Time series forecasting plays a pivotal role in numerous critical real-world domains, where accurately predicting future trends guides essential decision-making processes. Although recent advancements in deep learning, particularly Transformer-based architectures, have significantly enhanced forecasting capabilities, challenges persist in computational efficiency and long-term prediction accuracy. Additionally, existing patch embedding methods often fail to capture comprehensive temporal dependencies. To address these limitations, we propose the Global Context Mamba Network (GCMNet), a novel architecture designed to extract the global temporal information by using a Global Context Patch Embedding module and Mamba’s efficient sequence modeling capabilities. Our model incorporates Variate and Temporal Mamba layers to effectively capture inter-variable correlations and temporal dynamics while maintaining high computational efficiency. Experimental results on eight real-world datasets demonstrated that GCMNet outperforms eight state-of-the-art methods in long-term time series forecasting tasks.

ABSTRACT Establishing scientifically grounded soil erosion control targets and feasible vegetation restoration strategies is crucial for high quality development of the Chinese Loess Plateau which is severely affected by soil erosion. Currently, the commonly used tolerable soil loss (TSL) is inadequate for addressing the spatial heterogeneity in topography and temporal variability in climatic conditions attributed to its static characteristics. This study developed a dynamic assessment framework using the Revised Universal Soil Loss Equation (RUSLE) model to construct a monthly soil erosion rate raster dataset at a 500 m resolution for 2001–2024. Model validation was conducted through comparison with sediment yield data from 25 tributary hydrological stations. By analyzing the soil erosion rate and the prevailing climatic and land management factors, we established monthly erosion control targets (ECT) and the corresponding vegetation cover targets (VCT). Results showed that (1) the average erosion rate declined from 15.5 to 10.17 t ha −1 year −1 , with a mean of 10.53 t ha −1 year −1 , and peaks occurred in summer, especially July; (2) the estimated ECT values ranged from 13.4 to 8.9 t ha −1 year −1 , with a mean of 7.2 t ha −1 year −1 , peaking at 1.6 t ha −1 month −1 in July; (3) the estimated VCT ranged from 43.4% to 66.6%, averaging 59.8%. Overall, 51.6% of the Loess Plateau has already met the VCT requirements necessary to sustain the ECT. These findings provide scientific guidance for soil and water conservation and vegetation restoration tailored to local conditions, with potential applicability to other erosion‐prone regions.

Accurately extracting watershed boundaries is critical for hydrological modeling and environmental management. Traditional extraction methods from Digital Elevation Models (DEMs) rely on manually defined thresholds and supplementary terrain features, limiting adaptability and efficiency. To address these issues, this study developed a watershed boundaries extraction framework based on a Residual Bottleneck Attention Multi-feature Fusion Network (RBM-SegNet). The framework consists of three components: an input layer, a semantic segmentation model, and a post-processing module. Key contributions include: (1) utilizing the [DEM, Slope, Hillshade, and Aspect] functions as the optimal input combination; (2) introducing residual connections and the Bottleneck Attention Module (BAM) to enhance feature transmission and suppress irrelevant regions; (3) incorporating multi-feature fusion to refine structural and detail prediction; and (4) incorporating post-processing to improve output-completeness and hydrological consistency. The experimental results show that RBM-SegNet outperforms traditional and existing deep learning methods in accuracy, demonstrating strong potential for practical applications.

Slope length and steepness (LS) factors derived from Digital Elevation Models (DEMs) are critical for soil erosion modeling. However, acquiring large-scale, high-resolution DEMs remains challenging. Existing downscaling methods often fail to integrate global and local terrain features or recover high-frequency details in regions with large slope variations. This study proposes a terrain feature-aware downscaling model based on MambaIR (TfaDM_MambaIR), designed with large kernel attention module (LKA_Block) to capture global terrain structures and a texture detail feature extraction module (Res_MambaIR) to model long-range spatial dependencies in elevations and enhance fine-grained terrain relief, effectively reconstructing continuous terrain structures and complex details. A collaborative loss integrates explicit constraints on slope steepness and terrain structure lines to optimize global elevation precision while preserving complex topographic features. Experiments show that TfaDM_MambaIR reduces DEM/LS errors by 21%-80%/1%-41% and improves DEM/LS PSNR by 3%-25%/4%-15%, providing a feasible model for improving DEM and LS reliability at large scales.

