High-quality Datasets Research Articles

DETERMINING STUDENT'S ONLINE ACADEMIC PERFORMANCE USING MACHINE LEARNING TECHNIQUES

Predicting student's academic performance during online learning has been considered a major task during the pandemic period. During the online mode of learning, academic activities have been affected in such a way that the management of educational institutions has planned to design support systems for predicting the student's performance to reduce the dropout ratio of the students and bring improvement in academic activities. During COVID-19, the main challenge is maintaining student's grades by predicting their academic performance using different techniques such as Education Data Mining and Learning Analytics. Different features have been identified related to the teaching mechanisms in online learning, which have a great impact on the improvement of academic performance. A high-quality dataset helps us to generate productive results, which in turn helps us to make effective decisions for promoting high-quality education. In this research, five prediction models for predicting academic performance have been proposed by collecting an imbalanced dataset of 350 students from the same computer science domain. After applying pre-processing techniques for cleaning the data, machine learning models have been applied, including K-Nearest Neighbor Classifier, Decision Tree, Random Forest, Support Vector Classifier, and Gaussian Naive Bayes. Results have been predicted for an imbalanced and balanced dataset after feature selection. Support Vector classifier has produced the best results in a balanced dataset with selected features by giving an accuracy of 96.89%.

EnNet: Enhanced Interactive Information Network with Zero-Order Optimization.

Interactive image segmentation extremely accelerates the generation of high-quality annotation image datasets, which are the pillars of the applications of deep learning. However, these methods suffer from the insignificance of interaction information and excessively high optimization costs, resulting in unexpected segmentation outcomes and increased computational burden. To address these issues, this paper focuses on interactive information mining from the network architecture and optimization procedure. In terms of network architecture, the issue mentioned above arises from two perspectives: the less representative feature of interactive regions in each layer and the interactive information weakened by the network hierarchy structure. Therefore, the paper proposes a network called EnNet. The network addresses the two aforementioned issues by employing attention mechanisms to integrate user interaction information across the entire image and incorporating interaction information twice in a design that progresses from coarse to fine. In terms of optimization, this paper proposes a method of using zero-order optimization during the first four iterations of training. This approach can reduce computational overhead with only a minimal reduction in accuracy. The experimental results on GrabCut, Berkeley, DAVIS, and SBD datasets validate the effectiveness of the proposed method, with our approach achieving an average NOC@90 that surpasses RITM by 0.35.

Spatial interpolation of global DEM using federated deep learning

The digital elevation model (DEM) provides important data support for 3D terrain modeling. However, due to the complex and changeable terrain in the real world and the high cost of field measurement, it is extremely difficult to obtain continuous and high-density elevation data directly. Therefore, it is necessary to rely on spatial interpolation technology to restore the DEM overall picture in the original sampling area. The traditional spatial interpolation method usually has the characteristics of low model complexity and high computational cost, which leads to low real-time performance and low precision of the interpolation process. The interpolation operation based on DEM data can be considered as a special image generation process where the input is a DEM image with missing values and the output is a complete DEM image. At present, a large number of studies have proved that deep learning methods are very effective in image generation tasks. However, the training of deep learning models requires the support of a large number of high-quality data sets. DEM data in various countries, especially in key regions, are usually restricted by privacy protection regulations and cannot be disclosed. The emergence of Federated Learning (FL) provides a new solution, which supports local training on multiple end nodes, without sending local data to a remote center server for centralized training, effectively protecting data privacy. In this study, we propose a DEM interpolation model based on FL and multiScale U-Net. The experimental results show that compared with the traditional method, this model has faster processing speed and lower interpolation precision. At the same time, this research result provides a new way for efficient and secure use of terrain information, especially in those application scenarios that have strict requirements for DEM data privacy and security.

Machine Learning-Based Prediction of Clinical Outcomes in Microsurgical Clipping Treatments of Cerebral Aneurysms.

