Multi-modal sarcasm detection is a challenging research issue in the understanding of communication with multiple modalities, as different kinds of interactions between these modalities are needed to be captured. Although topics often play an important role in sarcasm understanding, limited efforts have been devoted to analyzing topical information for multi-modal sarcasm detection. Therefore, in this paper, we propose a topic-based input feature enhancement method with two-stage training process, which uses multi-modal topics to help sarcasm detection. Specifically, in the first stage of training, an unsupervised topic modeling approach named MultiTop-Clus is proposed. This approach leverages a multi-modal latent topic space to represent topics and is pre-trained with unlabeled video data. Then, in the second training stage, the MultiTopClus model is frozen, and a diffusion model is applied to effectively reconstruct a multi-modal feature representation which is a combination of two kinds of input features: topical features generated by the topic model, and original features used in an existing multi-modal sarcasm detection model. Finally, the reconstructed feature representation of each modality is fed to the existing multi-modal sarcasm detection network to make sarcasm predictions. Systematic experiments are carried out on twelve existing multi-modal sarcasm detection models and two multi-modal sarcasm datasets (a video-based dataset and a tweet-based dataset). The experimental results reveal a notable improvement in overall performance across twelve baseline sarcasm detection models. These findings highlight the effectiveness of our proposed approach.

Classifying each point in global navigation satellite system positioning trajectories as either in-field or on-road is pivotal for analyzing the operational performance of agricultural vehicles. This paper introduces a field-road trajectory segmentation method which significantly enhances segmentation robustness through a self-supervised learning approach. The method involves pre-training a trajectory representation model with self-supervised learning, which is subsequently fine-tuned for trajectory segmentation applications. Our model employs dual encoders: point-level and trajectory-level to capture essential multi-level spatio-temporal features for accurate trajectory segmentation. The point-level encoder focuses on extracting detailed features for individual points and performing point-density classification, while the trajectory-level encoder enriches these features by integrating trajectory similarity computations. Meanwhile, utilizing a combination of Convolutional Neural Networks and Transformer networks, the model adeptly handles both temporal and spatial dependencies in trajectory data, crucial for dynamic adaptation to various trajectories. The accuracy of our method achieves 93.95% and 89.32% on two manually labeled datasets, respectively, and experiments on the raw trajectory dataset demonstrate that a pre-training trajectory representation model can effectively capture the trajectory characteristic. Extensive validation confirms the superior efficacy of the proposed method and its potential impact on the evaluation efficiency of practical agricultural operations. The source code is available at the following address: https://github.com/peanut2code/PreTR-TS.

Manual fruit harvesting remains essential in modern horticulture for its ability to preserve fruit quality and reduce mechanical damage. However, the increasing labour intensity of these operations highlights the need for efficient assistive technologies to sustain productivity and ensure sustainability. Although fully automated systems offer alternatives, their limited adaptability in unstructured environments and high implementation costs restrict large-scale deployment. To address these limitations, a following collaborative robot harvesting-assisted transport system (FCR-HATS) is developed, incorporating a novel collaborative region constrained model (CRCM) to enhance harvesting efficiency and reduce manual workload through human-robot collaboration. CRCM introduces a decision-making framework that transitions the robot's behaviour from conventional passive following to adaptive cooperation by enabling three execution modes: following, stopping, and monitoring. CRCM overcomes the trajectory deviations inherent in the traditional direction-following constraints model (DFCM) by applying collaborative region constraints, allowing the FCR to better align with the human trajectory. Field experiments demonstrate that CRCM significantly improves the accuracy of human-robot trajectory overlap, achieving an average matching cost (AMC) ranging from 0.14 to 0.25 across four benchmark paths. This outperforms traditional DFCM models and reduces collision risk without requiring additional obstacle avoidance strategies. With a unit shrinkage distance of 0.05 m, the FCR achieves precise stopping with lateral and longitudinal errors of 0.17 m and 0.15 m, respectively. Validation in a greenhouse peach orchard further confirms the robustness and continuous operational capability of the proposed system under real harvesting conditions.