With the increasing size of datasets and the expansion of Deep Neural Networks (DNNs), the training process has become exceedingly time-consuming. Distributed training, specifically the Pipeline Model Parallelism (PMP) method, commonly mitigates this problem but suffers from bubble time delays. This paper proposes LIBPipe, a pipeline training framework that explicitly incorporates a load-imbalance method to reduce bubble time in PMP. Within LIBPipe, a model-unequal-partitioning method is designed from the perspective of load imbalance to reshape the pipeline execution pattern and significantly shorten idle periods during training. On top of this method, a performance-guided unequal-partitioning search algorithm is developed to efficiently identify near-optimal partitioning strategies under memory constraints. The paper theoretically proves the time efficiency of adopting load imbalance in pipeline models. Comprehensive experiments are conducted on an 8-GPU server to evaluate the efficiency of LIBPipe, using the IMDB and mini-ImageNet datasets as well as well-known models such as BERT and ResNet. The BERT-series models achieve a maximum throughput improvement of 60.3%, while the ResNet-series models achieve a maximum improvement of 74.1%.

Fragmented farmland and persistent cloud–snow interference in the high-altitude cold regions of the Qinghai–Tibet Plateau, coupled with unstable crop phenology, pose significant challenges for accurate cereal and oil crop identification using single-date imagery or low-dimensional features. This study focused on the agricultural areas of the Shigatse River Valley in the Qinghai–Tibet Plateau. Leveraging the Google Earth Engine (GEE) cloud computing platform, we integrated Sentinel-2 remote sensing data with field survey sampling data to extract the planting structures, distribution patterns, and cultivated areas of cereal and oil crops. Three machine-learning classifiers—Random Forest (RF), Support Vector Machine (SVM), and Gradient Boosted Trees (GBT)—were evaluated to investigate the influence of different feature sets and classifier combinations on mapping accuracy. The results indicated that when all feature bands were utilized, the RF classifier achieved the highest performance, with an overall accuracy of 84.77% and a kappa coefficient of 0.64, outperforming both the SVM and GBT models. The incorporation of phenological and topographic features further enhanced classification accuracy, providing a robust framework for identifying cereal and oil crops in high-altitude environments. Based on the optimal model estimation, the cultivated areas in 2021 were 581.52 km2 for highland barley, 295.39 km2 for wheat, and 386.81 km2 for rapeseed. Their spatial patterns closely aligned with the valley-terrace topography and local irrigation conditions. These findings offer novel insights and a reliable methodology for the rapid extraction of crop spatial information in regions with complex planting structures.

High-fidelity environmental mapping is a prerequisite for autonomous robots performing feed pushing and alley cleaning in modern cattle barns. To facilitate large-scale deployment, balancing sensing performance with hardware cost is critical, making low-beam (16-channel) LiDAR a more practical yet challenging sensing solution. Existing mapping frameworks struggle to accurately reconstruct thin-walled metal fences due to extreme point cloud sparsity and motion-induced "artificial thickening", while deep learning-based completion models incur prohibitive computational overhead for real-time execution on embedded agricultural platforms. To address these critical limitations, this study proposes Voxel-CattleMap, a lightweight structural enhancement framework designed to improve map completeness and consistency without upgrading hardware. The framework innovatively integrates two core modules: a distance-weighted spatio-temporal fusion strategy that accumulates geometric evidence from consecutive frames to fill sensing gaps, and an EVD-based voxel refinement module that leverages local geometric priors to restore the sharpness and integrity of structural boundaries. Extensive experiments conducted on a Handsfree robotic platform demonstrate that Voxel-CattleMap significantly outperforms state-of-the-art baseline methods, including A-LOAM, DLO, ROLO, and AG-LOAM. Specifically, the proposed method stabilizes the surface thickness within the range of 0.036–0.041 m, closely aligning with physical ground truth, and achieves more than double the structural continuity of baseline approaches. Notably, under rapid motion (0.42 m/s), Voxel-CattleMap nearly doubles the completion ratio of the agriculture-oriented AG-LOAM while maintaining a real-time processing rate of 9.94 FPS. These findings confirm that Voxel-CattleMap provides a robust and cost-effective perception foundation for the practical deployment of robots in unstructured dairy facilities.

