Effective weed control is crucial for Salvia miltiorrhiza cultivation, yet traditional methods are often inefficient, costly, or polluting. To address this, this study developed a laser weeding robot based on an improved object detection model capable of identifying weeds and implementing targeted strategies. First, a self-propelled laser weeding robot was constructed for Salvia miltiorrhiza fields to meet operational requirements. Second, a real-world field dataset was established for Salvia miltiorrhiza and five weed families. The detection model, optimized from the You Only Look Once (YOLO) architecture, integrates attention-based feature interaction, dynamic spatial attention, and small object feature enhancement modules. These improvements enhanced the features of small objects, improved occluded target localization, and strengthened similar object discrimination. Third, drawing on weed biological characteristics, a multi-level, differentiated laser weeding strategy was developed to precisely target growth points while ensuring crop safety. Finally, the model and strategy were deployed on the robot to perform real-time detection and intelligent laser weeding. Test results demonstrate the superior performance of the proposed model: the precision of object detection reached 78.09% (2.54% over baseline) and that of keypoint detection stood at 80.69% (8.22% over baseline). The mean average precision ( mAP50 ) metrics improved to 78.14% and 80.56%, representing increases of 2.34% and 2.88% respectively. Field tests achieved a 90.2% weed control rate alongside a low 1.9% damage rate to Salvia miltiorrhiza . These results validate the system's effectiveness and practicality, providing crucial technical support for intelligent weed management in Salvia miltiorrhiza and other high-value medicinal crops.

Nowadays, the importance of data privacy protection has grown significantly. Privacy Set Intersection (PSI) based on Fully Homomorphic Encryption (FHE) is widely applied in various privacy protection scenarios, such as federated learning and password verification. Nevertheless, the substantial computational demands of FHE and the vast scale of databases in PSI result in inefficient processing, thereby necessitating specialized accelerator architectures to enhance usability. Current general-purpose FHE accelerators do not adequately address the unique requirements of PSI applications, leading to suboptimal data handling and underutilization of hardware, which impedes their effective deployment for PSI acceleration. This paper introduces Ares, a practical hardware-software co-designed FHE-based PSI FPGA accelerator. We propose Lazy Relinearization to optimize redundant calculations in PSI and reduce computational complexity without changing the PSI protocol. At the same time, through the analysis and decoupling of the PSI computing pattern, we design an efficient hardware acceleration architecture that fully utilizes the bandwidth and computing resources of the hardware to achieve excellent acceleration performance. We highlight the following result: (1) a $47.99 \times$ speedup relative to CPU; (2) performance improvements of $1.79 \times$ and $1.93 \times$ over the state-of-theart FPGA FHE accelerators, Poseidon and FAB, respectively; (3) achieves $7.96 \times$ and $10.95 \times$ energy efficiency improvement compared to Poseidon and FAB, respectively.

Fully Homomorphic Encryption (FHE) neural networks has made significant strides in enabling privacy-preserving inference, yet it is frequently impeded by substantial computational overhead. Previous research has largely focused on optimizing neural network operations or managing FHE computation processes to reduce computational cost, but has overlooked addressing the interplay between FHE computation process and neural network characteristics as an integrated system. This paper introduces AceHomo, a novel approach utilizing reinforcement learning to optimize CKKS encryption level consumptions in neural networks, thereby enhancing performance and reducing latency without compromising network accuracy. Experimental results demonstrate that AceHomo significantly reduces latency by up to$1.47 \times$without compromising the network's inference accuracy. These encouraging results position AceHomo as a promising approach in optimizing FHE neural networks.

Private Set Intersection (PSI) is imperative in discovering the properties of the same data owned by two competitive parties, without revealing anything else of their respective data asset. Existing PSI solutions such as APSI and ORI-PSI suffer from severe communication and computation overhead due to inefficient communication and FHE polynomial evaluation, which hinders their deployment in practice. This issue is evident in both the upper-level protocol and the lower-level hardware platform. In this paper, we propose a novel software/hardware co-design acceleration architecture for PSI, termed as “Trident”, which includes two tightly coupled segments: from the protocol perspective, we investigate existing bottlenecks and propose a new PSI protocol with significantly less communication and computation under the security guarantee; besides, we re-architect the hardware platform by designing a PSI-specific accelerator, implemented with both FPGA and ASIC, targeting the key operations in the proposed protocol. We build a real-world experimental environment with two instantiated parties to verify the acceleration architecture, and highlight the following results: (1) up to 130$\boldsymbol{\times}$/145$\boldsymbol{\times}$speedup for the computation ofreceiverandsenderparties; (2) up to 37$\boldsymbol{\times}$reduction of communication overhead. (3) up to 93,651$\boldsymbol{\times}$and 74,326$\boldsymbol{\times}$higher energy efficiency over the CPU-based ORI-PSI and APSI, respectively.

