Detection Algorithm Research Articles

Comparison of deep learning algorithms for site detection of false data injection attacks in smart grids

Comparison of deep learning algorithms for site detection of false data injection attacks in smart grids

Acquisition of Data on Kinematic Responses to Unpredictable Gait Perturbations: Collection and Quality Assurance of Data for Use in Machine Learning Algorithms for (Near-)Fall Detection.

Slip, trip, and fall (STF) accidents cause high rates of absence from work in many companies. During the 2022 reporting period, the German Social Accident Insurance recorded 165,420 STF accidents, of which 12 were fatal and 2485 led to disability pensions. Particularly in the traffic, transport and logistics sector, STF accidents are the most frequently reported occupational accidents. Therefore, an accurate detection of near-falls is critical to improve worker safety. Efficient detection algorithms are essential for this, but their performance heavily depends on large, well-curated datasets. However, there are drawbacks to current datasets, including small sample sizes, an emphasis on older demographics, and a reliance on simulated rather than real data. In this paper we report the collection of a standardised kinematic STF dataset from real-world STF events affecting parcel delivery workers and steelworkers. We further discuss the use of the data to evaluate dynamic stability control during locomotion for machine learning and build a standardised database. We present the data collection, discuss the classification of the data, present the totality of the data statistically, and compare it with existing databases. A significant research gap is the limited number of participants and focus on older populations in previous studies, as well as the reliance on simulated rather than real-world data. Our study addresses these gaps by providing a larger dataset of real-world STF events from a working population with physically demanding jobs. The population studied included 110 participants, consisting of 55 parcel delivery drivers and 55 steelworkers, both male and female, aged between 19 and 63 years. This diverse participant base allows for a more comprehensive understanding of STF incidents in different working environments.

Information extraction of seasonal dissolved oxygen in urban water bodies based on machine learning using sentinel-2 imagery: An open access application in Baiyangdian Lake

Water bodies are crucial components of urban ecology. The development of rapid and timely water-quality assessment tools using easily measured variables is essential for the health management of urban water bodies. In this study, we focused on the dissolved oxygen (DO) of Baiyangdian Lake using 251 sets of empirically measured water quality data and corresponding Sentinel-2 satellite images. Nine machine learning algorithms were then used to develop a rapid detection algorithm for the spatial distribution of the DO concentration in Baiyangdian Lake. This study successfully applied these methods to invert the DO concentration in Baiyangdian Lake during spring, summer, and autumn. The results indicated that extra tree regression (ETR) provided the most accurate and stable results for inverting the DO concentration among the nine machine learning methods. In contrast, AdaBoost regression (ABR), Bayesian ridge regression (BRR), and support vector machines (SVM) exhibit relatively poor regression performance and lack sensitivity to DO concentrations. Moreover, the DO concentration in Baiyangdian Lake ranged from approximately 0 to 12 mg/L, with notable spatiotemporal variations. The highest overall DO concentration was observed in the spring, particularly in the southern region. The DO concentration significantly decreased during summer compared to that in spring, with higher values in the southwestern area and lower values in the northern region. The DO concentration reached its lowest value in autumn, with slightly higher values in the southern region. This study focused on the estimation and inversion of DO concentrations in the water bodies of Baiyangdian Lake. By introducing and comparing the performances of commonly used machine learning models, a rapid estimation of the DO concentration was achieved, thereby overcoming the limitations of traditional water quality monitoring methods in DO inversion. It not only intuitively explained the temporal and spatial variation patterns of DO concentration but also laid a foundation for further in-depth exploration of the interactions between DO and other water quality parameters.

Machine Learning for Patient-Based Real-Time Quality Control (PBRTQC), Analytical and Preanalytical Error Detection in Clinical Laboratory.

The rapidly evolving field of machine learning (ML), along with artificial intelligence in a broad sense, is revolutionising many areas of healthcare, including laboratory medicine. The amalgamation of the fields of ML and patient-based real-time quality control (PBRTQC) processes could improve the traditional PBRTQC and error detection algorithms in the laboratory. This narrative review discusses published studies on using ML for the detection of systematic errors, non-systematic errors, and combinations of different types of errors in clinical laboratories. The studies discussed used ML for detecting bias, the requirement for re-calibration, samples contaminated with intravenous fluid or EDTA, delayed sample analysis, wrong-blood-in-tube errors, interference or a combination of different types of errors, by comparing the performance of ML models with human validators or traditional PBRTQC algorithms. Advantages, limitations, the creation of standardised ML models, ethical and regulatory aspects and potential future developments have also been discussed in brief.

