High-quality Datasets Research Articles

Blending is all you need: Data-centric ensemble synthetic data

Deep generative models have gained increasing popularity, particularly in fields such as natural language processing and computer vision. Recently, efforts have been made to extend these advanced algorithms to tabular data. While generative models have shown promising results in creating synthetic data, their high computational demands and the need for careful parameter tuning present significant challenges. This study investigates whether a collective integration of refined synthetic datasets from multiple models can achieve comparable or superior performance to that of a single, large generative model. To this end, we developed a Data-Centric Ensemble Synthetic Data model, leveraging principles of ensemble learning. Our approach involved a data refinement process applied to various synthetic datasets, systematically eliminating noise and ranking, selecting, and combining them to create an augmented, high-quality synthetic dataset. This approach improved both the quantity and quality of the data. Central to this process, we introduced the Ensemble k-Nearest Neighbors with Centroid Displacement (EKCD) algorithm for noise filtering, alongside a density score for ranking and selecting data. Our experiments confirmed the effectiveness of EKCD in removing noisy synthetic samples. Additionally, the ensemble model based on the refined synthetic data substantially enhanced the performance of machine learning models, sometimes even outperforming that of real data.

Bi-directional information interaction for multi-modal 3D object detection in real-world traffic scenes

Multimodal 3D object detection methods are poorly adapted to real-world traffic scenes due to sparse distribution of point clouds and misalignment multimodal data during actual collection. Among the existing methods, they focus on high-quality open-source datasets, with performance relying on the accurate structural representation of point clouds and the precise mapping relationship between point clouds and images. To solve the above challenges, this paper proposes a multimodal feature-level fusion method based on the bi-directional interaction between image and point cloud. To overcome the sparsity issue in asynchronous multi-modal data, a point cloud densification scheme based on visual guidance and point cloud density guidance is proposed. This scheme can generate object-level virtual point clouds even when the point cloud and image are misaligned. To deal with the unalignment issue between point cloud and image, a bi-directional interaction module based on image-guided interaction with key points of point clouds and point cloud-guided interaction with image context information is proposed. It achieves effective feature fusion even when the point cloud and image are misaligned. The experiments on the VANJEE and KITTI datasets demonstrated the effectiveness of the proposed method, with average precision improvements of 6.20% and 1.54% compared to the baseline.

Mapping human footprint changes over Qingzang Plateau

Mapping human footprint (HF) is crucial for understanding the cumulative anthropogenic pressures on the environment. However, its effectiveness is often hampered by the absence of fine-scale, high-quality spatio-temporal datasets. Addressing this gap, we present the first annual HF maps for the Qingzang Plateau (QP) from 1990 to 2020, at a 100 m × 100 m resolution, boasting an accuracy exceeding 80%. Our findings reveal that the QP's human-induced modifications were relatively low, with a median HF of 3.39 in 2020, significantly lower than the global average of 7.56. The northeast and mid-south regions of the QP emerged as hotspots for anthropogenic impact. Notably, over two-thirds (68.2%) of the QP registered an HF score below 4, classifying them as largely intact. However, the QP's average HF score has escalated more swiftly than the global mean (0.0348/yr vs. 0.0186/yr), particularly in the last decade, indicating escalating human pressures. In terms of stability, montane grasslands and shrublands comprised 75.1% of moderately or highly altered biomes in 2020. In contrast, tropical and subtropical moist broadleaf forests witnessed the most substantial HF increases. Encouragingly, 26.29% of the QP experienced HF reductions over the past thirty years, highlighting potential conservation opportunities. Conversely, a significant HF increase across 36.59% of the QP underscores the pressing need to mitigate the adverse effects of intensified human activities. Our HF maps will be conducive to improve understanding the profound impacts wrought by human activities on the QP, and support the conservation planning and resource management needs, as well as decision-making related to restoration objectives.

