Agricultural machinery operation mode identification is crucial for advancing precision agriculture. This task classifies machinery trajectory points into “field” or “road” categories. Existing research on identifying agricultural machinery operation modes is hindered by insufficient mining of trajectory spatiotemporal features, inadequate propagation of trajectory semantic and graph topology features due to trajectory long-range dependencies, and a lack of capture of characteristic differences between nodes in a heterogeneous trajectory graph. To overcome the above shortcomings, we propose a multifiltered graph convolutional neural network with a self-learning graph topology model (MFGCN-SLT). First, we propose statistical feature enhancement via mathematical and statistical methods to capture the statistical features of trajectory points spatiotemporally. We employ an attention-based feature weighting strategy to adaptively emphasize discriminative spatiotemporal and statistical features of trajectory points. Third, we propose a self-learning adjacency matrix module that automatically constructs the adjacency matrix via an end-to-end graph structure learning paradigma to more effectively capture the graph topological features in the trajectory and the actual trajectory graph structure. Finally, an adaptive homophily graph convolutional network model is proposed, which is divided into two parts: we apply an augmented Laplacian operator that contains multiple adjacency matrices in time and space to better capture long-range semantic information in the trajectory; we incorporate multiple graph filters in graph convolution to capture the variability between different classes of trajectory points and establish a homogeneous graph after convolution to retain multiple types of trajectory information. To verify the effectiveness of the proposed model, experiments were carried out on the actual harvesting trajectories of two crops provided by the Key Laboratory of Agricultural Machinery Monitoring and Big Data Applications Ministry of Agriculture and Rural Affairs. The MFGCN-SLT model achieved accuracies of 98.86
Agricultural machinery trajectory time-series classification aims to infer machinery operation behaviors from GNSS data for operation monitoring, efficiency evaluation, and data-driven farm management. While GNSS-based methods are effective for field-road classification and coarse-grained operational state classification, they remain limited in fine-grained combined-harvester scenarios where different operation categories exhibit highly similar motion patterns in terms of position, speed, heading, and dwell time. This weak separability of trajectory patterns motivates the incorporation of complementary image and audio modalities. However, effectively exploiting image and audio modalities remains challenging for fine-grained agricultural machinery classification. First, the modality that is most informative for a given operation class is difficult to identify and exploit explicitly during classification. Second, long-tailed operation data can bias fixed shared fusion rules toward dominant classes or modalities, reducing the contribution of informative modalities for minority or process-dominated classes. To address these two difficulties, we propose TIM, a trajectory-image-audio multimodal model for eleven-class agricultural machinery trajectory time-series classification. The TIM encodes causal GNSS trajectories, synchronized image frames, and acoustic spectrograms with modality-specific branches to preserve motion, scene, and process cues that are unevenly distributed across modalities and then applies class-adaptive logit-level fusion so that different operation classes can emphasize the modality that is most informative for classification. Experiments on a real-world dataset collected during three months of maize harvesting in Ningxia, Shaanxi, and Inner Mongolia show that TIM-ClassGate achieves the highest overall metrics among the evaluated methods, with 81.36 https://github.com/kakushuu/TIM-Time-Series-Classification.git .
