The K Nearest Neighbors Query (KNNQ) is a fundamental distance-based query with applications in Geographic Information Systems (GIS) and many other domains. The K Nearest Neighbors Join Query (KNNJQ) combines KNNQ with the join operation to find the K nearest neighbors from a dataset P to every object in another dataset Q. Due to the large number of distance computations involved, KNNJQ is an expensive operation that poses a challenge for in-memory systems. In this paper, we propose and evaluate novel in-memory algorithms for KNNQ and KNNJQ using the compact data structure k^2 -tree. The proposed algorithms exploit hierarchical spatial decomposition together with grouping, pre-processing and progressive pruning techniques. In addition, we develop two new plane-sweep algorithms for KNNQ and KNNJQ, that improve the brute-force algorithms by reducing redundant computations. We conduct an extensive experimental evaluation using real-world spatial datasets to evaluate the performance of our KNNQ and KNNJQ algorithms, comparing them with KD-tree and R*-tree based algorithms. Our results highlight the efficiency of pruning techniques in k^2 -trees, which reduce distance computations and response time by up to an order of magnitude in KNNJQ, in comparison with the equivalent KD-tree, R*-tree and plane-sweep algorithms.
Data-driven prediction and decision-making increasingly depend on large-scale data lakes and machine learning (ML) and deep learning (DL) analytics across a wide range of application domains. Nevertheless, although geospatial vector data constitutes a substantial fraction of contemporary datasets, it has remained primarily managed within traditional GIS and database systems and has benefited only to a limited extent from recent advances in artificial intelligence (AI). This limitation has been largely attributable to the intrinsic characteristics of geospatial information, in particular when represented in vector format, like multidimensional spatial coordinates, scale and projection effects, variable-size geometries, and a broad spectrum of spatial operations, which are difficult to reconcile with the fixed-dimensional inputs typically required by DL models. This paper advanced DL-enabled geospatial analytics by enabling AI models and agents to provide native access, interpretation, and query processing over geospatial vector datasets. Rather than proposing novel architectures, we introduced a methodology for leveraging existing DL models by properly encoding both geospatial inputs and the spatial operations to be estimated. We investigated three representation/architecture families: (i) image-like, georeferenced histograms processed with ResNet and UNet, (ii) point- and graph-based variants of PointNet++, and (iii) fixed-length geometry embeddings obtained by extending Poly2Vec combined with a Transformer model. We instantiated these approaches on three representative geospatial operators, spatial synopsis, spatial clustering, and walkability estimation, and evaluated them using synthetic and real-world datasets. The results demonstrated that all three families could achieve high accuracy in specific configurations with distinct trade-offs: graph-based models were most effective for dense data and short-range interactions, whereas image- and vector-based approaches exhibited superior scalability due to fixed-size inputs, and the image-based approach was particularly effective for multi-dataset operations.
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