Green fruit detection methods: Innovative application of camouflage object detection and multilevel feature mining

Accurate detection of object fruits is essential for optimizing picking efficiency and predicting fruit yields. However, detecting early-stage ripening or green fruits, especially in complex fields, poses significant challenge due to their similarity to green leaves. This study introduces the TEAVit model, a novel camouflage object detection network specifically tailored for identifying green tomatoes in intricate agricultural environments. TEAVit incorporates a texture-edge-awareness module (TEAM) to enhance the extraction ability of texture feature by combining high-level and low-level features, an edge-guided feature module (EFM) to address background complexities and occlusions, and a context-aggregation module (CAM) to leverage contextual semantics. Experimental validation results demonstrate that the S-measure, E-measure, and F-measure performance metrics all exceed 90% on the tomato dataset, accompanied by a mean absolute error of 0.0245. These findings underpinned the effectiveness of the proposed green fruit camouflage object detection algorithm in offering new insights for agricultural target localization.

Diagnostic Accuracy of the Persyst Automated Seizure Detector in the Neonatal Population.

Neonatal seizures are diagnostically challenging and predominantly electrographic-only. Multichannel video continuous electroencephalography (cEEG) is the gold standard investigation, however, out-of-hours access to neurophysiology support can be limited. Automated seizure detection algorithms (SDAs) are designed to detect changes in EEG data, translated into user-friendly seizure probability trends. The aim of this study was to evaluate the diagnostic accuracy of the Persyst neonatal SDA in an intensive care setting. Single-centre retrospective service evaluation study in neonates undergoing cEEG during intensive care admission to Great Ormond Street Hospital (GOSH) between May 2019 and December 2022. Neonates with <44 weeks corrected gestational age, who had a cEEG recording duration >60 minutes, whilst inpatient in intensive care, were included in the study. One-hour cEEG clips were created for all cases (seizures detected) and controls (seizure-free) and analysed by the Persyst neonatal SDA. Expert neurophysiology reports of the cEEG recordings were used as the gold standard for diagnostic comparison. A receiver operating characteristic (ROC) curve was created using the highest seizure probability in each recording. Optimal seizure probability thresholds for sensitivity and specificity were identified. Eligibility screening produced 49 cases, and 49 seizure-free controls. Seizure prevalence within those patients eligible for the study, was approximately 19% with 35% mortality. The most common case seizure aetiology was hypoxic ischaemic injury (35%) followed by inborn errors of metabolism (18%). The ROC area under the curve was 0.94 with optimal probability thresholds 0.4 and 0.6. Applying a threshold of 0.6, produced 80% sensitivity and 98% specificity. The Persyst neonatal SDA demonstrates high diagnostic accuracy in identifying neonatal seizures; comparable to the accuracy of the standard Persyst SDA in adult populations, other neonatal SDAs, and amplitude integrated EEG (aEEG). Overdiagnosis of seizures is a risk, particularly from cEEG recording artefact. To fully examine its clinical utility, further investigation of the Persyst neonatal SDA's accuracy is required, as well as confirming the optimal seizure probability thresholds in a larger patient cohort.

A global analysis of ice phenology for 3702 lakes and 1028 reservoirs across the Northern Hemisphere using Sentinel-2 imagery

As existing global lake ice studies have predominantly focused on medium to large lakes, and reservoir ice studies have been limited to regional scales, very few studies of ice phenology have combined both lakes and reservoirs of different sizes. This study aims to characterize the freeze-up and break-up dates of 3702 lakes and 1028 reservoirs from 1 to 31,000 km2 across the Northern Hemisphere, and to analyze spatial patterns and relationships between ice phenological dates and driving factors. The freeze-up and break-up dates of these water bodies were retrieved from Sentinel-2 imagery using an ice detection algorithm through the Google Earth Engine platform from 2019 to 2023. The algorithm was verified by comparing phenology dates with an independent database based on observations from passive microwave sensors, with a mean absolute error of 18 days for both freeze-up and break-up dates. This newly established ice phenology database along with various geographic, morphometric, and climatic characteristics of the water bodies, was used to develop a random forest model for predicting ice phenology dates. While the predictive model performance is at a fair level (mean absolute error of 12 days for both freeze-up and break-up), challenges were encountered in certain high-elevation areas where cloudy conditions as well as black ice resulted in delayed freeze-up dates. Among the variables included in the random forest model, latitude and accumulation of freezing degree days were identified as the main drivers of ice phenology dates. Despite the challenges of applying a single, straightforward method on a global scale, this study has allowed the creation of a vast and comprehensive database of lake and reservoir freeze-up and break-up dates that can be used by the community to further analyze ice patterns.