Recent large language model (LLM) frameworks generate daily point-of-interest (POI) activity trajectories based on user profiles and historical context. These frameworks are typically evaluated using Jensen-Shannon divergence (JSD) on population-level mobility features. However, we demonstrate that this protocol incentivizes a problematic shortcut. When the evaluator provides the previous context, the method can achieve a low JSD score simply by copying it. To expose and quantify this bias, we introduce copy-overlap diagnostics, an aggregate copy-aware score, and a teacher-forced copy baseline. Our teacher-forced copy baseline reaches near-minimal JSD purely by copying, while an existing LLM agent already nearly duplicates the provided previous day on about 90
Social media platforms generate large volumes of real-time, geo-referenced content during natural hazards, enabling multimodal topic models to uncover latent themes and their spatial organisation. However, the evaluation of such models typically relies on unimodal metrics, primarily semantic coherence, which overlook spatial structure, temporal dynamics, and the practical usefulness of discovered topics. This paper proposes a multidimensional evaluation framework for disaster-related social media analysis that integrates semantic, spatial, temporal, and operational indicators of topic quality. The framework distinguishes between Utility Information Value (UIV), capturing the informational usefulness of topic representations, and Diagnostic Information Value (DIV), capturing the structural and hazard-consistent validity of spatial and spatio-temporal topic patterns. We examine the framework through a comparative benchmarking analysis of multimodal topic models (MultiGraph and JSTTS) alongside a strong unimodal baseline, using eight geo-referenced datasets from X (Twitter) and Bluesky covering earthquakes, floods, hurricanes, and wildfires under a unified experimental setup. Results show that multimodal models achieve higher actionability (up to 0.75), while the unimodal baseline shows greater semantic diversity (up to 0.99). Spatio-temporal interaction varies across hazards (0.07-0.45), and spatial patterns exhibit hazard-dependent structure. UIV and DIV are strongly correlated (Pearson r = 0.89 ). However, mean performance remains similar across models (UIV ≈ 0.49-0.50, DIV ≈ 0.52-0.53), with no statistically significant differences ( p> 0.05 ), indicating that performance differences are moderate and dataset-dependent rather than consistently model-driven.
In this work, we present the compact data structure k^2 -MS for the representation of raster coverages. k^2 -MS is based on a sequence of binary matrices (each represented by a k^2 -tree), which correspond to the binary encoding of the thematic variable values of the raster. The properties of the k^2 -MS data structure allow it to benefit from the processor instructions PDEP and PEXT, significantly reducing the access time to the structure. Through a series of experiments on different datasets, we evaluated the performance of our structure by comparing it with the k^2 -raster, one of the most competitive structures reported in the literature. On average, when comparing the best configurations of both approaches, k^2 -MS is 48 k^2 -raster.
Underwater noise pollution from shipping activities is widely recognised as a significant threat to marine life. Noise emitted by vessels can have various detrimental effects on fish and marine ecosystems. Accurately estimating and analysing vessel-generated underwater noise is therefore of critical importance for the protection and conservation of marine environments. In this paper, we present an enhanced version of our model for the spatiotemporal characterisation of vessel-generated underwater noise, with a focus on improving its scalability. The original model was limited to fishing vessels and relied on Automatic Identification System (AIS) data to reconstruct trajectories, as well as engine horsepower to estimate emitted noise. Here, we generalise the approach to include all vessel categories — including tankers, cruise ships, and recreational boats — still relying on AIS data, but estimating noise as a function of vessel length overall (LOA) and category, since horsepower information is not available for all vessels in the dataset. We broaden the study area to include the Central Adriatic Sea, in addition to the Northern part previously considered. The enlarged area and the substantially greater volume of AIS data introduce significant computational challenges, making scalability a primary concern. We address these challenges through a comprehensive analysis of optimisation strategies to improve query execution performance. In particular, we restructure the computational pipeline by implementing table partitioning and leveraging parallelisation techniques. Specifically, we employ PostgreSQL’s native parallel query execution and implement multiple partitioning strategies, including range, hash, and list partitioning. We further explore spatial partitioning through space tiling, comparing regular, adaptive, and k-d tree-based grids. Finally, we leverage the Citus extension to distribute computation across four and eight nodes. Our approach improves computational efficiency while preserving the accuracy of noise calculation, offering a scalable solution for large datasets.