A Multi-Scale Feature Fusion Deep Learning Network for the Extraction of Cropland Based on Landsat Data

In the face of global population growth and climate change, the protection and rational utilization of cropland are crucial for food security and ecological balance. However, the complex topography and unique ecological environment of the Qinghai-Tibet Plateau results in a lack of high-precision cropland monitoring data. Therefore, this paper constructs a high-quality cropland dataset for the YarlungZangbo-Lhasa-Nyangqv River region of the Qinghai-Tibet Plateau and proposes an MSC-ResUNet model for cropland extraction based on Landsat data. The dataset is annotated at the pixel level, comprising 61 Landsat 8 images in 2023. The MSC-ResUNet model innovatively combines multiscale features through residual connections and multiscale skip connections, effectively capturing features ranging from low-level spatial details to high-level semantic information and further enhances performance by incorporating depthwise separable convolutions as part of the feature fusion process. Experimental results indicate that MSC-ResUNet achieves superior accuracy compared to other models, with F1 scores of 0.826 and 0.856, and MCC values of 0.816 and 0.847, in regional robustness and temporal transferability tests, respectively. Performance analysis across different months and band combinations demonstrates that the model maintains high recognition accuracy during both growing and non-growing seasons, despite the study area’s complex landforms and diverse crops.

Designing a Compact Portable Electro-Cardio Device Enhanced by Machine Learning for Early Detection of Myocardial Infraction

Cardiovascular diseases remain one of the leading causes of death globally, underscoring the need for accessible, efficient, and accurate methods of heart health monitoring. Traditional heart monitoring tools, often confined to clinical settings, limit the ability to perform continuous, real-time assessments of heart function. This challenge has driven the demand for innovative, portable solutions that enable early detection and proactive management of heart conditions. The Heart Box project aims to revolutionize cardiac health monitoring through a portable, AI-integrated system that provides real-time analysis of electrocardiogram signals. Leveraging precision healthcare technology, this device enables early detection of heart conditions by continuous monitoring. Our ground breaking project encountered two primary challenges: establishing reliable hardware connections for data capture and sourcing appropriate datasets to train our Convolutional Neural Network model. Overcoming the intricacies of hardware integration demanded meticulous troubleshooting and innovative solutions to ensure seamless data acquisition. Additionally, identifying and curating high-quality datasets posed a significant hurdle, requiring exhaustive research and data pre-processing efforts. Despite these challenges, our perseverance and dedication led to the successful development of a robust Convolutional Neural Network model, revolutionizing cardiac assessment by leveraging cutting-edge technology and data-driven insights.

The Investigation of Hyperparameters Influence on Fruit and Vegetable Image Recognition Performance Using SVM

This study aims to improve the efficiency of automatic classification and quality control of fruits and vegetables through image recognition technology, to achieve efficient and accurate intelligent sorting in agricultural production, reduce labor costs and improve market competitiveness. A high-quality image dataset of 36 fruit and vegetable categories from Kaggle is used in this study. The images in the dataset have been preprocessed to ensure that the data is suitable for the classification task and sets the stage for efficient training and evaluation of the model. Logistic regression was first used as the baseline model in order to compare the performance with the Support Vector Machine (SVM) model. Subsequently, hyperparameter tuning is performed to optimize the model to achieve the best cross-validation accuracy. Next, the SVM model is trained with the selected hyperparameters, and the training time is recorded. The performance of the model was evaluated in detail by confusion matrix and classification reports, and the test set was used for final validation to ensure that the model would also perform well on unseen data. The SVM model achieves an accuracy of 96% on both validation and test sets, which is a very good performance. The hyperparameters optimized by GridSearchCV (C = 10, gamma = 0.1, kernel = rbf) effectively improve the performance of the model, verifying that reasonable hyperparameter selection is crucial to the SVM model. These results show that the model has good generalization ability and potential to be applied to real-time classification tasks.