Immunohistochemistry and RT-PCR as ancillary tools in the diagnosis of acquired oral syphilis.

To investigate Treponema pallidum detection using immunohistochemistry (IHC) and reverse transcription-polymerase chain reaction (RT-PCR) assays in acquired oral syphilis (AOS). Thirty-seven paraffin-embedded tissue specimens of AOS (32 secondary and five primary) were analyzed, integrating double-positive serological results with clinicodemographic and histopathological data. T. pallidum presence was semiquantitatively assessed by IHC, while RT-PCR targeted T. pallidum DNA. Sensitivity, specificity, and the area under the curve (AUC) were calculated with 95% confidence intervals (CI). The study included mostly females (62.2%) with a mean age of 27.1 years. T. pallidum was detected in all samples by IHC, predominantly in the epithelium across all layers (43.2%). RT-PCR identified T. pallidum DNA in 32 cases, with negative results observed in cases of secondary AOS. The AUC for IHC versus disease stage was 62.5% (95% CI: 45.1-77.8), and for RT-PCR, it was 57.8% (95% CI: 40.5-73.8). The AUC comparing IHC to RT-PCR was 83.8% (95% CI: 67.9-93.8). This study represents the first attempt to evaluate the proposed direct detection algorithm for AOS. IHC and RT-PCR serve as ancillary tools for detecting T. pallidum in both primary and secondary stages of AOS.

LUD-YOLO: A novel lightweight object detection network for unmanned aerial vehicle

Autonomous execution of tasks by unmanned aerial vehicles (UAVs) relies heavily on object detection. However, object detection in most images presents challenges such as complex backgrounds, small targets, and obstructions. Additionally, the limited computing speed and memory of the UAV processor affects the accuracy of conventional object detection algorithms. This paper proposes LUD-You Only Look Once (YOLO), a small and lightweight object detection algorithm for UAVs based on YOLOv8. The proposed algorithm introduces a new multiscale feature fusion mode that solves the degradation in feature propagation and interaction through the introduction of upsampling in the feature pyramid network and the progressive feature pyramid network. The application of the dynamic sparse attention mechanism in the Cf2 module achieves flexible computing allocation and content awareness. Furthermore, the proposed model is optimized to be sparse and lightweight, making it possible to deploy on UAV edge devices. Finally, the effectiveness and superiority of LUD-YOLO were verified on the VisDrone2019 and UAVDT datasets. The results of ablation and comparison experiments show that compared with the original algorithm, LUDY-N and LUDY-S have shown excellent performance in various evaluation indexes, indicating that the proposed improvement strategies make the model have better robustness and generalization. Moreover, compared with multiple other popular competitors, the proposed improvement strategies enable LUD-YOLO to have the best overall performance, providing an effective solution for UAVs object detection while balancing model size and detection accuracy.

Multi-Channel Multi-Scale Convolution Attention Variational Autoencoder (MCA-VAE): An Interpretable Anomaly Detection Algorithm Based on Variational Autoencoder.

With the rapid development of industry, the risks factories face are increasing. Therefore, the anomaly detection algorithms deployed in factories need to have high accuracy, and they need to be able to promptly discover and locate the specific equipment causing the anomaly to restore the regular operation of the abnormal equipment. However, the neural network models currently deployed in factories cannot effectively capture both temporal features within dimensions and relationship features between dimensions; some algorithms that consider both types of features lack interpretability. Therefore, we propose a high-precision, interpretable anomaly detection algorithm based on variational autoencoder (VAE). We use a multi-scale local weight-sharing convolutional neural network structure to fully extract the temporal features within each dimension of the multi-dimensional time series. Then, we model the features from various aspects through multiple attention heads, extracting the relationship features between dimensions. We map the attention output results to the latent space distribution of the VAE and propose an optimization method to improve the reconstruction performance of the VAE, detecting anomalies through reconstruction errors. Regarding anomaly interpretability, we utilize the VAE probability distribution characteristics, decompose the obtained joint probability density into conditional probabilities on each dimension, and calculate the anomaly score, which provides helpful value for technicians. Experimental results show that our algorithm performed best in terms of F1 score and AUC value. The AUC value for anomaly detection is 0.982, and the F1 score is 0.905, which is 4% higher than the best-performing baseline algorithm, Transformer with a Discriminator for Anomaly Detection (TDAD). It also provides accurate anomaly interpretation capability.