Cross-view geo-localization (CVGL) in real-world deployments suffers from orientation uncertainty: ground images are captured at arbitrary headings while aerial references remain North-aligned, causing severe retrieval instability under large azimuthal shifts. Existing approaches either rely on stochastic rotation augmentation, which does not explicitly enforce embedding consistency across viewing angles, or require explicit orientation labels and auxiliary prediction branches that increase model complexity. We propose DIRL (Direction-Inconsistency Robust Localization), a framework that models panoramic yaw rotation through cyclic panoramic shifts and implements rotation-consistent feature learning via multi-view contrastive training. DIRL introduces a symmetric Multi-View InfoNCE objective that jointly optimizes three complementary view pairs, combined with multi-rotation feature aggregation over uniformly sampled rotated variants to encourage consistent embeddings across heading changes, without requiring orientation labels or auxiliary prediction branches. Experiments on CVUSA, CVACT, VIGOR, and CV-Cities demonstrate that DIRL consistently improves rotation robustness: the mean Recall@1 degradation under arbitrary viewpoint rotations decreases from 11.10
High-resolution population data are essential for spatial analysis in Geographic Information Systems. However, census data are typically available only at coarse administrative scales. This paper presents popVAT, a geospatial representation-learning framework for fine-grained population mapping that integrates heterogeneous geospatial data using gridded population products as supervisory reference layers. The architecture combines a Variational Autoencoder for probabilistic spatial representation with pixel-level and atrous-convolution branches capturing local and large-scale spatial context. To address limitations of static multi-scale fusion, the model introduces adaptive gating mechanisms that dynamically regulate interactions between latent representations and contextual features. Using multi-source geospatial covariates, the framework generates 100 m resolution population estimates for Tunisia. Experimental results show strong predictive performance ( R^2= 0.9714; RMSE = 43,488 ), demonstrating stronger agreement with the supervisory reference surface while maintaining higher consistency with independent census aggregates than the evaluated baselines. The results highlight the potential of adaptive spatial representation learning for GIS-based population modeling. Our code and training dataset are available ^3
The rapid proliferation of location-based social media platforms has greatly accelerated the dissemination of geo-tagged information, but it has also facilitated the widespread propagation of localized rumors. Geo-textual Rumor Detection (GRD) has therefore become an important research topic in geoinformatics aimed at automatically identifying deceptive content tied to specific geographical contexts. However, most existing GRD methods rely on learning static patterns from offline datasets, which limits their ability to generalize to emergent local events characterized by rapidly evolving spatial-temporal information distributions. To better understand this limitation, we conduct preliminary analyses of model fitting behaviors during training and identify two critical issues: imbalanced fitting between real and fake classes, and low-rank feature representations caused by the model’s tendency to overfit to homogeneous real patterns. These phenomena directly lead to the severe loss of vital spatial information, which significantly constrains the model’s capacity to capture the diverse spatial and textual patterns inherent in localized rumors. To address these challenges, we propose a novel framework named Decomposing Orthogonal Spatial Subspaces for Emergent Geo-textual rumor detection (Doseg). Our approach decomposes model transformation matrices via singular value decomposition, explicitly separating linguistic semantic, geographical spatial information-aligned, and localized event-specific spatial components while enforcing orthogonality constraints to enhance spatial feature diversity. Extensive experiments on benchmark geo-textual datasets with strict spatial-temporal splits demonstrate that our method substantially improves detection performance and increases the number of dominant principal components in feature representations, leading to stronger generalization for emergent geo-textual rumor scenarios within the geospatial ecosystem.
The growing availability of large-scale transportation data has transformed urban mobility by enabling a wide range of preferences in route choices. One such problem is the maximization of a preference metric (e.g., safety, ease of navigation, or scenic appeal) while adhering to certain upper bound on travel-time. Such route planning problems have diverse use-cases in urban navigation. However, determining routes which maximize a preference metric, while adhering to an upper bound on travel-time, is computationally hard. This problem can be reduced to an instance of the Arc Orienteering Problem (AOP) which is known to be an NP-Hard problem. Although there have been works which attempted to develop heuristics for this problem, they are limited in their solution quality due their myopic search strategy. Moreover, their inherent serial nature limits their ability to harness the benefits of increasingly available parallel processing offered by modern multi-core processors. This shortcoming of the current state of the art severely limits their use in large-scale deployments. To address these limitations, this paper proposes a novel recursive algorithm that explores the solution space more comprehensively while intelligently reusing the intermediate results (across recursive calls) to improve performance. Furthermore, the algorithm’s computational structure is inherently parallelizable, enabling it to fully exploit modern multi-core architectures through existing scheduling frameworks. Extensive experiments on large, real-world road networks demonstrate that this approach yields, on average, a 1.29× improvement over the state of the art in terms solution quality, while maintaining (or better in some case) comparable runtime. Lastly, our algorithm exhibits near-linear speedup as the number of processing cores increases, underscoring both its scalability and practical applicability.
