STILT transport reference and outperform deterministic ML. Finally, we highlight GenGHG’s compatibility with any global meteorology product at a 0.25° resolution. This flexibility, combined with its low computational cost, makes it straightforward to run footprints using multiple meteorology products and subsequently evaluate the possible effects of meteorological uncertainties.

Spatial databases are the main means to manage geo-big data, and learned spatial indices are a novel approach to improve the spatial retrieval performance of spatial databases by modeling the data distribution. However, the complex hierarchical structures in current learning models pose significant limitations, including prolonged construction times, slow data updates, and suboptimal dynamic query performance. Consequently, improving the efficiency of both index construction and updates is essential. We addressed these challenges by introducing a new method, the Spatial Uniform Partition Learned Index (SUPLI). SUPLI utilizes an iterative uniform partitioning algorithm that simplifies data distribution by uniformly segmenting space and applies a linear regression function-instead of a neural network model-to enable efficient index construction. Additionally, SUPLI incorporates query load optimization and historical query learning strategies, which dynamically adjust the spatial query algorithm to enhance query efficiency. Furthermore, a buffer structure is employed to store change information, facilitating efficient updates. Comparative evaluations conducted on three synthetic datasets and two real-world datasets show that SUPLI outperforms the classic R-tree by an order of magnitude in construction, query, and update performance, and demonstrates additional advantages over similar spatial learned indices, such as SPRIG and LISA.

Accurately predicting PM2.5 concentrations remains a major challenge due to the complex and nonlinear nature of its formation and transport processes. Traditional models often struggle to capture the spatiotemporal dynamics of PM2.5, particularly under varying meteorological conditions. In recent years, Koopman operator theory has attracted increasing attention for its ability to transform nonlinear systems into linear representations, thereby enhancing model interpretability and stability. However, most existing Koopman models primarily focus on temporal dynamics and overlook the spatial correlations. To address this limitation, we propose a novel framework called Graph Attention Physics-Constrained Learning (GAPCL). This method combines Graph Attention Networks—designed to model spatial dependencies between PM₂.₅ monitoring stations—with the Physics-Constrained Learning framework grounded in Koopman theory. The model employs an attention mechanism to dynamically weight PM2.5 monitoring stations, uncover spatial relationships, and integrates graph topology with Koopman eigenfunctions to represent the spatiotemporal evolution of complex nonlinear dynamical systems. The effectiveness of GAPCL was validated using hourly data from 2019 to 2021 in the Beijing-Tianjin-Hebei region. The results demonstrated that the model achieved superior prediction accuracy, particularly for short-term forecasts. Compared to the Spatial Physics Constrained Learning (SPCL), GAPCL improved the RMSE, MAE, IA, and r by an average of 12.07

This paper presents the NTIRE 2026 image super-resolution (×4) challenge, one of the associated competitions of the NTIRE 2026 Workshop at CVPR 2026. The challenge aims to reconstruct high-resolution (HR) images from low-resolution (LR) inputs generated through bicubic downsampling with a ×4 scaling factor. The objective is to develop effective super-resolution solutions and analyze recent advances in the field. To reflect the evolving objectives of image super-resolution, the challenge includes two tracks: (1) a restoration track, which emphasizes pixel-wise fidelity and ranks submissions based on PSNR; and (2) a perceptual track, which focuses on visual realism and evaluates results using a perceptual score. A total of 194 participants registered for the challenge, with 31 teams submitting valid entries. This report summarizes the challenge design, datasets, evaluation protocol, main results, and methods of participating teams. The challenge provides a unified benchmark and offers insights into current progress and future directions in image super-resolution.

Sea level rise (SLR) critically impacts coastal safety. Analyzing 70 years of data from 10 stations along the Zhejiang coast, we find that in 2020, sea levels stood 40.3 cm above the national datum, with higher levels in estuaries than eastern islands. While the 70-year average rise rate is moderate at 0.25 cm/yr, it has accelerated significantly in the last two decades.Future projections excluding acceleration estimate rises of 7–18 cm by 2050 and 17–39 cm by 2100; however, if acceleration persists, nonlinear extrapolation suggests increases of 16–29 cm by 2050 and 54–88 cm by 2100, matching high-emission climate scenarios simulated by the coupled global climate models. Northern and estuarine regions exhibit higher SLR rates driven primarily by vertical land movement (VLM). Additionally, significant interdecadal fluctuations, linked to 10–20 year periodic signals and strongly correlated with the Pacific Decadal Oscillation (PDO), show a sign reversal around 2003.

The exponential growth of geographic data reveals limitations in traditional spatial indices. Spatial learned indices that incorporate machine learning models have been proposed to enhance index performance. However, due to the considerable overhead of fine-grained data partitioning and the complexity of hierarchical model structures, existing spatial learned indices still exhibit bottlenecks in index construction and query processing. To address the aforementioned issues, we propose TBSI, an in-memory Transformer-based spatial learned index with an end-to-end structure. TBSI employs an enhanced quadtree to optimize data partitioning and utilizes a Transformer-based position prediction model to manage each data partition, preserving a simple yet effective index structure. TBSI exhibits superior performance in both index construction and query processing. We also design spatial query algorithms based on a filtering-refinement mechanism and data update algorithms based on buffers and flag arrays to support efficient query processing and index maintenance. Extensive experiments on real-world and synthetic datasets demonstrated that, compared to baselines, TBSI achieved up to 23.4 times speedup in build time, up to 24.3 times reduction in index size, up to 5.9 times improvement in range queries, and up to 4.5 times improvement in kNN queries. Also, TBSI exhibited robust adaptability to dynamic data updates.