Membership inference attacks have emerged as a significant privacy concern in the training of deep learning models, where attackers can infer whether a data point was part of the training set based on the model's outputs. To address this challenge, we propose a novel defense mechanism, AdaMixup. AdaMixup employs adaptive mixup techniques to enhance the model's robustness against membership inference attacks by dynamically adjusting the mixup strategy during training. This method not only improves the model's privacy protection but also maintains high performance. Experimental results across multiple datasets demonstrate that AdaMixup significantly reduces the risk of membership inference attacks while achieving a favorable trade-off between defensive efficiency and model accuracy. This research provides an effective solution for data privacy protection and lays the groundwork for future advancements in mixup training methods.

Multi-modal sarcasm detection is crucial for understanding human communications. A key aspect of multi-modal sarcasm detection is the analysis of emotion incongruity. However, the advancement of emotion analysis in video is hindered by the scarcity of labeled datasets, which are limited in both scale and diversity due to high human annotation cost. In this paper, to deal with different kinds of emotion distributions of open-topic in video data, we propose a simple yet remarkably effective method named Prompted Emotion Distribution Enhancement (PEDE). This method leverages large-scale pre-trained models to generate emotion distributions, thereby enriching the input features for sarcasm detection models. Then, intra- and inter-modality emotion graphs are constructed and a graph attention network (GAT) is used to learn emotion incongruity in input. Extensive experiments demonstrate that our approach can significantly enhance the performance of existing multi-modal sarcasm detection approaches on a sarcasm video dataset.

In precision agriculture, the integration of spatial information through Geographic Information System (GIS) plays a crucial role in optimizing agricultural operations. However, the increasing scale of spatial data poses computational challenges to traditional processor-centric solutions used in GIS due to memory limitations: limited memory capacity and low bandwidth. An architecture shift from processor-centric to memory-centric is necessary for modern data-intensive applications. To overcome the memory bottleneck issues, this paper introduces a novel, callable, pipelined solution in Processing Near Memory (PNM) configuration with large memory capacity and high inner bandwidth tailored for memory-intensive precision agriculture applications. The solution leverages a scalable Computational Memory Module (CMM) with a Geolocation Query Engine (GQE), which addresses the k-Nearest Neighbors (kNN) search problem, a fundamental computation used in GIS for various geolocation queries. Our experimental evaluation, conducted on three distinct datasets, showcased the effectiveness of our proposed CMM. These datasets include a substantial Wheat big data set consisting of 17,847 daily trajectories, as well as smaller Paddy and Corn datasets containing 1,634 and $\text{1, 2 9 0}$ daily trajectories, respectively. In these experiments, our CMM consistently achieved an impressive $\text{9 5 \%}$ and $\text{9 9 \%}$ average reduction in query times and energy consumption compared to conventional processor-centric approaches, respectively. By efficiently managing large-scale spatial data, the proposed solution empowers precision agriculture, contributing to increased productivity, sustainability, and profitability in modern agricultural practices.

With over five million contaminated sites worldwide, accurately characterizing the three-dimensional (3D) distribution of soil contamination is critical for effective risk assessment and site remediation. However, current 3D interpolation methodologies often fail to simultaneously account for spatial correlation and spatial heterogeneity, both of which are critical for capturing the complex spatial structure of subsurface contamination. This study developed a refined 3D interpolation model that integrates site characteristics, spatial position, spatial correlation, and spatial heterogeneity to simulate site contamination and quantify prediction uncertainty. The proposed machine learning (ML) model achieved high predictive performance, with coefficient of determination (R2) values above 0.73 for four heavy metals (HMs). To enhance model generalizability, a stability analysis framework was developed alongside a novel model selection strategy based on random dataset partitioning and random ordering of input covariates, and 1000 random simulations could provide a reliable basis for model screening. This study introduces a new, precise 3D spatial interpolation method. Owing to the easy accessibility of its covariates, it offers high versatility, making a significant contribution to site assessment and remediation efforts.