Weighted intervals are ubiquitous, because many objects are associated with temporal and numeric dimensions. As interval datasets are usually large, efficient management and processing of large weighted interval data are required. This article addresses the problem of top-k range search on weighted interval data, which retrieves k intervals with the largest weight among a set of intervals overlapping a given query interval. It finds important analytical applications for vehicles, events, and cryptocurrencies. Existing algorithms for range search on interval data are inefficient for this problem, because they need to search for all intervals that overlap a given query interval. To overcome this inefficiency issue, we first provide a baseline algorithm and then propose three data structures, along with their associated algorithms. Our first proposed algorithm is practically fast but requires O(nlog k) time, where n is the number of intervals, whereas the others require less than O(nlog k) time. Furthermore, we address a duration-constrained variant of the top-k range search problem. To solve this variant efficiently, we extend our algorithms and present how to maintain a time complexity of less than O(nlog k) . We conduct extensive experiments on real-world datasets, and the results show that our algorithms outperform baseline techniques in most cases.
Detecting anomalies in spatiotemporal trajectory data is pivotal for understanding urban mobility and identifying irregular travel behaviors. Existing methods primarily employ reconstruction-based frameworks to encode trajectories into latent representations and then reconstruct the original trajectory. However, they often struggle to capture the complex spatiotemporal dependencies inherent in vehicle movements because they rely heavily on trajectory embeddings. To address this challenge, we introduce CRDiff, a novel trajectory anomaly detection framework via spatial-temporal complementary reconstruction diffusion model. Specifically, we propose a complementary trajectory masking mechanism that partitions trajectory points into observed and unobserved subsets, formulating the problem as a conditional imputation task. By utilizing dual complementary branches, the model comprehensively leverages context from the entire trajectory to capture intricate correlations. Furthermore, we incorporate a Travel Pattern-aware Trajectory Encoder to extract macro-level movement semantics, coupled with a spatiotemporal trajectory denoiser to model fine-grained spatiotemporal dependencies. Finally, anomalies are identified through an ensemble of multi-step diffusion reconstruction errors, significantly enhancing detection robustness. Extensive experiments on two public vehicle trajectory datasets demonstrate that CRDiff outperforms state-of-the-art baselines in detection accuracy.
Unregulated urban sprawl drives environmental degradation, infrastructure stress, and reduced quality of life, making reliable urban expansion modeling (UEM) critical for sustainable planning and evidence-based land-use (LU) governance. However, many cellular automata (CA), machine learning, and shallow deep learning (DL) models rely on fixed spatial windows, simplified neighborhoods, and implicit transition mechanisms, limiting their ability to capture genuine neighborhood effects and transition potential. This study proposes a ResNet+ConvLSTM+Efficient Channel–Position Attention (ECPA) framework that combines deep residual learning for multi-scale spatial feature extraction, ConvLSTM for spatiotemporal dependency modeling, and position-aware attention for adaptive feature recalibration. The framework is evaluated in Colombo and Kandy, Sri Lanka, across two temporal intervals (2005–2015–2025) using two input configurations: eight driving variables and a reduced set of five predictors, enabling sensitivity analysis of transferable transition learning beyond increased input dimensionality. In Colombo, the proposed model achieves 94.75
The analysis of congested areas in an urban environment is important for several applications, including public transportation management, urban planning as well as situational-awareness. Various spatial analysis methods and algorithms have been proposed for this purpose and are integrated in libraries for spatial analytics. One of the most popular methods is hot-spot analysis, where the problem is to identify statistically significant spatial clusters. This problem is studied mainly for spatial and spatio-temporal point data, whereas hot-spots generated by trajectories are equally important as they can directly capture congestion due to mobility patterns. In this paper, we study the problem of hot-spot analysis in road networks. Technically, our approach relies on an adaptation of the Getis-Ord index for hot spot discovery, so that it becomes applicable for graphs (representing the road network). To address scalable processing, we present the design and implementation of Spark-based algorithms for hot spot discovery, including an approximate algorithm that is both efficient and produces highly accurate results. Experiments carried out on real-world public transportation data from the city of Riga demonstrate the effectiveness of our approach.