Improved lightweight infrared road target detection method based on YOLOv8

Infrared-based road scene object detection algorithms often face issues with excessive parameters and computational demands, making them incompatible with edge devices having constrained computational capabilities. This paper introduces an enhanced lightweight infrared-based road object detection algorithm based on YOLOv8n. Firstly, a streamlined network architecture is devised by merging YOLOv8n’s C2f module with PConv, creating a lighter module and reducing the neural network’s downsampling rate of infrared images. This strategy reduces redundant computations and memory access, preventing the loss of fine details in infrared images caused by deep convolutional neural networks. Additionally, the model’s accuracy in detecting infrared targets is significantly enhanced through the integration of the coordinate attention mechanism. Finally, replacing CIoU with Wise-IoU for bounding box regression in YOLOv8n accelerates the model’s convergence. Empirical findings indicate that in contrast to the YOLOv8n algorithm, the optimized model showcases a 34.17 % reduction in model size, a 40.35 % decrease in parameters, and a 4.8 % increase in average detection accuracy. This enhanced algorithm not only achieves a lightweight profile but also delivers superior performance on embedded edge devices.

Rapid detection technology of sewing thread based on DeblurGAN-BSV3 defuzzification algorithm and ST-FPN detection algorithm

Stitch defect detection is an important means to ensure the quality of stitched fabrics. In order to realize the intellectualization of the sewing process, a rapid defect detection method based on machine vision was proposed. In order to solve the problem of motion blur caused by sewing speed, an image DeblurGAN-BSV3 algorithm is proposed. Firstly, different attention fusion mechanisms are introduced to achieve accurate extraction of image features. Secondly, the depth-separable convolution of the trunk network is replaced by blueprint convolution to ensure the processing speed of the detection algorithm and the clarity of the output image. In order to solve the difficulty of rapid detection of pin defects, a student–teacher feature pyramid matching method was proposed. Firstly, the abnormal region interpretation technology was used to explicitly visualize the defect area of the identified abnormal instances, and then the abnormal detection score was performed on the detected images to make a quality judgment. The experimental results show that the combination of the above two methods has a good ability for rapid defect detection. Rapid detection of common pin defects is realized.

Concrete Surface Crack Detection Algorithm Based on Improved YOLOv8.

Concrete surface crack detection is a critical research area for ensuring the safety of infrastructure, such as bridges, tunnels and nuclear power plants, and facilitating timely structural damage repair. Addressing issues in existing methods, such as high cost, lengthy processing times, low efficiency, poor effectiveness and difficulty in application on mobile terminals, this paper proposes an improved lightweight concrete surface crack detection algorithm, YOLOv8-Crack Detection (YOLOv8-CD), based on an improved YOLOv8. The algorithm integrates the strengths of visual attention networks (VANs) and Large Convolutional Attention (LCA) modules, introducing a Large Separable Kernel Attention (LSKA) module for extracting concrete surface crack and local feature information, adapted for features such as fracture susceptibility, large spans and slender shapes, thereby effectively emphasizing crack shapes. The Ghost module in the YOLOv8 backbone efficiently extracts essential information from original features at a minimal cost, enhancing feature extraction capability. Moreover, replacing the original convolution structure with GSConv in the neck network and employing the VoV-GSCSP module adapted for the YOLOv8 framework reduces floating-point operations during feature channel fusion, thereby lowering computational complexity whilst maintaining model accuracy. Experimental results on the RDD2022 and Wall Crack datasets demonstrate the improved algorithm increases in mAP50 by 15.2% and 12.3%, respectively, and in mAP50-95 by 22.7% and 17.2%, respectively, whilst achieving a reduced model computational load of only 7.9 × 109, a decrease of 3.6%. The algorithm achieves a detection speed of 88 FPS, enabling real-time and accurate detection of concrete surface crack targets. Comparison with other mainstream object detection algorithms validates the effectiveness and superiority of the proposed approach.