Beef cattle behavior provides valuable information regarding their health status. Recently, deep convolutional network-based methods have achieved considerable results in beef cattle behavior recognition. However, their robustness under low-light or dark conditions remains limited, which restricts their application in real farm environments. To address this issue, this study constructed a realistic beef cattle behavior dataset in the dark, named Dark Beef Cattle Actions, which was collected under real nighttime farm conditions. The constructed dataset contains 1097 video clips collected from 30 beef cattle and covers 6 behavioral classes, including running, feeding, drinking, grooming, mounting, and fighting. Based on this dataset, we proposed a novel neural network architecture based on spatio-temporal dark enhancement and dynamic fusion for beef cattle behavior recognition in the dark. First, a spatio-temporal dark enhancement module was designed to improve dark video quality while preserving motion features. Second, a dynamic fusion module was introduced to adaptively fuse features from different branches and obtain more discriminative representations. In addition, a joint loss was adopted to optimize both dark enhancement and action recognition. Experimental results on the constructed dataset show that the proposed method achieved a weighted-averaged precision score of 88.47%, a weighted-averaged recall score of 80.18%, an accuracy score of 83.80%, and a weighted-averaged F1-score of 84.12%. Compared with other state-of-the-art methods, the proposed method achieved competitive performance in the recognition of night-time beef cattle behavior. These findings would provide support for intelligent livestock behavior recognition and monitoring in precision farming.

With the rapid development of CNN or Transformer, the present mainstream approaches regard an image patch as the reference of the target to perform tracking, which is known as template matching-based trackers. However, most existing template matching-based trackers only consider the per-frame localization accuracy, neglecting the potential distractor (similar object) dependencies among multiple video frames, which poses a fundamental challenge in template matching-based tracking. In this work, we propose a novel comprehensive framework with multi-frame distractor suppression for visual object tracking (MFDSTrack), which explicitly models the temporal history of both the target object and potential distractors. Specifically, we utilize a universal target candidate generation module to detect target candidates (both target and distractors), providing a holistic view of the scene. In addition, a temporal and distractor-aware association module is designed to suppress multi-frame distractors by adopting a simple encoder-decoder Transformer architecture. The encoder accepts inputs of target candidates' history, while the decoder takes current target candidate queries and the output of the encoder as inputs to associate current target candidate queries with historical trajectories. We extensively evaluate our trackers, MFDSTrack-SD, MFDSTrack-OS, MFDSTrack-GRM, and MFDSTrack-LT on the LaSOT, LaSOText, TrackingNet, GOT-10k, UAV123, NFS, and OTB100 benchmark. Extensive experiments show that our methods outperform previous state-of-the-art trackers on seven tracking benchmarks.

Change detection in synthetic aperture radar (SAR) images is a hot yet highly challenging task in remote sensing. Existing unsupervised SAR change detection methods often struggle with inherent speckle noise and insufficiently utilize pseudo-labels, particularly neglecting uncertain areas. In this paper, we propose a multilevel difference-enhanced denoise dual-branch network (MDDNet), comprising representation learning and change detection branches. First, fuzzy c-means clustering is employed to generate pseudo-labels, categorizing the image areas as changed, nochanged, and uncertain. Second, we design a denoise representation loss function in the representation learning branch to maximize the use of pseudo-labels, while mitigating speckle noise. Furthermore, a multilevel difference computation module is proposed to focus on changes in ground objects and capture more comprehensive change information. Experimental results on three public SAR datasets show that the proposed method outperforms six state-of-the-art methods, achieving the best performance with an average overall accuracy of 98.86% and an average Kappa coefficient of 89.36%.