Trajectory tracking is a method that tracks the current position of a moving object and simplifies its trajectory. It combines two fundamental techniques of moving objects databases, i.e., position tracking and trajectory simplification, in one routine to avoid sending redundant trajectory data to the databases, such that the network, storage, and computing resources are saved. There are some distinct trajectory tracking algorithms, such as LDRH and GRTS. However, they still suffer from issues of effectiveness and efficiency. Moreover, they only track a moving object in discs and cannot satisfy the varied requirements of trajectory tracking in shapes beyond discs. To solve these problems, this paper presents two novel one-pass trajectory tracking algorithms, i.e., CITT and BITT, based on sector intersection and spatio-temporal cone intersection that effectively and efficiently track a moving object in discs and beams, respectively. Using four real-life trajectory datasets, we experimentally show that our CITT and BITT are feasible to track moving objects in discs and beams, respectively. The output sizes of CITT and BITT are comparable with GRTS, and they are 34.7
With advancements in deep learning, convolutional neural networks (CNNs) have been increasingly applied to digital rock super-resolution (SR), achieving significant results. However, compared to general image SR networks, digital rock SR networks are relatively underdeveloped. To enhance the reconstruction performance of digital rock SR networks, we propose an Enhanced Bi-dimensional Large Kernel Hybrid Attention Network (EBHN). We propose Enhanced Bi-dimensional Large Kernel Hybrid Attention Building Block (EBHB), and introduce two novel attention mechanisms: Enhanced Multi-scale Spatial Attention (EMSA) and Enhanced Contrast-aware Blueprint Residual Channel Attention (EBRA). EMSA optimizes attention allocation through a multi-scale spatial squeeze and excitation mechanism, while EBRA enhances channel weight allocation and deep feature extraction via an information supplement branch. Additionally, we incorporate Large Kernel Residual Attention (LKRA) to capture long-range dependencies, expanding the receptive field and improving feature extraction in both spatial and channel dimensions. Experiments on carbonate and sandstone datasets demonstrate that EBHN achieves competitive results with a smaller network size and lower computational complexity. The code and datasets are available GitHub repository and Dataset repository respectively, ensuring reproducibility and facilitating further research in this area.