Background: This study investigates the application of Machine Learning techniques to predict clinical outcomes in microsurgical clipping treatments of cerebral aneurysms, aiming to enhance healthcare processes through informed clinical decision making. Methods: Relying on a dataset of 344 patients' preoperative characteristics, various ML classifiers were trained to predict outcomes measured by the Glasgow Outcome Scale (GOS). The study's results were reported through the means of ROC-AUC scores for outcome prediction and the identification of key predictors using SHAP analysis. Results: The trained models achieved ROC-AUC scores of 0.72 ± 0.03 for specific GOS outcome prediction and 0.78 ± 0.02 for binary classification of outcomes. The SHAP explanation analysis identified intubation as the most impactful factor influencing treatment outcomes' predictions for the trained models. Conclusions: The study demonstrates the potential of ML for predicting surgical outcomes of ruptured cerebral aneurysm treatments. It acknowledged the need for high-quality datasets and external validation to enhance model accuracy and generalizability.

Non-Contact Human Body Dimension Estimation Methods Based on Deep Learning: A Critical Analysis

With the advancement of information technology, Human Body Dimension Estimation (HBDE) is moving towards digitalization and non-contact methods. This paper critically reviews non-contact HBDE methods based on deep learning, primarily covering image-based methods, the Adaptive Body Structure Segmentation (ABSS) algorithm with height and weight measurements, and 3D meshes or point cloud data methods. Convolutional Neural Networks (CNNs) excel in image measurement, while the ABSS algorithm with adaptive segmentation achieves high accuracy. Recent studies, such as the Grey Wolf Optimizer-Elman Neural Network (GBWO-ENN) model, have significantly improved prediction accuracy by optimizing neural network structures. Additionally, 3D point cloud data combined with the Skinned Multi-Person Linear (SMPL) model has proven effective in complex scenarios. This paper summarizes the performance of various methods, analyzes common datasets, and explores future development directions for non-contact HBDE. Despite significant progress, this field still faces challenges such as data quality, scarcity of extreme samples, and privacy issues. Future research should focus on building high-quality datasets and developing lightweight measurement networks to meet practical application needs.

Interpretable prediction of remanence in sintered NdFeB through machine learning strategy

The effects of composition on magnetic properties of sintered NdFeB are generally revealed by trial-and-error method, which is actually time-consuming and resource-intensive. It has been demonstrated that data-driven machine learning (ML) strategy has a great potential in accelerating processes of material research. In this study, accurate prediction of remanence in sintered NdFeB magnets has been achieved by a high-quality and few-shot dataset. We used a set of physical and chemical properties of element as inputs to build ML models based on element classification knowledge. Specifically, according to the position of elements, we divided them into three categories and then calculated the average properties of three positions separately. The strong generalization ability of best-performing ML model was further verified in nineteen unknown samples with only 0.724 % error. In addition, the symbolic regressor (SR) and SHapley Additive exPlanation (SHAP) algorithm were employed to explain the model results from physical perspective. In short, a high-efficiency and feasible ML strategy was developed for predicting remanence of sintered NdFeB magnets. We believe that this work will show great prospects in the composition design and performance screening of sintered NdFeB magnets.

Foundation models in ophthalmology: opportunities and challenges.

Last year marked the development of the first foundation model in ophthalmology, RETFound, setting the stage for generalizable medical artificial intelligence (GMAI) that can adapt to novel tasks. Additionally, rapid advancements in large language model (LLM) technology, including models such as GPT-4 and Gemini, have been tailored for medical specialization and evaluated on clinical scenarios with promising results. This review explores the opportunities and challenges for further advancements in these technologies. RETFound outperforms traditional deep learning models in specific tasks, even when only fine-tuned on small datasets. Additionally, LMMs like Med-Gemini and Medprompt GPT-4 perform better than out-of-the-box models for ophthalmology tasks. However, there is still a significant deficiency in ophthalmology-specific multimodal models. This gap is primarily due to the substantial computational resources required to train these models and the limitations of high-quality ophthalmology datasets. Overall, foundation models in ophthalmology present promising opportunities but face challenges, particularly the need for high-quality, standardized datasets for training and specialization. Although development has primarily focused on large language and vision models, the greatest opportunities lie in advancing large multimodal models, which can more closely mimic the capabilities of clinicians.

