With the rapid development of autonomous driving technology, road traffic detection, as a core task of the perception module, directly impacts the safety and reliability of autonomous driving systems. Although deep learning-based methods have become a research hotspot, challenges such as low detection accuracy and poor model generalization remain. To address these issues, this paper proposes an improved YOLO11n-based road traffic detection method. The proposed approach enhances the detection accuracy of small objects by adding a small object detection layer, optimizes the existing dual DWConv structure by introducing a GhostConv+DWConv detection head combination, and designs an Inner-CIoU loss function better suited for small objects to improve model generalization and the accuracy of bounding box regression. Experimental results show that, compared with the existing YOLO11n algorithm, the proposed model achieves detection accuracy improvements of 1.1% and 1.9% on the KITTI and BDD100K datasets, respectively, with detection speeds of 125 FPS and 124 FPS. This demonstrates the model’s effectiveness in detecting low-resolution small objects and its strong generalization capability across diverse traffic scenarios.

Due to the complexity of traffic flow and the insufficient extraction of spatio-temporal features, it is difficult for D2STGNN to capture the dynamic changes of traffic networks, which limits the improvement of prediction accuracy. In this paper, a Bi-EMHGRU model combining an efficient multi-head self-attention mechanism (EMHSA) and a bidirectional gated recurrent unit (BiGRU) is proposed. This model captures the sequential dependencies of both forward and backward timings through BiGRU and dynamically allocates weights to each time step by using the multi-head self-attention mechanism to focus on key sequential features. Meanwhile, a multi-scale time feature extraction module is introduced, which enhances the modeling ability for short-term fluctuations and long-term trends and improves the modeling effect of complex spatio-temporal dynamics. The experimental results show that Bi-EMHGRU performs excellently on the PEMS04 and PEMS08 datasets. The root mean square error value has decreased by approximately 0.55~1.55, the mean absolute error has decreased by approximately 0.89~1.40, and the mean absolute percentage error has decreased by approximately 0.86~1.77 percentage points. It can still maintain stable prediction performance when the prediction step length increases and has strong generalization ability. Compared with the existing benchmark models, Bi-EMHGRU can capture the dynamic spatio-temporal features of traffic flow more effectively and significantly improves the prediction accuracy and robustness.

Nowadays, the safety of electric vehicle (EV) riders has now become a focal issue in society, and wearing safety helmets was proven to be an effective way to reduce injury in accidents. In order to enhance road traffic safety and improve regulatory efficiency, an UAV-assisted helmet intelligent detection algorithm based on intelligent communication and deep learning is proposed. By combining intelligent communication technology, UAVs can transmit video data in real time and analyze it quickly by intelligent algorithms. First, an improved Outlook-C2f architecture was proposed to enhance the algorithm’s focus on the small targets; Second, CARAFE is proposed to used in the Feature Pyramid Network (FPN) to dynamically generate weights for precise feature reconstruction and improved spatial resolution; Finally, WIoU (Wise Intersection over Union) was integrated to improve the accuracy of positional information. The experimental results show that, based on the road real-time dataset, the improved YOLOv8 algorithm achieves 96.7% mAP and 26.91 FPS, which are significantly better than the traditional method, demonstrating its potential for application in complex traffic scenarios.

Non-intrusive load monitoring (NILM) facilitates the rational energy allocation and fine-grained management using real-time load data monitor and analysis. To improve load identification performance in NILM under conditions of limited labeled data, this paper presents a novel method based on Bi-LSTM feature fusion and fine-tuned few-shot learning (FT-FSL). First, weighted pixel voltage-current (V-I) image features and multidimensional time-frequency sequence features are fused using Bi-LSTM feature fusion method. Then, FT-FSL is employed to enable the load classification model to be trained with only a small number of labeled samples. Finally, the proposed method is evaluated on the PLAID dataset and compared with four mainstream FSL approaches (Matching Network, Prototypical Network, Relation Network, and MAML). Experimental results show that the proposed method achieves an accuracy of 92.46%, outperforming the comparison models by 12.21, 4.18, 5.90, and 9.04 percentage points, respectively. These results demonstrate the effectiveness of the proposed approach in identifying load types with limited labeled data.