Addressing the challenges associated with manual measurement of body sizes for beef cattle, the potential for inducing stress responses in animals, and the inefficiencies inherent in such labor-intensive tasks, a novel threedimensional (3D) point-based deep learning (DL) network named CattlePartNet is proposed. This innovative network is redesigned to segment point cloud data (PCD) of cattle into crucial body regions, thereby facilitating and optimizing the measurement process. The newly proposed CattlePartNet adopts the base network of PointNet++ as its backbone network, where the parallelized patch-aware attention module and depth-wise separable convolutions are freshly incorporated to lower the risk of overfitting. Additionally, it incorporates the Sophia optimizer, a novel and highly efficient optimization algorithm, instead of the Adam optimizer in PointNet++. Impressively, CattlePartNet outperforms the PointNet++ backbone network with a 2.4 % improvement in mean Intersection over Union (mIoU), as demonstrated on the ShapeNetPart dataset, which is widely regarded as an established benchmark for part segmentation tasks within the domain of PCD analysis. Leveraging the established network, CattlePartNet is trained on a dataset of cattle PCD for automated segmentation of key body regions, achieving an impressive mIoU of 91.7 %. The body sizes of cattle are categorized into two types: linear and curvilinear. For linear body sizes: body height (BH), body length (BL), and hip height (HH), measurement points are identified and extracted using sophisticated contour extraction techniques such as Alpha Shapes, mean curvature, and Gaussian curvature. Curvilinear body sizes: chest girth (CG) and abdominal circumference (AC) are measured through slice interception and cubic B-spline curve methods. The mean relative errors for the 5 body sizes-BH, BL, CG, AC, and HH-are reported as 4.96 %, 5.47 %, 6.04 %, 5.68 %, and 5.49 %, respectively. In comparison to traditional measurement of beef cattle, CattlePartNet precisely extracts key regions of body size from PCD, demonstrating robustness in generalizing segmentation across diverse cattle breeds and showing potential for segmenting PCD of other large livestock species. Furthermore, the measurement algorithm of 5 body sizes accurately localizes key measurement points within each region, providing essential support for breeding applications such as health assessment and production performance measurement.

Gradient Inversion attacks pose a significant threat to Federated Learning (FL) by recovering users' private data from shared gradients. Recent studies have demonstrated the feasibility of reconstructing private data in image classification tasks within FL. However, these methods usually reconstruct images with limited fidelity or rely on unrealistic assumptions. In this paper, we propose a high-fidelity Gradient Inversion method using Diffusion Model (GIDM), which incorporates latent diffusion model into gradient inversion attack. Rather than using the diffusion model's image generation capability as a black box, we disassemble the model and treat each sampling step as a fundamental operating unit, optimizing the intermediate output. Additionally, we design a feature regularization to serve as an auxiliary part of the optimization objective. By reconstructing the high-dimensional features of truth images, GIDM uncovers additional private information, thereby enhancing the fidelity of the reconstructed images. Extensive experiments demonstrate that GIDM achieves pixel-level image reconstruction with superior fidelity and visual quality compared to existing GAN-based methods. Furthermore, GIDM proves effective against defensive strategies and large data batches, underscoring the need for more advanced privacy protection mechanisms in FL.

The accurate extraction of valley bottoms from digital elevation models (DEMs) is crucial for hydrological and geomorphological analyses of mountainous landscapes. However, threshold settings rely on manual intervention; roads near valley bottoms resemble valley-bottom features, and thresholds cannot effectively adapt to valleys of various shapes, leading to low extraction accuracy in existing methods, particularly in narrow V-shaped valleys. To address these issues, this study developed a semantic segmentation approach called a Dense-based Attention Merging Context Guided Network (DAM-CGNet). Without relying on thresholds, this method effectively excludes roads on hillslopes and enhances the recognition of steep feature changes at valley boundaries, enabling the extraction of valley bottoms of various shapes. Key improvements include: (1) incorporating the convolutional block attention module (CBAM) to enhance feature reuse in the information flow, employing attention mechanisms to suppress irrelevant feature responses and focus on valley boundary features; (2) using the dense connection strategy of DenseNet to rebuild the feature flow, helping the model keep important valley-bottom details in deep layers and better recognize small and narrow valleys; and (3) addressing the limitations of single-channel DEM representation by evaluating various input combinations, ultimately selecting DEM, topographic position index (TPI), and slope as effective inputs for valley-bottom extraction. Experiments using semantic segmentation models and conventional methods validated the effectiveness of the proposed method. Specifically, DAM-CGNet achieved high accuracy on the test set (MPA: 90.15 %, MIoU: 84.18 %, FWIoU: 92.99 %) and outperformed conventional methods in extracting valley bottoms of various shapes. This method, without a manual threshold setting as in conventional approaches, enhances valley bottom extraction precision and provides a new perspective for subsequent valley bottom width calculations.