Improving machine-learning models in materials science through large datasets

The accuracy of a machine learning model is limited by the quality and quantity of the data available for its training and validation. This problem is particularly challenging in materials science, where large, high-quality, and consistent datasets are scarce. Here we present alexandria, an open database of more than 5 million density-functional theory calculations for periodic three-, two-, and one-dimensional compounds. We use this data to train machine learning models to reproduce seven different properties using both composition-based models and crystal-graph neural networks. In the majority of cases, the error of the models decreases monotonically with the training data, although some graph networks seem to saturate for large training set sizes. Differences in the training can be correlated with the statistical distribution of the different properties. We also observe that graph-networks, that have access to detailed geometrical information, yield in general more accurate models than simple composition-based methods. Finally, we assess several universal machine learning interatomic potentials. Crystal geometries optimised with these force fields are very high quality, but unfortunately the accuracy of the energies is still lacking. Furthermore, we observe some instabilities for regions of chemical space that are undersampled in the training sets used for these models. This study highlights the potential of large-scale, high-quality datasets to improve machine learning models in materials science.

How big is big data?

Big data has ushered in a new wave of predictive power using machine-learning models. In this work, we assess what big means in the context of typical materials-science machine-learning problems. This concerns not only data volume, but also data quality and veracity as much as infrastructure issues. With selected examples, we ask (i) how models generalize to similar datasets, (ii) how high-quality datasets can be gathered from heterogenous sources, (iii) how the feature set and complexity of a model can affect expressivity, and (iv) what infrastructure requirements are needed to create larger datasets and train models on them. In sum, we find that big data present unique challenges along very different aspects that should serve to motivate further work.

Nighttime Pothole Detection: A Benchmark

In the field of computer vision, the detection of road potholes at night represents a critical challenge in enhancing the safety of intelligent transportation systems. Ensuring road safety is of paramount importance, particularly in promptly repairing pothole issues. These abrupt road depressions can easily lead to vehicle skidding, loss of control, and even traffic accidents, especially when water has pooled in or submerged the potholes. Therefore, the detection and recognition of road potholes can significantly reduce vehicle damage and the incidence of safety incidents. However, research on road pothole detection lacks high-quality annotated datasets, particularly under low-light conditions at night. To address this issue, this study introduces a novel Nighttime Pothole Dataset (NPD), independently collected and comprising 3831 images that capture diverse scene variations. The construction of this dataset aims to counteract the insufficiency of existing data resources and strives to provide a richer and more realistic benchmark. Additionally, we develop a baseline detector, termed WT-YOLOv8, for the proposed dataset, based on YOLOv8. We also evaluate the performance of the improved WT-YOLOv8 method and eight state-of-the-art object detection methods on the NPD and the COCO dataset. The experimental results on the NPD demonstrate that WT-YOLOv8 achieves a 2.3% improvement in mean Average Precision (mAP) over YOLOv8. In terms of the key metrics—AP@0.5 and AP@0.75—it shows enhancements of 1.5% and 2.8%, respectively, compared to YOLOv8. The experimental results provide valuable insights into each method’s strengths and weaknesses under low-light conditions. This analysis highlights the importance of a specialized dataset for nighttime pothole detection and shows variations in accuracy and robustness among methods, emphasizing the need for improved nighttime pothole detection techniques. The introduction of the NPD is expected to stimulate further research, encouraging the development of advanced algorithms for nighttime pothole detection, ultimately leading to more flexible and reliable road maintenance and road safety.

A globally sampled high-resolution hand-labeled validation dataset for evaluating surface water extent maps

Abstract. Effective monitoring of global water resources is increasingly critical due to climate change and population growth. Advancements in remote sensing technology, specifically in spatial, spectral, and temporal resolutions, are revolutionizing water resource monitoring, leading to more frequent and high-quality surface water extent maps using various techniques such as traditional image processing and machine learning algorithms. However, satellite imagery datasets contain trade-offs that result in inconsistencies in performance, such as disparities in measurement principles between optical (e.g., Sentinel-2) and radar (e.g., Sentinel-1) sensors and differences in spatial and spectral resolutions among optical sensors. Therefore, developing accurate and robust surface water mapping solutions requires independent validations from multiple datasets to identify potential biases within the imagery and algorithms. However, high-quality validation datasets are expensive to build, and few contain information on water resources. For this purpose, we introduce a globally sampled, high-spatial-resolution dataset labeled using 3 m PlanetScope imagery (Planet Team, 2017). Our surface water extent dataset comprises 100 images, each with a size of 1024×1024 pixels, which were sampled using a stratified random sampling strategy covering all 14 biomes. We highlighted urban and rural regions, lakes, and rivers, including braided rivers and coastal regions. We evaluated two surface water extent mapping methods using our dataset – Dynamic World (Brown et al., 2022), based on Sentinel-2, and the NASA IMPACT model (Paul and Ganju, 2021), based on Sentinel-1. Dynamic World achieved a mean intersection over union (IoU) of 72.16 % and F1 score of 79.70 %, while the NASA IMPACT model had a mean IoU of 57.61 % and F1 score of 65.79 %. Performance varied substantially across biomes, highlighting the importance of evaluating models on diverse landscapes to assess their generalizability and robustness. Our dataset can be used to analyze satellite products and methods, providing insights into their advantages and drawbacks. Our dataset offers a unique tool for analyzing satellite products, aiding the development of more accurate and robust surface water monitoring solutions. The dataset can be accessed via https://doi.org/10.25739/03nt-4f29 (Mukherjee et al., 2024).