Federated Learning (FL) enables collaborative model training across devices, but data exchanges pose privacy risks. Homomorphic Encryption (HE) is widely used to enhances privacy in FL but incurs significant communication and computation latency. Prior work reduced this latency using compressions, but sacrificed learning accuracy and overlooked the impact of the number of participating devices on latency. Over-the-air computation (AirComp) leverages wireless channels' superposition property to achieve high spectral efficiency and efficient aggregation irrespective of device number. In this paper, we propose HEAirFed, integrating AirComp with the state-ofthe-art HE scheme CKKS for efficient privacy-preserving FL. In HEAirFed, we develop a ciphertext-oriented wireless communication module to ensure homomorphic operations leverage AirComp's superposition property, enabling correct decryption. We further build a rigorous error analysis model, derive the worst-case upper bound of approximation error, and characterize this bound's impact on the convergence guarantee of HEAirFed, measured by the optimality gap with bounded approximation error. Then, we minimize this gap and derive a near-optimal solution in semi-closed form. Extensive experimental results on real-world datasets validate the ciphertext-oriented design's necessity, the error analysis's correctness, and demonstrate that HEAirFed achieves a substantial reduction in communication and aggregation latency compared to baseline, with minimal learning accuracy loss.

In the era of big data , information overload poses a challenge, complicating user decision-making. Recommender systems aim to assist in this process. In recent years, research on point-of-interest (POI) recommendations has been gaining momentum with some studies pointing to issues that need to be resolved. Previous studies often used heterogeneous graphs to learn across different entity types, overlooking same-type entity relationships. Certain studies solely extract raw node features from a single source, thus disregarding information diversity, whereas others employ inappropriate methods that fail to preserve the inherent characteristics of the relevant information in the design of raw inputs. The integration of multiple sources of information can introduce a certain amount of noise into the data; however, the approaches used in related research may not be effective in handling this situation. To address these issues, we propose a hybrid structural graph attention network (HS-GAT) for POI recommendation. In this approach, multisource data are first preprocessed and relevant raw features are initialized. Subsequently, heterogeneous graphs are built for user-POI-POI attributes and POI-user-user attributes. These heterogeneous graphs are aggregated using a dual-attention mechanism, to create embedding matrices for users and POIs, which are then used to construct user-user and POI-POI homogeneous graphs. These graph structures are then combined with user and POI embeddings obtained from heterogeneous graphs and fed into a graph attention network (GAT), which yields the final embedding representations for users and POIs. Finally, recommendations for POIs are made in the form of inner products. A comprehensive performance evaluation of HS-GAT on the Yelp, Boston, Chicago and London datasets demonstrated that the proposed approach outperforms other state-of-the-art methods.

Natural language processing (NLP) has been shown to perform well in various tasks, such as answering questions, ascertaining natural language inference and anomaly detection. However, there are few NLP-related studies that touch upon the moral context conveyed in text. This paper studies whether state-of-the-art, pre-trained language models are capable of passing moral judgments on posts retrieved from a popular Reddit user board. Reddit is a social discussion website and forum where posts are promoted by users through a voting system. In this work, we construct a dataset that can be used for moral judgement tasks by collecting data from the AITA? (Am I the A*******?) subreddit. To model our task, we harnessed the power of pre-trained language models, including BERT, RoBERTa, RoBERTa-large, ALBERT and Longformer. We then fine-tuned these models and evaluated their ability to predict the correct verdict as judged by users for each post in the datasets. RoBERTa showed relative improvements across the three datasets, exhibiting a rate of 87% accuracy and a Matthews correlation coefficient (MCC) of 0.76, while the use of the Longformer model slightly improved the performance when used with longer sequences, achieving 87% accuracy and 0.77 MCC.

Micro-expressions are brief facial movements characterized by short duration, involuntariness and low intensity. Recognition of spontaneous facial micro-expressions is a great challenge. In this paper, we propose a simple yet effective Main Directional Mean Optical-flow (MDMO) feature for micro-expression recognition. We apply a robust optical flow method on micro-expression video clips and partition the facial area into regions of interest (ROIs) based partially on action units. The MDMO is a ROI-based, normalized statistic feature that considers both local statistic motion information and its spatial location. One of the significant characteristics of MDMO is that its feature dimension is small. The length of a MDMO feature vector is 36 × 2 = 72, where 36 is the number of ROIs. Furthermore, to reduce the influence of noise due to head movements, we propose an optical-flow-driven method to align all frames of a micro-expression video clip. Finally, a SVM classifier with the proposed MDMO feature is adopted for micro-expression recognition. Experimental results on three spontaneous micro-expression databases, namely SMIC, CASME and CASME II, show that the MDMO can achieve better performance than two state-of-the-art baseline features, i.e., LBP-TOP and HOOF.