Real-time and accurate beef cattle detection is essential for effective livestock management. Traditional manual observation methods are labor-intensive and inefficient. Recent studies have shown that deep learning has significantly improved beef cattle detection accuracy. However, achieving robust beef cattle detection remains challenging due to single farming scenarios, occlusions, and dense cattle groups. As an effective solution, this paper proposes a novel method for efficient feature selection and fusion for real-time beef cattle detection (EFSF-RBCD). Specifically, we begin by developing a feature extraction network based on multipath cooperative and poly kernel inception (MPCPKI), which is designed to optimize the feature extraction capabilities. The network includes an efficient P4 feature-layer selection module based on the multipath cooperative gating mechanism (EP4MCGM), which integrates low-level features from shallow layers and enhances fine detail detection. Additionally, the P5 feature layer selection module, based on the cross-stage partial poly kernel inception network (CSPPKINetP5), enables efficient target feature extraction while reducing the computational load. Furthermore, we propose a frequency-domain context feature fusion network (FDCFN), a novel framework that integrates the frequency-domain branch (FDB) and context feature fusion branch (CFFB) to capture local and global contextual information better. Additionally, to enhance detection accuracy, a novel bounding box regression loss function, SIoU, was introduced, which improves bounding box position and size estimation by incorporating orientation information between the ground truth and predicted boxes. Experimental results show that EFSF-RBCD achieves an AP@0.5 of 90.3% and an AP@0.5-0.95 of 59.6%, with 26.4M parameters, a computational cost of 50.8 GFLOPs, and a processing speed of 100.3 FPS. The proposed method outperforms existing state-of-the-art methods in terms of AP@0.5 and AP@0.5-0.95 while maintaining a low parameter count and computational load. Additionally, it demonstrated competitive performance in terms of FPS. This study provides a new approach for beef cattle detection in complex environments and lays a theoretical foundation for the development of technologies related to smart-farm deployment.

The rapid development of 3D technology has significantly impacted precision livestock management, enabling the evaluation and optimization of beef cattle growth by analyzing their three-dimensional geometric characteristics. This study presents a non-contact beef cattle surface reconstruction method using multi-view point clouds, captured by a synchronized multi-camera system, followed by point cloud preprocessing, the Intrinsic Shape Signature (ISS) keypoints extraction, and Principal Components Analysis (PCA) feature fusion for accurate alignment and reconstruction. In order to realize the point cloud alignment for 4 angles, the Super 4-Points Congruent Sets (Super4PCS) is used for coarse alignment, and then the Iterative Closest Point (ICP) algorithm accelerated by K-dimension tree (KD-Tree) is used to complete the fine alignment. Aiming at the missing point cloud caused by the obstruction of the railings, the repair method of bilateral filtering combined with the improved cubic B-spline of statistical analysis is proposed. Based on the highest point of the withers, the symmetry plane is determined to realize the reconstruction from unilateral point cloud to complete point cloud. Finally, Withers Height (WH), Hip Height (HH), Oblique Body Length (OBL), Heart Girth (HG), Abdominal Girth (AG), Cannon Girth (CG), Ischial Width (IW), and Hip Width (HW) values were measured by Euclidean Distance and B-spline curve fitting. The mean absolute error of the predicted values for the eight body scales for 100 beef cows was 4.09 cm, and the mean absolute percentage error was 4.01%. The maximum absolute error was 9.09 cm and the minimum absolute error was 0.74 cm. The experimental results show that this method can provide a new and accurate method for the reconstruction of multi-view livestock point clouds.