Field-road classification that automatically identifies the activity (either in-field or on-road) of each point in Global Navigation Satellite System (GNSS) trajectories is a critical process in the behavior analysis of agricultural vehicles. To capture movement patterns specific to agricultural operations, we propose a multi-view field-road classification method, which extracts a physical and a visual feature vector to represent a trajectory point. We propose a task-specific approach using a pre-trained visual model to effectively extract visual features. Firstly, an image is generated based on a point plus its neighboring points to provide the contextual information of the point. Then, an image recognition model, a fine-tuned ResNet model is developed using the pretraining-finetuning paradigm. In such a paradigm, a pre-training process is used to train an image recognition model (ResNet) with natural image datasets (e.g., ImageNet), and a fine-tuning process is applied to update the parameters of the pre-trained model using the trajectory point images, enabling the model to have both general knowledge and task-specific knowledge. Finally, a visual feature is extracted for a point by the fine-tuned model, thereby overcoming the limitations caused by the small-scale generated images. To validate the effectiveness of our multi-view field-road classification, we conducted experiments on four trajectory datasets (Wheat 2021, Paddy, Wheat 2023, and Wheat 2024). The results demonstrated that the proposed method achieves competitive accuracy performance, i.e., 92.56%, 87.91%, 90.31%, and 94.23% on four trajectory datasets, respectively. Extensive experiments demonstrate that our approach can consistently perform better than the existing state-of-the-art method on the four trajectory datasets by 2.99%, 4.42%, 2.88%, and 2.77% in the F1-score, respectively. In addition, we conduct an in-depth analysis to verify the necessity and effectiveness of our method.

Field-road classification, which automatically identifies in-field activities and out-of-field activities in global navigation satellite system (GNSS) recordings, is an important step for the performance evaluation of agricultural machinery. Although several field-road classification methods based only on GNSS recordings have been proposed, there is a trade-off between time consumption and accuracy performance for such methods. To obtain an optimal balance, it is important to choose a suitable field-road classification method for each trajectory based on its GNSS trajectory quality. In this article, a trajectory classification task was proposed, which classifies the quality of GNSS trajectories into three categories (high-quality, medium-quality, or low-quality). Then, a trajectory classification (TC) model was developed to automatically assign a quality category to each input trajectory, utilizing global and local features specific to agricultural machinery. Finally, a novel field-road classification method is proposed, wherein the selection of field-road classification methods depends on the trajectory quality category predicted by the TC model. The comprehensive experiments show that the proposed trajectory classification method achieved 86.84% accuracy, which consistently outperformed current trajectory classification methods by about 2.6%, and the proposed field-road classification method has obtained a balance between efficiency and effectiveness, i.e., sufficient efficiency with a tolerable accuracy loss. This is the first attempt to examine the balance problem between efficiency and effectiveness in existing field-road classification methods and to propose a trajectory classification specific to these methods.

Soil potassium is a vital macronutrient and element for both crops and biogeochemical cycles. However, studies on variety of soil potassium and environmental factors in long term cultivated farmlands are still lack. In this study, both soil properties of topsoil (0-20 cm) and environmental factors of ten soil health monitoring station in China were investigated from 1990 to 2018. Seven learning methods were used to build soil total potassium (TK) prediction model. Results showed that the mean value of soil TK content in 0-20 cm was 15.64 +/- 0.09 g center dot kg(-1)with the range of 3.17-31.04 g center dot kg(-1) . TK significantly increased with both the increasing of soil pH and elevation, while significantly decreased with the increasing of total organic carbon, organic matters, air temperature, amount of precipitation, humidity, and atmospheric pressure. The order of R-2 of soil TK prediction models from low to high was Linear Regression < Support Vector Regression < Decision Tree < Light Gradient Boosting Machine < Random Forest < eXtreme Gradient Boosting < Feedforward neural networks with the highest R-2 at 0.91 and the lowest values of RMSE and EF. Soil pH, air temperature, and precipitation could be important environmental factors affecting significantly soil total potassium. Machine learning methods generally showed better performances than that linear regression one. Based on long-term and in-situ monitor of soil properties and environmental factors, a sustainability management strategy for farmlands could come true.

Static word embedding is still useful, particularly for context-unavailable tasks, because in the case of no context available, pre-trained language models often perform worse than static word embeddings. Although dimension is a key factor determining the quality of static word embeddings, automatic dimension selection is rarely discussed. In this paper, we investigate the impact of word frequency on the dimension selection, and empirically find that word frequency is so vital that it needs to be taken into account during dimension selection. Based on such an empirical finding, this paper proposes a dimension selection method that uses a metric (Mixed Product Distance, MPD) to select a proper dimension for word embedding algorithms without training any word embedding. Through applying a post-processing function to oracle matrices, the MPD-based method can de-emphasize the impact of word frequency. Experiments on both context-unavailable and context-available tasks demonstrate the better efficiency-performance trade-off of our MPD-based dimension selection method over baselines.