Aiming at the problems of high computational complexity, large number of parameters, slow segmentation speed and low segmentation accuracy existing in the photovoltaic cell defect segmentation model, a photovoltaic cell defect segmentation model based on lightweight improved U-Net is proposed. First of all, the MobitNetv3_Large network is used to replace the backbone network of the original U-Net, which reduces the computational amount and the number of parameters of the model while retaining the feature extraction ability of the original network. Secondly, the G-DConv module is designed by integrating the DynamicConv module into the GhostConv module, replacing the ordinary convolutional module used in the upsampling part of the original U-Net, which maximally reduces the network parameters and computational amount while improving the inference speed of the model. Finally, by introducing the ECA attention mechanism after network upsampling, the interference of complex background on the detection effect is reduced. The experimental results show that the number of parameters of this model is only 2.43×10⁶, the computational amount is only 3.03×10⁹, and the inference speed reaches 61 frame/s. Compared with the baseline model, the improved model increases MIoU and MPA by 0.12 and 2.17 percentage points respectively, meeting the requirements for industrial equipment deployment.

To address the issues of low detection accuracy, high model complexity, and excessive computational costs in small-target defect detection of printed circuit boards (PCBs), which hinder deployment on edge devices, a lightweight algorithm based on YOLO11n was proposed. Firstly, the BiMAFPN (Bi-Directional Multi-Branch Auxiliary Feature Pyramid Network) architecture is employed to reconstruct the network structure. Subsequently, the C3k2_Faster module is implemented to reduce model complexity while maintaining detection accuracy. Finally, the LSCD (Lightweight Shared Convolutional Detection) head is introduced to enhance precision. Experimental results demonstrate that the proposed model achieves 93.0% precision and 82.8% recall, with a compact model size of 3.8 MiB. Enhancements include a 0.6 percentage points increase in precision. The mean average precision (mAP) values reach 89.9% (mAP@0.5) and 47.1% (mAP@0.5:0.95), representing improvements of 1.4 and 0.6 percentage points respectively compared with the baseline YOLO11n model while reducing model size, computational complexity, and parameter count by 30.9%, 19.0% and 34.6% respectively. These optimizations enable the improved algorithm to maintain competitive detection performance while achieving significant lightweight characteristics, demonstrating strong potential for practical deployment in edge computing environments.

Fake news can have serious consequences if not dealt with in a timely manner. Currently, various attention mechanisms are mainly employed in multimodal fake news detection methods to fuse unimodal features. The semantic gaps that may exist between different modal features are not taken into account, nor is the potential of multimodal pre-training models fully exploited. To this end, a new multimodal fake news detection model that performs multistage fusion of features is proposed in this paper. The pretrained multimodal model is utilized by the proposed model to extract the aligned features. Then the features are enhanced by each other through the attention mechanism, and the enhanced features are spliced to achieve early fusion. Finally, the interaction information between different modal features is captured by the multiscale fusion module, and the fusion weights are learned to realize the late fusion of features. It is shown by the experimental results that the model proposed in this paper achieves better results than similar models, and the effectiveness of the attention mechanism and the multiscale fusion module is also verified by the experimental results.

Emotion plays a significant role in fake news detection. Existing research mainly focuses on extracting emotional features from a linguistic perspective, which fails to explore the relationships between emotions from a psychological perspective.In addition, the connection between sentiment features and text features is ignored by existing work, and the potential semantic information of news can not be explored fully. To address the above issues, a fake news detection mode(FNEK) is proposed by this paper, which integrates psychological and emotional knowledge. The Plutchik’s wheel of emotions theory from psychology is incorporated by the model to extract emotional features, and the emotional features are combined with textual features from local and global perspectives, which enhances the accuracy and reliability of fake news detection.Experimental results on publicly available Politifact, Weibo16, and Weibo20 datasets show that the proposed model improves accuracy by 2.1, 0.7 and 2.5 percentage points, respectively, compared with state-of-the-art baseline models.