Anomaly detection for atomic clocks using unsupervised machine learning algorithms

Atomic clocks are the most accurate timekeeping instruments in the world and play a crucial role in numerous applications, notably in satellite navigation. Nonetheless, the accuracy of these clock signals is subject to anomalies, highlighting the importance of anomaly characterization and detection. In our study, we identify three types of anomalies in the phase data of atomic clocks—outliers, phase jumps and inflection points, and also frequency jumps in the frequency data, which corresponds to the inflection points. The novelty of our work lies in the application of machine learning methods for atomic clock anomaly detection (AD), as opposed to traditional algorithms. We introduce a variety of unsupervised AD and change point detection (CPD) algorithms from the realms of statistics and machine learning, propose a customized thresholding function for AD, and evaluate AD and CPD algorithms on both synthetic and real-world atomic clock data. Our experiments reveal that the subsequence local outlier factor algorithm achieves excellent performance in detecting outliers and phase jumps, but struggles with inflection points. Crucially, we found that the application of a CPD algorithm specifically for detecting the frequency jumps in frequency data substantially improves performance. Our study adopts unsupervised machine learning methods for AD provides robust detection strategies, thereby contributing an expanded understanding of the use of machine learning for clock data monitoring.

MFNet: Multi-scale feature enhancement networks for wheat head detection and counting in complex scene

Developing a high-precision wheat head detection algorithm is challenging due to the dense distribution and diverse sizes of wheat head in the field, as well as serious coverage from weeds. In this work, we propose multi-scale feature enhancement networks for wheat head detection and counting in complex scene (MFNet). Firstly, we introduce a deformable spatial attention mechanism (DSAM) and embed it in the backbone network to enhance the extraction of wheat head features while suppressing irrelevant features, effectively improving detection of wheat head in occluded environments. Secondly, we design a multi-scale receptive field feature fusion (MRFF) module in combination with an improved light-weight feature pyramid module to achieve more accurate detection of wheat head of different sizes and improve localization accuracy simultaneously. Additionally, the modified detection head with deformable convolution is able to adapt to different shapes of wheat head features and accurately predict the bounding boxes. Our method achieves 94.2% AP@50 at a speed of 30 FPS on the GWHD dataset, and to verify the generalization of the proposed method, we constructed a dense wheat head detection (DWHD) dataset with annotations, conducted experiments on the DWHD and SPIKE datasets and compared them with state-of-the-art algorithms. The experimental results show that the proposed method outperformed most of the existing methods, which further proved the superiority and robustness of the proposed method, and demonstrated that our model possesses excellent adaptability, enabling it to flexibly cope with scenarios of significant scale differences and severe occlusion in complex field environments for wheat head, and provide a technological reference for monitoring the wheat phenotype.

A hand-crafted φ-OTDR event recognition method based on space-temporal graph and morphological object detection

In the long-range pipeline monitoring application of Fiber Optic Distributed Vibration Sensor (FODVS), instantaneous vibration event with high energy generates a pair of invert-V signature in the space-temporal graph, providing clear characteristic for event recognition. This paper proposed an event recognition method utilizing FODVS and hand-crafted morphological object detection algorithm. Phase-sensitive Optical Time Domain Reflectometer (φ-OTDR) was employed as the distributed sensor to collect the real field space-temporal graph data containing invert-V signature. Bandpass filter and Gaussian blur were utilized to suppress random noise and enhance the invert-V signature, then a hand-crafted morphological object detection method was proposed to recognize the invert-V pattern. The proposed method was proved to be able to accurately capture the invert-V signature in the space-temporal graph. This paper provided a concept-proof preliminary demonstration on the hand-crafted FODVS event recognition scheme for pursuing model with good interpretability and strong adaptability to varied environments.

Seismocardiography and echocardiography: the correlation in the systolic complex

Aim. This study aimed to investigate the correlation between seismocardiographic and echocardiographic systolic variables and whether a decrease in preload could be detected by the seismocardiography (SCG). Methods. This study included a total of 34 subjects. SCG and electrocardiography were recorded simultaneously followed by echocardiography (echo) in both supine and 30◦ head-up tilted position. The SCG signals was segmented into individual heartbeats and systolic fiducial points were defined using a detection algorithm. Statistical analysis included correlation coefficient calculations and paired sample tests. Results. SCG was able to measure a decrease in preload by almost all of the examined systolic SCG variables. It was possible to correlate certain echo variables to SCG time intervals, amplitudes, and peak to peak intervals. Also, changes between supineand tilted position of some SCG variables were possible to correlate to changes in echo variables. LVET, IVCT, S’, strain, SR, SV, and LVEF were significantly correlated to relevant SCG variables. Conclusion. This study showed a moderate correlation, between systolic echo and systolic SCG variables. Additionally, systolic SCG variables were able to detect a decrease in preload.