Improving data quality by implementing an electronic medical record seems to depend on the context

Abstract Background The European healthcare system is reliant on its digital transformation to deal with challenges like rising expenditures or workforce shortage. The digital transformation is inevitably accompanied by the implementation of electronic medical records (EMR) and the ongoing adaptation of existing ones. Systematic reviews indicate that this process can have an impact on medical records data quality (DQ) [1]. At micro level, DQ is essential to ensure high quality of care. At macro level, sufficient DQ is a prerequisite for big data analyzability. In this field, completeness is a commonly analyzed dimension of DQ and empirical results indicate that completeness can improve but also deteriorate as a result of the described implementation or adoption of EMRs [2]. The aim of this work was to investigate the implementation of EMRs in comparable settings and to observe and discuss possible differences in the change in DQ. Methods Data was collected on three surgical clinics of a German academic teaching hospital before and after the implementation of an EMR. Paper-based and electronic medical records were compared. Analysis focused on ten items that were commonly documented in both record types (e.g. pain). T-tests and χ²-tests were used to compare average completeness per record type and percentage of completeness per item. Results A total of N = 659 records was analyzed. Overall, results show a significant improvement in completeness from an average of 6.0/10 items in the paper-based record type to 7.2/10 in the EMR (p<.05). At clinic level, improvement rates vary from 0.9 to 1.4. At the level of the specific items, significant deteriorations are visible in certain clinics. Conclusions Results suggest that DQs variability is context-dependent (e.g. on the clinic’s turnover rate or its patient’s length of stay). Due to the unavoidable digital transformation, a detailed context and needs analysis involving all stakeholders should be carried out before any changes are made. Key messages • The application of advanced analytics such as big data or AI training is reliant on the availability of high-quality datasets. • Electronic medical records have been demonstrated to enhance data quality, but it remains uncertain how and why improvements appear to be context-dependent.

WaterGPT: Training a Large Language Model to Become a Hydrology Expert

This paper introduces WaterGPT, a language model designed for complex multimodal tasks in hydrology. WaterGPT is applied in three main areas: (1) processing and analyzing data such as images and text in water resources, (2) supporting intelligent decision-making for hydrological tasks, and (3) enabling interdisciplinary information integration and knowledge-based Q&A. The model has achieved promising results. One core aspect of WaterGPT involves the meticulous segmentation of training data for the supervised fine-tuning phase, sourced from real-world data and annotated with high quality using both manual methods and GPT-series model annotations. These data are carefully categorized into four types: knowledge-based, task-oriented, negative samples, and multi-turn dialogues. Additionally, another key component is the development of a multi-agent framework called Water_Agent, which enables WaterGPT to intelligently invoke various tools to solve complex tasks in the field of water resources. This framework handles multimodal data, including text and images, allowing for deep understanding and analysis of complex hydrological environments. Based on this framework, WaterGPT has achieved over a 90% success rate in tasks such as object detection and waterbody extraction. For the waterbody extraction task, using Dice and mIoU metrics, WaterGPT’s performance on high-resolution images from 2013 to 2022 has remained stable, with accuracy exceeding 90%. Moreover, we have constructed a high-quality water resources evaluation dataset, EvalWater, which covers 21 categories and approximately 10,000 questions. Using this dataset, WaterGPT achieved the highest accuracy to date in the field of water resources, reaching 83.09%, which is about 17.83 points higher than GPT-4.

DMFGAN: a multifeature data augmentation method for grape leaf disease identification.