In multimodal sentiment analysis tasks, effective fusion of sentiment information between different modalities is hindered due to three challenges: nonverbal modal information being insufficiently utilized, fine-grained associative modeling for cross-modal interactions being inadequately established, and hierarchical semantic fusion mechanisms being imperfectly designed. To address these issues, a multimodal sentiment analysis method with cross-modal feature enhancement and hierarchical MLP communication is proposed in this paper. A progressive fusion architecture is constructed, where nonverbal modal information is first enhanced through a cross-modal attention mechanism, enabling many-to-many cross-modal fine-grained interactions to be captured. Subsequently, a lightweight hierarchical MLP communication module is designed to implement hierarchical feature interactions in both horizontal and vertical dimensions, through which cross-modal deep semantic fusion is achieved. It is demonstrated by the experimental results that compared with the suboptimal model on CMU-MOSI, The Acc2 and F₁ values increase 0.89 percentage points and 0.77 percentage points compared with the suboptimal model. Furthermore, all metrics in comparative experiments on CMU-MOSEI are shown to surpass those of baseline models, with the Acc2 and F₁ values being elevated to 86.34% and 86.25%.

In this paper, the degenerate Hopf bifurcation is investigated at high-order singular point in a class of three-dimensional systems. Based on the center manifold theorem, a formal series method for directly calculating singularity quantities is proposed, which avoids the tedious process of converting the original three-dimensional system into the plane reduction equations. The corresponding linear recursive formula of this algorithm is easy to execute. A class of fourth-order system is specifically studied, its center problem of high-order singular points is solved and the cyclicity of degenerate Hopf bifurcation is determined.

For high-dimensional sparse linear regression models, this paper proposed an FDR rule for model selection based on false discovery rate (FDR) from the perspective of posterior estimation, and then introduced a dynamic signal-to-noise ratio (SNR) change factor on this basis. The FDR_R rule, which was more robust to SNR variations and was invariant to data scale, was proposed. Combined with the OMP algorithm, simulation experiments compared the probabilities of successfully selecting all true variables and the FDR values for the FDR rule, FDR_R rule, and existing rules. The results show that the FDR_R rule is more robust than the other rules in high SNR or large sample sizes, more resistant to data scaling issues, and achieves the lowest FDR. Finally, the proposed method was applied to real data from patients with mantle cell lymphoma, identifying genes associated with cell proliferation.

This paper investigates the blow-up criteria for the Navier-Stokes equations of inhomogeneous incompressible fluids. While the existence of weak solutions to these equations has been extensively studied, the existence of strong solutions remains an open question, particularly for highly regular data that meet specific compatibility conditions, where only local existence results are currently available. The focus of this study is on the blow-up criteria within a two-dimensional bounded smooth domain, which refers to the conditions that cause solutions to lose regularity. By deriving priori estimates independent of the density lower bounds and establishing the existence and uniqueness of local strong solutions to the initial value problem or the initial boundary value problem, this paper offers new theoretical insights into the stability and existence of solutions in the dynamics of inhomogeneous incompressible fluids.

In order to further explore the physiological mechanism of cassava resistance to bacterial wilt disease, 20 cassava germplasm with different resistance abilities were selected as the research objects, and cultured at Seedling stage (1.5M) and the grown up stage (3.5M) after planting, and the resistance evaluation was carried out 16 days later. The activities of phenylalanine ammoniase (PAL), superoxide dismutase (SOD), polyphenol oxidase (PPO), catalase (CAT), peroxidase (POD), malondialdehyde (MDA) and soluble sugar in cassava leaflets were determined. The contents of sugar, soluble protein (sPRO), proline (proline) were analyzed, and the relationship between the activity of defense-related enzymes and the content of contents and the infection process of bacterial wilt disease and the disease resistance of varieties were analyzed. The results showed that PAL, PPO, POD activities and MDA, soluble sugar, soluble protein and proline contents increased first and then decreased, and there were significant differences between them and healthy leaflets under 3.5M treatment. SOD and CAT activities were opposite. The proportion of lobular lesions was significantly negatively correlated with SOD and CAT, and positively correlated with other indexes. The defense enzyme activity and contents of resistant varieties were generally higher than those of susceptible varieties. The activities of SOD, POD, PPO, soluble sugar and proline in SC124 tetraploid were higher than those in SC124, and the activities of PAL, CAT and PPO in SC205 tetraploid were higher than those in SC205. The results indicated that PAL, PPO, POD, SOD, CAT activities and MDA, soluble sugar, soluble protein and proline contents were involved in the disease resistance control process of cassava bacterial blight.