Research on traffic sign detection algorithm based on improved SSD in complex environments

Abstract In complex traffic sign environments, detection challenges include uneven size distribution, insufficient valid information, and difficulties in identifying targets under resource constraints, leading to missed and false detections. This study proposes an enhanced lightweight traffic sign detection algorithm based on SSD (Single Shot MultiBox Detector). By replacing the original backbone network with MobileNetV2, the model is streamlined to have fewer parameters, which improves generalization in complex environments. This modification significantly boosts the recall rate, achieving a better balance between precision and recall. By introducing the FPN(Feature Pyramid Network) combined with the CBAM(Convolutional Block Attention Module) attention mechanism, the detailed and semantic information between deep and shallow layers is fully integrated, reducing the loss of feature information, thus enhancing the strengthening of key information of traffic signs and the adaptability to different scales of traffic signs. Finally, by integrating the cross-attention mechanism, the algorithm's anti-interference ability in complex environments is improved, and the positioning accuracy of traffic signs is enhanced by capturing the dependency between different positions. Through ablation experiments and comparative experiments on a public traffic sign dataset, our improved SSD algorithm achieved an mAP of 89.97%. Compared with the original algorithm, the mAP increased by 12.41%, the recall rate increased by 18.38%, and the sum of precision and recall F1 increased by 14.6%. These improvements significantly enhance the performance of traffic sign detection in complex environments, thereby meeting the performance requirements of traffic sign detection.&amp;#xD;

YOLO-TSL: A lightweight target detection algorithm for UAV infrared images based on Triplet attention and Slim-neck

When performed from an Unmanned Aerial Vehicle (UAV) perspective, infrared target detection often suffers from low accuracy, excessive model parameters, and slow processing speed. To address these challenges, this paper proposes a lightweight infrared target detection algorithm, named YOLO-TSL, which is based on an improved version of YOLOv8n. Building on an in-depth analysis of UAV infrared image characteristics, this paper introduces Triplet Attention to the model backbone to enhance target detection accuracy and effectively suppress background interference. Moreover, YOLO-TSL incorporates the Slim-Neck architecture, which features GSConv and GSbottleneck in the neck structure and employs a one-time aggregation method to design the cross-level partial network known as the VoV-GSCSP module. This architecture significantly reduces the model’s computational complexity while maintaining high detection accuracy. Furthermore, this paper introduces an innovative inner-MPDIoU loss function that optimizes IoU loss computation by enhancing bounding box similarity and adaptive auxiliary bounding box scale adjustments, based on both inner-IoU and MDPIoU. Following experimental validation, YOLO-TSL demonstrate significant improvements over YOLOV8n, including a 3.9% increase in mAP50 and a 3% increase in recall. Additionally, YOLO-TSL has 15.7% fewer parameters, requires 11% less computation, and offers 26% faster inference. Comparative experiments on the FLIR dataset reveal that the algorithm not only outperforms YOLOv8n by 1.4% in mAP50, but also offers significant advantages over other algorithms in terms of parameter reduction and computational efficiency, thus demonstrating YOLO-TSL’s superiority in accuracy, efficiency, and speed.

An unmanned aerial system benchmark object detection dataset for deep learning in outfall surveys

ABSTRACT Deep-learning-based object detection in UAS imagery is crucial for outfall surveys and basin environmental protection. Comprehensive UAS image datasets serve as a foundation for creating deep-learning-based outfall-detection models and evaluating outfall-detection algorithms. However, existing remote sensing deep learning datasets lack specific outfall data. This study introduces a benchmark UAS image dataset of outfalls, categorized into three main and seven subcategories. Over 10,000 labeled images were collected from UAS images with a 10 cm resolution for the Yangtze River, Yellow River, and other basins from 2019 to 2022, covering nearly all types of outfalls in China. Each sample was matched with digital surface model (DSM) data generated through photographometry or the DSM transfer method. Evaluation with seven widely used deep learning-based object detection algorithms demonstrated the dataset's viability, achieving an average precision ( A P 50 ) of 64.6, surpassing performance on Microsoft Common Objects in Context. Further experiments indicated that DSMs attached to this dataset could benefit from geo-deep-learning-based object detection algorithms for outfalls, achieving an A P 50 exceeding 70. This study presents a UAS benchmark image dataset for outfall surveys, potentially advancing the application of deep learning in environmental protection.