The use of deep learning techniques to identify grape leaf diseases relies on large, high-quality datasets. However, a large number of images occupy more computing resources and are prone to pattern collapse during training. In this paper, a depth-separable multifeature generative adversarial network (DMFGAN) was proposed to enhance grape leaf disease data. First, a multifeature extraction block (MFEB) based on the four-channel feature fusion strategy is designed to improve the quality of the generated image and avoid the problem of poor feature learning ability of the adversarial generation network caused by the single-channel feature extraction method. Second, a depth-based D-discriminator is designed to improve the discriminator capability and reduce the number of model parameters. Third, SeLU activation function was substituted for DCGAN activation function to overcome the problem that DCGAN activation function was not enough to fit grape leaf disease image data. Finally, an MFLoss function with a gradient penalty term is proposed to reduce the mode collapse during the training of generative adversarial networks. By comparing the visual indicators and evaluation indicators of the images generated by different models, and using the recognition network to verify the enhanced grape disease data, the results show that the method is effective in enhancing grape leaf disease data. Under the same experimental conditions, DMFGAN generates higher quality and more diverse images with fewer parameters than other generative adversarial networks. The mode breakdown times of generative adversarial networks in training process are reduced, which is more effective in practical application.

Multi-Step Temperature Prognosis of Lithium-Ion Batteries for Real Electric Vehicles Based on a Novel Bidirectional Mamba Network and Sequence Adaptive Correlation

The battery systems of electric vehicles (EVs) are directly impacted by battery temperature in terms of thermal runaway and failure. However, uncertainty about thermal runaway, dynamic conditions, and a dearth of high-quality data sets make modeling and predicting nonlinear multiscale electrochemical systems challenging. In this work, a novel Mamba network architecture called BMPTtery (Bidirectional Mamba Predictive Battery Temperature Representation) is proposed to overcome these challenges. First, a two-step hybrid model of trajectory piecewise–polynomial regression and exponentially weighted moving average is created and used to an operational dataset of EVs in order to handle the problem of noisy and incomplete time-series data. Each time series is then individually labeled to learn the representation and adaptive correlation of the multivariate series to capture battery performance variations in complex dynamic operating environments. Next, a prediction method with multiple steps based on the bidirectional Mamba is suggested. When combined with a failure diagnosis approach, this scheme can accurately detect heat failures due to excessive temperature, rapid temperature rise, and significant temperature differences. The experimental results demonstrate that the technique can accurately detect battery failures on a dataset of real operational EVs and predict the battery temperature one minute ahead of time with an MRE of 0.273%.

A multispectral camera in the VIS–NIR equipped with thermal imaging and environmental sensors for non invasive analysis in precision agriculture

Multispectral imaging (MSI) is a technique used to inspect materials properties in different domains, ranging from industrial to medical and cultural heritage and, recently, precision agriculture. Even though several MSI solutions are already commercially available, the research community is working to optimize multispectral cameras in terms of performance and cost. Systems for the agricultural field are usually very compact, combined with drones for large areas acquisition. In this work, we detail the implementation of an innovative, modular and low-cost solution of a multispectral camera based on three core camera systems in the optical (VIS–NIR) and thermal (LWIR) range. Multispectral imaging is performed with a rotating wheel of interchangeable band-pass filters. The system is also equipped with a set of environmental sensors to acquire CO2 concentration values, light intensity, temperature, and relative humidity of the surrounding environment. The technology and the measurement protocol were experimentally validated in laboratory and in open field. Advantages with respect to the available MSI cameras mounted on UAV is the integrated imaging in both the reflectance and the thermal emissive band in a close-up imaging setup and the use of environmental sensors. From the multispectral stack the spectral signature of the plants can be obtained and various vegetation indices (e.g., NDVI, NDRE) can be calculated for investigating the health status of the plant, while thermography provide additional monitoring. Close-up multispectral imaging is expected to tackle the new challenges of precision agriculture by enabling the acquisition of high-quality dataset on single plants.