Morphological analysis is an important research issue in the field of natural language processing. In this study, we propose a context-free morphological analysis task, namely word-level prefix/suffix sense detection, which deals with the ambiguity of sense expressed by prefix/suffix. To research this novel task, we first annotate a corpus with prefixes/suffixes expressing negation (e.g., il-, un-, -less) and then propose a novel few-shot learning approach that applies an input-augmentation prompt to a token-replaced detection pre-training model. Empirical studies demonstrate the effectiveness of the proposed approach to word-level prefix/suffix negation sense detection.

Adversarial attack serves as a major challenge for neural network models in NLP, which precludes the model's deployment in safety-critical applications. A recent line of work, detection-based defense, aims to distinguish adversarial sentences from benign ones. However, the core limitation of previous detection methods is being incapable of giving correct predictions on adversarial sentences unlike defense methods from other paradigms. To solve this issue, this paper proposes TextShield: (1) we discover a link between text attack and saliency information, and then we propose a saliency-based detector, which can effectively detect whether an input sentence is adversarial or not. (2) We design a saliency-based corrector, which converts the detected adversary sentences to benign ones. By combining the saliency-based detector and corrector, TextShield extends the detection-only paradigm to a detection-correction paradigm, thus filling the gap in the existing detection-based defense. Comprehensive experiments show that (a) TextShield consistently achieves higher or comparable performance than state-of-the-art defense methods across various attacks on different benchmarks. (b) our saliency-based detector outperforms existing detectors for detecting adversarial sentences.

Few-shot learning remains a great challenge for the task of acceptability judgment that identifies whether a sentence is acceptable or unacceptable. In this paper, we propose a prompt-free learning approach, namely PF-ELECTRA, to few-shot acceptability judgment. First, we leverage a pre-trained token replaced detection model, ELECTRA, as our basic few-shot learner to deal with the challenge of data distribution difference. Second, we design a prompt-free few-shot learning strategy that uses both the maximal unacceptability score for a single token and the overall unacceptability score for the whole sentence to judge the acceptability. Empirical studies validate the effectiveness of PF-ELECTRA on challenging few-shot acceptability judgment. To the best of our knowledge, it is the first work that improves the performance of few-shot acceptability judgment based on standard fine-tuning.

The understanding of agricultural machinery, whether it is used "in field" or "on road", plays an essential role in optimizing the efficiency of cross-regional agricultural mechanization services that have been available in China for decades. With the widespread availability of Global Navigation Satellite System (GNSS)-enabled devices, developing a method for automatically identifying the activity associated with each point in GNSS recorded trajectories can significantly enhance the optimization of these services. In this paper, we propose a novel fieldroad classification method that utilizes two feature vectors to represent each point in a GNSS trajectory. The first feature vector is a statistical feature vector extracted with existing methods, while the second feature vector is a visual feature vector obtained through an image segmentation model. To extract visual features, we first convert each GNSS trajectory into a trajectory image that encodes the motion information (e.g., speed, direction) of each point and the spatiotemporal relationship between points. Then, an image segmentation model specifically for field-road classification was developed, which can effectively extract a visual feature vector for each pixel in a trajectory image, corresponding to each point in the GNSS trajectory. The image segmentation model used in our approach has a robust feature extraction capability that enhances the statistical feature representation of each point by adding pixel-level visual features. We evaluated the effectiveness of our multi-view field-road classification method using trajectories of corn, wheat, and paddy harvesters. The results demonstrated that our method achieved accuracy scores of 94.17%, 90.93%, and 83.43% for respective datasets. Our method outperformed state-of-the-art field-road classification methods, especially for trajectories with high-frequency GNSS acquisition.

The classification that distinguishes whether machines are driving on roads or working in fields based on their global navigation satellite system (GNSS) trajectories is essential for effective management of cross-regional agricultural machinery services in China. In this paper, a novel field–road classification method utilizing multiple deep neural networks (MultiDNN) is proposed to enhance the accuracy of field and road point classification. The MultiDNN model incorporates a bi-directional long short-term memory network (BiLSTM), a topology adaptive graph convolution network (TAG), and a self-attention network (ATT) to effectively extract spatio-temporal features for field–road classification. The BiLSTM is used to capture temporal relationships along the time axis of a trajectory, providing global contextual information for each point. Then, the TAG network is used to obtain the spatio-temporal relationships between adjacent points in a trajectory, offering local contextual information for each point. Finally, the ATT network assigns varying weights to features to emphasize important characteristics. The performance of the MultiDNN model was evaluated using a wheat harvesting trajectory dataset, and the results showed that it achieved a high degree of accuracy, up to 89.75%, outperforming the best baseline method (GCN) by 2.79%.