In order to explore the influence mechanism of different Ca/Cd ratios in soil on plant Cd absorption and accumulation and its physiological and biochemical characteristics in carbonate areas with high Ca and Cd background. In this study, the Cd accumulation characteristics, physiological and biochemical characteristics of Capsicum annuum L. (capsicum) were discussed based on different ratios of Ca/Cd in production substrates by using regional simulated pot experiments. The results showed that with the continuous growth of capsicum, the effects of different Ca/Cd treatments on the growth rate of capsicum were gradually obvious, and the photosynthesis, root activity and biomass were significantly improved. The increase of Ca/Cd ratio in the substrate promoted Cd accumulation in roots, reduced Cd transfer to stems and leaves, At the same time, it promoted the accumulation of Ca in leaves, increased the Ca/Cd ratio in leaves, promoted the antioxidant enzyme activity (SOD, POD, and CAT activities increased by 17%-69%, 16%-268%, and 9%-136%, respectively) and proline content of leaves, and the malondialdehyde and protein carbonyls were significantly reduced. In conclusion, Ca regulates Cd tolerance in capsicum by enhancing Cd fixation by roots, reducing the transfer process of Cd to aerial parts, and enhancing growth and development and stress resistance, and this detoxification mechanism is closely related to different Ca/Cd stoichiometry.

Soil nutrient supply imbalance leads to nutrient limitation for microorganisms; however, the microbial response strategies employed in different types of soils and aggregates towards it, as well as the impact on soil organic carbon (SOC), remain unclear. This study selected typical karst neutral calcareous soil (LS) and acidic red soil (RS) under secondary forest vegetation coverage in the subtropical monsoon region of China for comparative analysis. Nutrient limitation characteristics were evaluated using indicators such as the stoichiometric ratios of carbon (C), nitrogen (N), and phosphorus (P) in total amounts, bioavailable fractions, microbial biomass and extracellular enzymes as well as the chemical stoichiometric imbalances between microbes and resources in soil and aggregates. Additionally, the response strategies of different soil microbial communities to nutrient limitation and their impacts on SOC were analyzed. The results indicated that different indicators reflected various nutrient limitation characteristics. Notably, the stoichiometric ratios of bioavailable fractions and extracellular enzymes, along with the chemical stoichiometric imbalances determined the microbial response strategies. Overall, both soil types exhibited co-limitation of N and P, with RS experiencing a comparatively stronger limitation of both elements and stronger N and P limitations were observed in its macro-and medium-aggregates. Driven by soil dissolved organic carbon (DOC), the macro-and medium-aggregates microorganisms tended to enhance their N and P utilization efficiency (NUE and PUE) by increasing microbial phosphorus entropy (qMBP) and lowering the intracellular nutrient homeostasis to adapt to nutrient limitations, a process that resulted in reduced carbon utilization efficiency (CUE). To the contrary, LS kept a higher CUE. The SOC content was not influenced by the chemical stoichiometric imbalance or the microbial response strategies; instead, soil physicochemical properties, particularly pH and calcium content, emerged as key driving factors for SOC levels. The findings of this study provide a theoretical basis for the management and restoration of soil nutrients and carbon pools in subtropical forest ecosystems in China.