A compound-target pairs dataset: differences between drugs, clinical candidates and other bioactive compounds

Providing a better understanding of what makes a compound a successful drug candidate is crucial for reducing the high attrition rates in drug discovery. Analyses of the differences between active compounds, clinical candidates and drugs require high-quality datasets. However, most datasets of drug discovery programs are not openly available. This work introduces a dataset of compound-target pairs extracted from the open-source bioactivity database ChEMBL (release 32). Compound-target pairs in the dataset either have at least one measured activity or are part of the manually curated set of known interactions in ChEMBL. Known interactions between drugs or clinical candidates and targets are specifically annotated to facilitate analyses of differences between drugs, clinical candidates, and other active compounds. In total, the dataset comprises 614,594 compound-target pairs, 5,109 (3,932) of which are known interactions between drugs (clinical candidates) and targets. The extraction is performed in an automated manner and fully reproducible. We are providing not only the datasets but also the code to rerun the analyses with other ChEMBL releases.

Enhanced stereodivergent evolution of carboxylesterase for efficient kinetic resolution of near-symmetric esters through machine learning

Carboxylesterases serve as potent biocatalysts in the enantioselective synthesis of chiral carboxylic acids and esters. However, naturally occurring carboxylesterases exhibit limited enantioselectivity, particularly toward ethyl 3-cyclohexene-1-carboxylate (CHCE, S1), due to its nearly symmetric structure. While machine learning effectively expedites directed evolution, the lack of models for predicting the enantioselectivity for carboxylesterases has hindered progress, primarily due to challenges in obtaining high-quality training datasets. In this study, we devise a high-throughput method by coupling alcohol dehydrogenase to determine the apparent enantioselectivity of the carboxylesterase AcEst1 from Acinetobacter sp. JNU9335, generating a high-quality dataset. Leveraging seven features derived from biochemical considerations, we quantitively describe the steric, hydrophobic, hydrophilic, electrostatic, hydrogen bonding, and π-π interaction effects of residues within AcEst1. A robust gradient boosting regression tree model is trained to facilitate stereodivergent evolution, resulting in the enhanced enantioselectivity of AcEst1 toward S1. Through this approach, we successfully obtain two stereocomplementary variants, DR3 and DS6, demonstrating significantly increased and reversed enantioselectivity. Notably, DR3 and DS6 exhibit utility in the enantioselective hydrolysis of various symmetric esters. Comprehensive kinetic parameter analysis, molecular dynamics simulations, and QM/MM calculations offer insights into the kinetic and thermodynamic features underlying the manipulated enantioselectivity of DR3 and DS6.

A Scoping Review on the Application of Artificial Intelligence in Treating Rare Diseases

Rare diseases affect a small percentage of the population but often present significant challenges in diagnosis, treatment, and management due to their complex and heterogeneous nature. Recent advancements in artificial intelligence (AI) have shown promising potential to address these challenges by improving early diagnosis, personalized treatment strategies, and drug discovery. This scoping review explores the current landscape of AI applications in rare disease treatment. We examine the role of AI-driven tools in enhancing diagnostic accuracy, optimizing treatment pathways, and facilitating drug repurposing for rare conditions. Additionally, we highlight the limitations and ethical concerns associated with AI implementation, such as data privacy, the need for high-quality datasets, and algorithmic transparency. The findings suggest that while AI holds transformative potential, its integration into clinical practice for rare diseases requires multidisciplinary collaboration, ongoing research, and regulatory oversight.

The Zenith Total Delay Combination of International GNSS Service Repro3 and the Analysis of Its Precision

Currently, ground-based global navigation satellite system (GNSS) techniques have become widely recognized as a reliable and effective tool for atmospheric monitoring, enabling the retrieval of zenith total delay (ZTD) and precipitable water vapor (PWV) for meteorological and climate research. The International GNSS Service analysis centers (ACs) have initiated their third reprocessing campaign, known as IGS Repro3. In this campaign, six ACs conducted a homogeneous reprocessing of the ZTD time series spanning the period from 1994 to 2022. This paper primarily focuses on ZTD products. First, the data processing strategies and station conditions of six ACs were compared and analyzed. Then, formal errors within the data were examined, followed by the implementation of quality control processes. Second, a combination method is proposed and applied to generate the final ZTD products. The resulting combined series was compared with the time series submitted by the six ACs, revealing a mean bias of 0.03 mm and a mean root mean square value of 3.02 mm. Finally, the time series submitted by the six ACs and the combined series were compared with VLBI data, radiosonde data, and ERA5 data. In comparison, the combined solution performs better than most individual analysis centers, demonstrating higher quality. Therefore, the advanced method proposed in this study and the generated high-quality dataset have considerable implications for further advancing GNSS atmospheric sensing and offer valuable insights for climate modeling and prediction.