• A multi-view density-based field-road classification method was proposed. • An object detection method was used for field-road classification. • The combination of two field-road classification results used the DBI metric. Field-road classification that automatically identifies the operation modes (either in-field or on-road) of GNSS (Global Navigation Satellite System) points plays an important role for the operational performance analysis of agricultural vehicles. Intuitively, a field often has high point density because in-field driving speed is rather low and the distance between consecutive strips is closed. In this paper, two methods were used to capture the in-field high-density characteristic: DBSCAN (Density-Based Spatial Clustering of Applications with Noise) and an object detection model. DBSCAN is a widely-used density-based clustering algorithm, which clusters the points with high point density into a cluster. Alternatively, a trajectory can be transformed into an image, and an object detection model can be applied to detect objects in the image, where an object is a set of pixels with high pixel density (i.e., a set of points with high point density). Finally, the two field-road classification results are combined using DBI (Davis Bouldin index), a metric which can evaluate the quality of either classification result. The developed method was validated by the harvesting trajectories of two crops (wheat and paddy), and the density-based field-road classification achieved 85.97% and 73.34% accuracy on the wheat data and the paddy data, respectively.

针对全国范围农机作业动态监测和量化统计的应用需求,该研究通过在农业机械上安装北斗终端,制订数据传输规范,完成了基于北斗的农机作业大数据系统建设.该系统由农业机械及北斗终端、农机制造企业物联网平台和农机作业大数据管理服务平台3部分组成.系统共接入农机290153辆.经数据清洗、轨迹分割和参数提取3个数据处理步骤,可获得农机的工作时长、行驶里程和作业面积等基本统计量.以2021年夏小麦机械化收割为例,利用该系统进行数据获取、处理和统计分析,输出收割机分布热力图和作业重心转移图,进行了收割时长、收割效率与收割面积等统计,分析了小麦主产区对跨区作业的依赖程度.麦收期间在线收割机累计35243辆,日均18568辆,收割时长中位数均值为8.3 h/d,收割面积中位数均值为5.5 hm2/d,约75％的小麦收割机进行了跨区作业,跨区距离中位数约为597 km.应用结果表明,农机作业大数据系统可准确开展数据处理和作业统计,可以向农业农村部门、农机制造企业、农机合作社和农机手提供作业动态监测和数据分析服务.

Although soil total phosphorus (TP) is a primary and essential large element reflecting the soil fertility in agricultural ecosystems, studies on model development of TP and its differences between wheat and paddy lands after a long cultivation history at a regional scale are still limited. Hence, a comparison model of TP with different learning methods and datasets were built, and the relationship between environmental factors and TP were discussed. TP from a long cultivation of either wheat or paddy agriculture systems was investigated, and the regression between TP and climate parameters (air temperatures, precipitation, humidity, and atmospheric pressure) and latitude were analyzed. A comparison of model development with six learning methods, including one statistical learning method (linear) and five machine learning methods (support vector regression, decision tree, random forest, XGBoost, and LightGBM), and two datasets (0–20 and 0–170-cm soil layers) was made. The models were evaluated by the root mean squared error (RMSE), mean deviation (RMD), mean absolute error (MAE), and model effective (EF). The results showed that the TP content of the top soil layer in wheat lands (0.89 ± 0.01 g kg−1) was significantly higher than that of paddy lands (0.63 ± 0.01 g kg−1). The annual average precipitation, humidity, and air temperature had significant negative relationships with TP content, while the annual average atmospheric pressure and latitude had significant positive relationships with TP. Most machine learning methods showed better performances than that of a statistical learning method with the highest r2 of 0.82. The different datasets used for model development had no significant effect on model performances. The average TP content of the top soil layer tends to be greater in wheat lands than that in paddy lands after a long cultivation. Other than the different statistical parameters (the average, maximum, and minimum values) of each climate parameter, comprehensive climate parameters including the annual, semiannual, quarterly, and monthly air temperature, precipitation, humidity, and atmospheric pressure should be considered for further model development. Although different datasets in variable soil depth had no significant effect on model performances, machine learning methods such as random forest, XGBoost, and LightGBM are recommended for better performance than a linear learning method for soil TP model development. It is recommended that a comparison of different machine learning methods will help build a stronger model in similar studies.