The study aims to investigate the effects of Fenlong tillage on the soil fungal community in the rhizosphere and bulk soils during the tuberous root formation period of cassava, and to reveal its yield-enhancing mechanism, providing theoretical basis for optimizing the existing cassava cultivation methods. Using South China 205 as the material, the conventional tillage as the control, the cultivation of cassava with Fenlong tillage, combined with high-throughput sequencing technology and bioinformatics methods, combined with soil agrochemical analysis methods, the study investigated the impact of Fenlong tillage on the fungal community diversity in the rhizosphere and bulk soils during the period of cassava root tuber formation. The results showed that the dry weight yield of cassava tubers root was significantly increased by the powder tillage, reaching 7.94 ± 1.13 t/hm², which was 61.1% higher (P＜0.01) than that of the conventional tillage (4.93 ± 0.73 t/hm²), indicating that the Fenlong tillage had a significant yield-increasing effect on cassava. Fenlong tillage significantly affected the Alpha(α) diversity (P<0.05) and community composition of soil fungal communities. The dominant fungal phyla in the rhizosphere and bulk soils were ascomycetes, SAR hypergroup, basidiomycetes, mollispores, and chytrids, but there were differences in relative abundance. The correlation analysis and redundancy analysis results of the relative abundance of fungal communities and environmental factors showed that nitrase activity, available potassium content, urease activity, ammonium nitrogen content, nitrous oxide activity, catalase activity, soil pH value, organic matter content, and available phosphorus content had significant effects on the changes of fungal communities (P<0.05), with a more significant effect on the ascomycetes, basidiomycetes, and SAR hypergroup. Nitrase activity, available potassium content, urease activity, and ammonium nitrogen content were the main influencing factors. Cassava Fenlong tillage can change the environmental factors that had a greater impact on fungal community structure, such as nitrase activity, available potassium content, urease activity, alkali-soluble nitrogen nitrogen content, nitrous oxide enzyme activity, catalase enzyme activity, soil pH value, organic matter content, and available phosphorus content, thereby changing the Alpha (α) and Beta (β) diversity of soil fungal communities.

In response to the urban heat island problem caused by the rapid urbanization development of the Pinglu Canal Economic Belt. Based on the GEE platform, and by using spatiotemporal cube models, emerging spatiotemporal hotspot analysis, and spatial analysis methods, this article aims to study the spatiotemporal evolution characteristics and driving factors of urban heat islands in the Pinglu Canal Economic Belt from 2010 to 2020. The following conclusions can be drawn: 1) The proportion of high-temperature areas in core cities increased from 10% in 2010 to 21% in 2020, with newly added high-temperature patches transformed from medium temperature areas. 2) Emerging hotspot analysis identified 17 spatiotemporal patterns of cold hotspots, with the exception of Fangchenggang City, the other four cities showing significant new hot spots around the built-up areas, ranked in the order of Beihai City>Nanning City>Qinzhou City>Guigang City; 3) The intensity of the urban heat island had significant spatial autocorrelation, with high to high clustering overlapping highly with impermeable surfaces. The intensity of the urban heat island was highly matched with land use types in space, such as the average surface temperature of impermeable surfaces and bare land being higher, while the average surface temperature of water bodies was the lowest; 4) In 2020, the order of the impact of the three driving factors on urban heat island effect was NDVI>DEM>POP. The dual factor interaction was most strongly driven by the combination of NDVI and DEM. This study has important scientific implications for how cities in the Pinglu Canal Economic Belt can cope with the intensified heat wave risk, and contributes to exploring the comprehensive mechanism of urban heat island formation in the Pinglu Canal Economic Belt.