Graph positive-unlabeled learning is an important task that tries to learn binary classification models from only positive and unlabeled (PU) nodes. While state-of-the-art methods focus on training graph neural networks, they frequently rely on weak objective functions derived solely from a given class prior probability or inferred exclusively from the graph structure, leading their performance significantly lags behind that of fully labeled counterparts. In this paper, we fill this gap by treating unlabeled nodes as samples ambiguously labeled as both positive and negative, and by introducing a learning method called Bootstrap Label Disambiguation (BLD), which progressively resolves label ambiguities during the training of binary classifiers. BLD comprises a node representation learning module via bootstrapping and a novel central region-based label disambiguation strategy. The learning module leverages both previous representations and the derived positive centriod as targets to train positive-aligned representations, eliminating the need for a prior. Consequently, the disambiguation strategy constructs a central-region to identify ambiguous nodes and steadily transforms them into effective supervision. Extensive experiments on a range of real-world datasets show that our BLD method significantly outperforms existing approaches and in many cases even surpasses fully labeled classification models. The source code is available at https://github.com/yunyun85/BLD.

Embedded microcontrollers with low power consumption, high performance, and rich interfaces widely use FCFS scheduling, which suffers from long average wait times and inefficiency for short tasks, increasing energy consumption. To address the lack of clear priority rules in existing ERPT scheduling, this paper proposes a Hierarchical Priority Preemptive Threshold (HPPT) algorithm based on Analytic Hierarchy Process (AHP). HPPT establishes task priorities through energy metrics, builds a low-energy model for prioritized scheduling, and optimizes system-level energy efficiency. Physical experiments show HPPT reduces device power consumption and extends operational lifespan by approximately 24.08%.

Recently, a novel visual state space (VSS) model, referred to as Mamba, has demonstrated significant progress in modeling long sequences with linear complexity, comparable to Transformer models, thereby enhancing its adaptability for processing visual data. Although most methods aim to enhance the global receptive field by directly modifying Mamba's scanning mechanism, they tend to overlook the critical importance of local information in dense prediction tasks. Additionally, whether Mamba can effectively extract local features as convolutional neural networks (CNNs) do remains an open question that merits further investigation. In this paper, We propose a novel model, AtrousMamba, which effectively balances the extraction of fine-grained local details with the integration of global contextual information. Specifically, our method incorporates an atrous-window selective scan mechanism, enabling a gradual expansion of the scanning range with adjustable rates. This design shortens the distance between adjacent tokens, enabling the model to effectively capture fine-grained local features and global context. By leveraging the atrous window scan visual state space (AWVSS) module, we design dedicated end-to-end Mamba-based frameworks for binary change detection (BCD) and semantic change detection (SCD), referred to as AWMambaBCD and AWMambaSCD, respectively. Experimental results on six benchmark datasets show that the proposed framework outperforms existing CNN-based, Transformer-based, and Mamba-based methods. These findings clearly demonstrate that Mamba not only captures long-range dependencies in visual data but also effectively preserves fine-grained local details.

Remote sensing change detection (RSCD), which aims to identify differences between bitemporal images, has made great progress through the application of deep learning methods. Convolutional neural networks and transformers are extensively employed in remote sensing image change detection, achieving promising results. However, current models predominantly focus on visual representation learning while neglecting the potential of multimodal learning methods. Consequently, this leads to issues such as inaccurate identification of nonsemantic changes, incomplete boundary extraction due to the degradation of local feature details, and the loss of small targets. Recently, a novel method for efficiently learning from natural language supervision, known as contrastive language image pretraining, has been proposed. Inspired by this work, we aim to leverage pixel-text relationships to guide the training of detection (RSCD) models. We propose a novel end-to-end multimodal dual-branch Siamese network, named MDS-Net, specifically designed for RSCD. In addition, we propose a U-Shaped vision-language guided transformer multimodal decoder to enhance the interactions between visual and textual information. To capture detailed semantic change features, we also introduce a novel difference feature enhancement. module. Extensive experiments demonstrate that our model outperforms other state-of-the-art methods, with IoU values of 85.60%, 76.81%, 89.55%, 94.51%, and 71.77% on the LEVIR-CD, LEVIR-CD+, WHUCD, CDD, and SYSU-CD datasets, respectively.