PrescDRL: deep reinforcement learning for herbal prescription planning in treatment of chronic diseases

Treatment planning for chronic diseases is a critical task in medical artificial intelligence, particularly in traditional Chinese medicine (TCM). However, generating optimized sequential treatment strategies for patients with chronic diseases in different clinical encounters remains a challenging issue that requires further exploration. In this study, we proposed a TCM herbal prescription planning framework based on deep reinforcement learning for chronic disease treatment (PrescDRL). PrescDRL is a sequential herbal prescription optimization model that focuses on long-term effectiveness rather than achieving maximum reward at every step, thereby ensuring better patient outcomes. We constructed a high-quality benchmark dataset for sequential diagnosis and treatment of diabetes and evaluated PrescDRL against this benchmark. Our results showed that PrescDRL achieved a higher curative effect, with the single-step reward improving by 117% and 153% compared to doctors. Furthermore, PrescDRL outperformed the benchmark in prescription prediction, with precision improving by 40.5% and recall improving by 63%. Overall, our study demonstrates the potential of using artificial intelligence to improve clinical intelligent diagnosis and treatment in TCM.

TSAE-UNet: A Novel Network for Multi-Scene and Multi-Temporal Water Body Detection Based on Spatiotemporal Feature Extraction

The application of remote sensing technology in water body detection has become increasingly widespread, offering significant value for environmental monitoring, hydrological research, and disaster early warning. However, the existing methods face challenges in multi-scene and multi-temporal water body detection, including the diverse variations in water body shapes and sizes that complicate detection; the complexity of land cover types, which easily leads to false positives and missed detections; the high cost of acquiring high-resolution images, limiting long-term applications; and the lack of effective handling of multi-temporal data, making it difficult to capture the dynamic changes in water bodies. To address these challenges, this study proposes a novel network for multi-scene and multi-temporal water body detection based on spatiotemporal feature extraction, named TSAE-UNet. TSAE-UNet integrates convolutional neural networks (CNN), depthwise separable convolutions, ConvLSTM, and attention mechanisms, significantly improving the accuracy and robustness of water body detection by capturing multi-scale features and establishing long-term dependencies. The Otsu method was employed to quickly process Sentinel-1A and Sentinel-2 images, generating a high-quality training dataset. In the first experiment, five rectangular areas of approximately 37.5 km2 each were selected to validate the water body detection performance of the TSAE-UNet model across different scenes. The second experiment focused on Jining City, Shandong Province, China, analyzing the monthly water body changes from 2020 to 2022 and the quarterly changes in 2022. The experimental results demonstrate that TSAE-UNet excels in multi-scene and long-term water body detection, achieving a precision of 0.989, a recall of 0.983, an F1 score of 0.986, and an IoU of 0.974, significantly outperforming FCN, PSPNet, DeepLabV3+, ADCNN, and MECNet.

Implementation of an Automated System Using Machine Learning Models to Accelerate the Process of In Silico Identification of Small Molecules As Drug Candidates.