Timely building damage detection is crucial in disaster decision-making. Unlike existing studies using sub-meter resolution images, we employ stable, mid-high resolution (1-3 meters) remote sensing data to develop building flood damage classification methods for quick disaster response. A mid-high-resolution dataset of 973 Planet-sliced images from three floods in different countries is presented to facilitate building damage classification. Based on this, we propose a Cross Fusion Pyramid Swin Transformer (CFP-SwinT), which balances local and global representations in flood scenes. The model employs feature pyramids to classify multi-scale building groups, and integrates spatial-channel features from dual-time images via a cross-fusion block to jointly capture building groups (pre-disaster) and damage severity (post-disaster). Through extensive experiments, CFP-SwinT outperformed all other backbones and settings and achieved 80.2% accuracy on four-level classification, demonstrating its effective focus on multi-scale disaster information and dual-temporal changes. To enable rapid adaptation to different disaster-affected regions, we combined an unsupervised domain adaptation method with CFP-SwinT, allowing a model trained on one disaster to achieve 77.8% accuracy on unseen flood events. This approach avoids sample labeling, which can benefit decision-making efficiency during disasters. The code for the proposed method is available at https://github.com/WuChuYi/Cross-Fusion-Pyramid-Swin-Transformer.

Multivariate time series forecasting is ubiquitous in the real world. The performance of prediction model is determined by its representation ability. At present, self-supervised pre-training is the main method to improve the representation ability of prediction models. However, the periodic characteristics of time series are rarely considered in the existing pre-training models. Our experimental study shows that the periodic characteristics of time series have a great impact on the performance of self-supervised pre-training models. To address this issue, we propose a novel self-supervised prediction model, SMformer. SMformer has two distinctive features: (1) A new patch partition Module is innovatively introduced into backbone model transformer using the periodic property of time series. (2) Two pretext tasks, shuffle and mask, are design for the self-supervised pre-training of the model SMformer. We conducted extensive experiments on seven benchmark datasets, and the experimental results demonstrate that SMformer significantly outperforms prior comparison baselines.

Understanding the interaction between complex urban environments and human mobility flow patterns underpins adaptive transport systems, resilient communities, and sustainable urban developments, yet inter-regional origin-destination mobility flow information from traditional surveys are costly to update. The satellite imagery offers up-to-date information on urban sensing and opens avenues to examine urban morphology-mobility dynamics. This study develops a deep learning model, Imagery2Flow for predicting fine-grained human mobility flows in urban areas using 10 to 30-meter medium resolution satellite imagery in a timely and low-cost manner. Extensive experiments demonstrate good performance and flexible spatial-temporal generalizability on the top-10 largest metropolitan statistical areas of the United States. Through exploring the spatial heterogeneous effects, we investigate the urban factors (centrality and compactness) influencing human movement flow distributions, enhancing our comprehension of their interactions. The spatial transferability of Imagery2Flow helps reduce regional inequality by informing decisions in data-poor regions, learning from data-rich ones. Interestingly, the typologies of urban sprawl can help explain the cross-city model generalization capability. The temporal transferability proves that human dynamics of cities and the process of urbanization can be well captured from the observed built environment by remote sensing.

Since its successful launch in 2011, the VIIRS (Visible Infrared Imaging Radiometer Suite) on Suomi NPP (Suomi National Polar-orbiting Partnership) has been providing continuous global Earth observations with 22 spectral bands spanning 0.41 to 12.01 μm for over a decade. Originally designed for a 7-year lifespan, Suomi NPP VIIRS has consistently exceeded both its operational specifications and user expectations, successively served as the primary, secondary, and now tertiary satellite imaging radiometer for the NOAA's Joint Polar Satellite System (JPSS) program. Meanwhile, VIIRS on NOAA-20 (launched 2017) and NOAA-21 (launched 2022) have become the secondary and primary imaging radiometers in succession. Building upon lessons learned from SNPP, NOAA-20 VIIRS has distinguished itself as having the most stable reflective solar bands with less than 0.15% responsivity degradation per year in the first 5 years before any correction, and has been endorsed by the World Meteorological Organization/Global Space-based Calibration System (WMO/GSICS) as the on-orbit stability reference. More recently, NOAA-21 VIIRS has become the primary instrument, demonstrating robust performance, especially after the second mid-mission outgassing. This paper evaluates the on-orbit performance of VIIRS instruments, incorporating comprehensive assessments with the calibration methodologies encompassing onboard, vicarious, intercalibration, and recalibration techniques. It focuses on the long-term stability, accuracy, and intersatellite bias assessments using various methods. The root cause for the radiometric biases for the reflective solar bands between Suomi NPP and NOAA-20/-21 is explored, and the recalibration methodology to improve stability and consistency is discussed. Collaborative initiatives with partner satellite programs such as MODIS, the recently launched PACE, EMIT, and the to-be-launched METImage on EPS-SG have also been initiated. These activities underscore the efforts towards achieving consistent, long-term global Earth observations, which are crucial for facilitating time series analyses that support studies on atmosphere, land, ocean dynamics, and artificial nightlight, thereby enhancing global climate change research and operational support for numerical weather predictions.