Estimation of canopy biophysical and biochemical parameters using remote sensing data is important for regional crop-growth condition monitoring and yield assessment. The inversion of the radiative transfer model PROSAIL based on the look-up table (LUT) approach is widely used for this purpose, taking remotely sensed reflectance as input. For the LUT-based parameter mapping, the main part is searching for the optimal solution from a large LUT. Due to the computational complexity of the sorting algorithm and size of the LUT for the solution search, a substantial amount of time is normally needed for estimation. In order to speed up the mapping of parameters using remote sensing observations, a faster method is developed for searching the LUT by introducing a binary search algorithm. The results of the experiments based on SPOT-5 imagery show that the proposed method can increase the mapping speed by about 70 times compared to the sorting algorithm.

This paper presents Template-Guided Classification (TGC), a technique for using the class labels of existing large-area land-cover maps to automatically classify spatially highresolution images. TGC uses land-cover images as templates to guide hierarchical clustering and labeling. To test TGC, 10-m SPOT 5 HRG images and 1-m colour orthophotos of the Vermilion River watershed, Canada were classified into forest/non-forest classes using the 25-m Earth Observation for the Sustainable Development of forests (EOSD) landcover map as a template. Although the average accuracies of the 10-m SPOT classifications were poor, the 1-m orthophoto accuracies were much higher (87% forest user's accuracy, 82% forest producers accuracy, 93% overall accuracy). TGC classification accuracies were highly variable. Further investigation is needed to determine whether TGC can be made into a robust procedure.

Tree species composition is an indicator of forest type. It is also a required attribute in forest inventory, biomass and stand volume estimation. Accurate mapping tree species is essential for forest management purposes. In this paper the performances of LiDAR, RapidEye data, and their combination on tree species classification were investigated in a boreal forest. Both Random forest (RF) and support vector machine (SVM) classification methods were performed. Results indicated that combined LiDAR and RapidEye data improved the classification accuracy significantly, compare to using each type of data separately. The RF classifier outperformed SVM for tree species classification. Six variables that contributed most to classification accuracy were digital elevation model, slope, canopy height, red-edge NDVI, and red-edge and Near infrared bands of RapidEye data.

Fractional abundances predicted for a given pixel using spectral mixture analysis (SMA) are most accurate when only the endmembers that comprise it are used, with larger errors occurring if inappropriate endmembers are included in the unmixing process. This paper presents an iterative implementation of SMA (ISMA) to determine optimal per-pixel endmember sets from the image endmember set using two steps: 1) an iterative unconstrained unmixing, which removes one endmember per iteration based on minimum abundance and 2) analysis of the root-mean-square error as a function of iteration to locate the critical iteration defining the optimal endmember set. The ISMA was tested using simulated data at various signal-to-noise ratios (SNRs), and the results were compared with those of published unmixing methods. The ISMA method correctly selected the optimal endmember set 96% of the time for SNR of 100 : 1. As a result, per-pixel errors in fractional abundances were lower than for unmixing each pixel using the full endmember set. ISMA was also applied to Airborne Visible/Infrared Imaging Spectrometer hyperspectral data of Cuprite, NV. Results show that the ISMA is effective in obtaining abundance fractions that are physically realistic (sum close to one and nonnegative) and is more effective at selecting endmembers that occur within a pixel as opposed to those that are simply used to improve the goodness of fit of the model but not part of the mixture

Spectral mixture analysis (SMA) has been used extensively in the hyperspectral remote sensing community for the subpixel abundance estimation of targets. However, the task of defining every endmember can be difficult, as evident from the importance attributed to the topic in the recent literature. The effectiveness of SMA can be compromised when the required spectral endmembers are not well constrained in terms of their spectral magnitude and shape. The spectral magnitude of the endmembers is more difficult to obtain than their spectral shape, in part because the effects of the atmosphere and topography are difficult to constrain. This paper presents a derivative spectral unmixing (DSU) model, which is an extension of the spectral mixture analysis and derivative analysis. Using a DSU approach, it is possible to estimate the fraction of an endmember characterized by one or more diagnostic absorption features despite having only a general knowledge of the spectral shapes of the remaining endmembers. The DSU is assessed using spectral data acquired for a lichen-covered rock sample, and the estimated fractions of lichen and rock are assessed against that obtained from a high spatial resolution digital photograph. The results of the laboratory experiments suggests that the DSU is a promising algorithm for the quantitative analysis of hyperspectral data, but experiments on airborne/spaceborne imagery are now required to assess its value for geological mapping.