The Jianghuai Region, situated in a transitional zone between northern and southern China, is characterized by ecological significance encompassing biodiversity, climate regulation, agricultural productivity, and ecosystem services. Chuzhou City, a critical transportation hub with interlaced distributions of terraces, hills, and plains, was selected as a case study due to its representative Jianghuai geographical features. Based on land use data from 2000, 2010, and 2020, the spatiotemporal evolution of land use and landscape patterns in Chuzhou City during 2000-2020 was analyzed using land transfer matrices, landscape pattern indices, and gray correlation analysis. A Flus model was employed to predict land use scenarios for 2030. A comprehensive evaluation was conducted on land use evolution, landscape pattern dynamics, ecosystem service value (ESV) changes, and correlations between ESV and landscape patterns. The results indicated: 1)Cropland remained the dominant landscape type, while construction land expanded continuously. Forest area increased, whereas cropland and grassland decreased during 2000-2020. 2)Landscape patterns exhibited notable fragmentation and complexity over the study period. 3)Total ESV initially rose then declined, showing an overall downward trend. Water bodies contributed predominantly to ESV, with spatially clustered distributions. 4)At the landscape level, ESV correlated closely with the largest patch index (LPI), juxtaposition index (IJI), landscape shape index (LSI), and patch density (PD). At the patch type level, associations between landscape indices and ESV demonstrated both consistency and divergence. 5)By 2030, land use change are projected to feature outward expansion from Langya District’s urban core. Construction land dispersion is predicted to intensify, accompanied by improved ecological risk resilience, cyclical fluctuations in landscape metrics, and a 78-million-yuan increase in total ESV. This study provides theoretical references for optimizing landscape configurations and formulating land use policies in Jianghuai cities.

Ecosystem service flow is related to the healthy development of regional natural ecosystems and human well-being. From the perspective of ecosystem service supply and demand, studying ecosystem service flow in the Pinglu Canal Economic Belt is the basis for finding the construction of the ecological canal and the high-quality development of the economic belt. In this paper, the potential supply of ecosystem services is estimated by an equivalent factor, and the demand index of ecosystem services is constructed to calculate the demand for ecosystem services based on population density and average land GDP, so as to determine the supply and demand ratio of ecosystem services in Pinglu Canal Economic Belt in 2000, 2010 and 2020. The breakpoint-field strength model was used to quantify the flow and intensity of ecosystem service flows. The results show that: 1) During the past 20 years, the supply value of ecosystem services in the study area had little change (increasing from 267.6 billion yuan in 2000 to 275.3 billion yuan in 2020) and is in a spatial stable state, with the lowest value being 200 million yuan in Haicheng District in 2000 and the highest value being 22.5 billion yuan in Guiping in 2020. The districts under the jurisdiction of Nanning are the agglomeration areas with a strong demand for ecosystem services. 2) During the study period, the demand for ecosystem services increased continuously, from 40 665 people · yuan /km² in 2000 to 57 682 people · yuan /km² in 2020. The demand for ecosystem services formed three high-value clusters in the districts under the jurisdiction of Nanning, Guigang and Beihai, with the supply-demand ratio above -1 for most of them. 3) During the study period, the ecosystem flow intensity of the Pinglu Canal Economic Belt generally increased to a stable trend, and the ecosystem service flow increased from 19.813 billion yuan in 2000 to 21.4 billion yuan in 2010, and slightly decreased to 21.16 billion yuan in 2020. Guiping of Guigang is an important ecosystem service supply area. In 2000,2010 and 2020, the ecosystem service flows exported by Guiping are respectively, and the situation of ecosystem supply/deficit in each region is becoming increasingly significant.

Land use data can provide important basis for many scientific researches, and remote sensing images are widely used as data source for land use type interpretation. In order to improve the accuracy of remote sensing image classification, it is necessary to involve some other digital information. The terrain in the karst area of Guangxi is complex, together with quick expansion of plantation in recent years, which brings difficulties to the interpretation of remote sensing images. Based on the random forest algorithm, this study conducted land use interpretation of remote sensing images in Chongzuo, Guangxi. Two classification procedures were run: the first one only used remote sensing image data for classification, and altitude and slope data were added in the second procedure. Results show that the classification accuracy of using remote sensing images alone is 0.849, and after adding altitude and slope data, the overall accuracy improved to 0.961. The discrimination among land use types such as natural forests, artificial forests and farmland is largely increased, which is particularly useful in rugged landforms such as karst. This research provides a better solution for land use montoring.