Accurate weed segmentation and classification are critical for precision agriculture to reduce herbicide usage and enhance crop yield. Traditional approaches like manual scouting and simple image processing are inefficient and unreliable in large-scale farming. Moreover, challenges such as visual similarities between crops and weeds, varying lighting conditions, soil interference, and motion blur hinder accuracy and robustness of automated systems. To address these limitations, this research proposes an innovative deep learning framework for semantic weed segmentation and classification using UAV-captured field images. The proposed model, called Patches Convolution Catkin Sheaf Attention Network (PCCSAN)-based Densenet121 U-Net, combines multi-scale patch learning with hierarchical attention mechanisms for effective feature extraction and segmentation. Sparse Nonnegative Shearlet Feature Decomposition (SNSFD) is employed to enhance feature representation under noisy or blurred conditions. Further, the Willow Catkin Optimization (WCO) algorithm is utilized to fine-tune model training and boost performance. Experiments are conducted on sorghum weed-crop field imagery, and the framework demonstrates high accuracy and generalization capability. Quantitatively, it achieves 98.9
Air quality monitoring is important for environmental management, especially to forecast PM2.5-level pollution concentration. Classical DL models involve heavy computational power, inefficient task offloading, and poor adaptability. Hence, it is ill-fitted for real-time edge application scenarios. To this end, Multi-objective Artificial Afterimage deep Q Self-Attention and Inter-sample Attention Transformer (MAAQ-SAINT) is proposed for PM2.5 prediction on edge devices such as Raspberry Pi 4B and 3B + . The design implements Regression Relief Feature Selection (RRFS) for optimal feature extraction, and Optimal Stopping Theory (OST) for task offloading. The Multi-objective Artificial Afterimage Algorithm (MAAA) is used to enhance prediction accuracy while reducing computational complexity. A self-attention mechanism and inter-sample attention mechanism learn spatial and temporal dependencies by maintaining robust performance. Quantization represents applied optimization for processing in resource-constrained environments. A performance evaluation using MAE, RMSE, and execution time shows that MAAQ-SAINT outperforms traditional techniques in terms of classification (prediction) accuracy (99
Complex Event Recognition (CER) systems have the ability to process streams of events by detecting event patterns with minimal latency. Typically, these patterns have a temporal structure, often resembling the sequential structure of regular expressions. A pattern advances to the next state by checking various conditions on the current and possibly previous events of the stream. CER systems are very efficient in tracking all the possible paths that a pattern may follow and report when a path is complete and a complex event must be reported. In some cases, the conditions that need to be checked may be spatial. For example, in maritime situational awareness, a condition may need to check whether a vessel is close to any other vessel. Such conditions are not easily expressed directly as regular expressions. For such spatio-temporal tasks, there exist dedicated modules which can evaluate this type of conditions efficiently. Thus, we can integrate such a spatio-temporal module within a CER system in order to take advantage of both worlds: the CER engine can accommodate and process complex regular expressions and delegate the evaluation of expensive spatio-temporal tasks to a dedicated module whenever it needs to. We present an approach towards such an integration. We describe how a CER engine, based on symbolic automata, can cooperate with a spatio-temporal link discovery (stLD) module such that the former can leverage the spatio-temporal capabilities of the latter. This cooperation can take place in an online manner rendering the whole system suitable for real-time processing of event streams. We discuss two different communication schemes between the CER engine and the spatio-temporal module and explore when each one should be preferred. We provide a theoretical estimation of the predicted performance of the system under each communication scheme. Our extensive experimental evaluation confirms most of our theoretical predictions.
In the immediate aftermath of a disaster, acquiring information regarding the status of surviving infrastructure is imperative. Among the swiftest sources of such information are first-responders and volunteers engaged in a crowdsensed map generation system. However, data collected through such means often exhibits inconsistencies, omissions, and errors, necessitating the development of robust path combination systems tailored to such exigencies. In this paper, we introduce a novel system termed Proteus , designed to proficiently validate and amalgamate crowd-sourced paths. Utilizing collected GPS trajectories, herein referred to as mobility traces, Proteus adeptly merges these trajectories to generate a cohesive set of connected edges, thereby expeditiously and accurately delineating ground truth irrespective of data acquisition modalities. Leveraging distance thresholds, maximum angle deviation constraints, road-building heuristics, and contextual information extracted from prior maps, Proteus assembles a precise map from mobility traces. Diverging from conventional graph merging techniques, Proteus is characterized by its spatial awareness, preservation of location data integrity, and fault tolerance mechanisms. The system incorporates spatial coverage directives for volunteers to optimize region surveillance efficiency. We demonstrate and assess Proteus through the development of an agent-based model simulating volunteer movements during map generation within an urban environment. Through ANOVA analysis, we ascertain optimal input parameters for the model, validating our system’s efficacy in accurately reconstructing road networks from volunteer-collected GPS trajectories.