Drugs are commonly utilized to diagnose, cure, or prevent the occurrence of diseases, as well as to restore, alter, or change organic functions. Drug discovery is a time-consuming, costly, difficult, and inefficient process that yields very few medicinal breakthroughs. Drug research and design involves the capturing of structural information for biological targets and small molecules as well as various in silico methods, such as molecular docking and molecular dynamic simulation. This article proposes the idea of expediting computational drug development through a collaboration of scientists and universities, similar to the Human Genome Project using machine learning (ML) strategies. We envision an automated system where readily available or novel small molecules (chemical or plant-derived), as well as their biological targets, are uploaded to an online database, which is constantly updated. For this system to function, machine learning strategies have to be implemented, and high-quality datasets and high quality assurance of the ML models will be required. ML can be applied to all computational drug discovery fields, including hit discovery, target validation, lead optimization, drug repurposing, and data mining of small compounds and biomolecule structures. Researchers from various disciplines, such as bioengineers, bioinformaticians, geneticists, chemists, computer and software engineers, and pharmacists, are expected to collaborate to establish a solid workflow and certain parameters as well as constraints for a successful outcome. This automated system may help speed up the drug discovery process while also lowering the number of unsuccessful drug candidates. Additionally, this system will decrease the workload, especially in computational studies, and expedite the process of drug design. As a result, a drug may be manufactured in a relatively short time.

A LSTM algorithm-driven deep learning approach to estimating repair and maintenance costs of apartment buildings

PurposeThis study proposes a deep learning algorithm-based model to predict the repair and maintenance costs of apartment buildings, by collecting repair and maintenance cost data that were incurred in an actual apartment complex. More specifically, a long short-term memory (LSTM) algorithm was adopted to develop the prediction model, while the robustness of the model was verified by recurrent neural networks (RNN) and gated recurrent units (GRU) models.Design/methodology/approachRepair and maintenance cost data incurred in actual apartment complexes is collected, along with various input variables, such as repair and maintenance timing (calendar year), usage types, building ages, temperature, precipitation, wind speed, humidity and solar radiation. Then, the LSTM algorithm is employed to predict the costs, while two other learning models (RNN and GRU) are taught to validate the robustness of the LSTM model based on R-squared values, mean absolute errors and root mean square errors.FindingsThe LSTM model’s learning is more accurate and reliable to predict repair and maintenance costs of apartment complex, compared to the RNN and GRU models’ learning performance. The proposed model provides a valuable tool that can contribute to mitigating financial management risks and reducing losses in forthcoming apartment construction projects.Originality/valueGathering a real-world high-quality data set of apartment’s repair and maintenance costs, this study provides a highly reliable prediction model that can respond to various scenarios to help apartment complex managers plan resources more efficiently, and manage the budget required for repair and maintenance more effectively.

Accelerated Data Engine: A faster dataset construction workflow for computer vision applications in commercial livestock farms

Large-scale, high-quality dataset was the foundation of developing advanced artificial intelligence applications. However, creating such a benchmark dataset in a professional field, such as precision management of animals, was always a challenge because of the costly and labor-intensive process of annotation and review. This study introduced a novel workflow named Accelerated Data Engine (ADE), designed to efficiently produce representative and high-quality computer vision datasets from raw animal surveillance footage. By incorporating referring and grounding models (R&G models) as auto-annotators, along with a distillation mechanism for dataset-auditors, ADE significantly speeded up the dataset construction process. The new workflow received natural language inputs as referrals to identify animal instances, delineated their body shapes, and then refined the auto-annotated data through a selection process. To demonstrate the efficacy of ADE, three 30-minute surveillance video samples featuring pigs, sheep, and cattle were discussed in this study. The results indicated the R&G models effectively annotated animals across various farms, while distillation mechanisms could identify various detection errors, balance the data representations, refine annotations, and verify the data quality. Two high-quality cattle datasets (6.5 k and 486 frames), including 26 k and 2.5 k cattle instances, were generated through the ADE workflow from 24-hour surveillance videos on a commercial cattle farm and made publicly available. The proposed dataset has achievable performance between 74.6 %∼84.1 %. The ADE workflow saved 78.4 % of manual work compared to the traditional dataset construction workflow (approximately 141 h). This pioneering approach empowered the fast creation of benchmark animal datasets and would enhance computer vision applications in the livestock production industry in the future.