Linking transcriptome and morphology in bone cells at cellular resolution with generative AI

Abstract Recent advancements in deep learning (DL) have revolutionized the capability of artificial intelligence (AI) by enabling the analysis of large-scale, complex datasets that are difficult for humans to interpret. However, large amounts of high-quality data are required to train such generative AI models successfully. With the rapid commercialization of single-cell sequencing and spatial transcriptomics platforms, the field is increasingly producing large-scale datasets such as histological images, single-cell molecular data, and spatial transcriptomic data. These molecular and morphological datasets parallel the multimodal text and image data used to train highly successful generative AI models for natural language processing and computer vision. Thus, these emerging data types offer great potential to train generative AI models that uncover intricate biological processes of bone cells at a cellular level. In this Perspective, we summarize the progress and prospects of generative AI applied to these datasets and their potential applications to bone research. In particular, we highlight three AI applications: predicting cell differentiation dynamics, linking molecular and morphological features, and predicting cellular responses to perturbations. To make generative AI models beneficial for bone research, important issues, such as technical biases in bone single-cell datasets, lack of profiling of important bone cell types, and lack of spatial information, need to be addressed. Realizing the potential of generative AI for bone biology will also likely require generating large-scale, high-quality cellular-resolution spatial transcriptomics datasets, improving the sensitivity of current spatial transcriptomics datasets, and thorough experimental validation of model predictions.

A deep learning dataset for metal multiaxial fatigue life prediction

Multiaxial fatigue failure of metals, a common issue in industrial production, often leads to significant losses. Recently, many researchers have applied deep learning methods to predict the multiaxial fatigue life of metals, achieving promising results. Due to the high costs of fatigue testing, training data for deep learning is scarce and labor-intensive to collect. This study meets this need by creating a large-scale, high-quality dataset for multiaxial fatigue life prediction, consisting of 1167 samples from 40 materials collected from literature. The dataset includes key mechanical properties (elastic modulus, yield strength, tensile strength, Poisson’s ratio) and 48 loading paths, along with additional relevant information (composition ratios, processing conditions). Common deep learning models validated the dataset’s effectiveness. This dataset aims to support researchers applying deep learning to fatigue life prediction, addressing the long-standing issue of data scarcity, thereby advancing the intersection of artificial intelligence and metal fatigue research.

Group-common and individual-specific effects of structure-function coupling in human brain networks with graph neural networks.

The human cerebral cortex is organized into functionally segregated but synchronized regions bridged by the structural connectivity of white matter pathways. While structure-function coupling has been implicated in cognitive development and neuropsychiatric disorders, it remains unclear to what extent the structure-function coupling reflects a group-common characteristic or varies across individuals, at both the global and regional brain levels. By leveraging two independent, high-quality datasets, we found that the graph neural network accurately predicted unseen individuals' functional connectivity from structural connectivity, reflecting a strong structure-function coupling. This coupling was primarily driven by network topology and was substantially stronger than that of the correlation approaches. Moreover, we observed that structure-function coupling was dominated by group-common effects, with subtle yet significant individual-specific effects. The regional group and individual effects of coupling were hierarchically organized across the cortex along a sensorimotor-association axis, with lower group and higher individual effects in association cortices. These findings emphasize the importance of considering both group and individual effects in understanding cortical structure-function coupling, suggesting insights into interpreting individual differences of the coupling and informing connectivity-guided therapeutics.

New intelligent models for predicting wax appearance temperature using experimental data – Flow assurance implications

Wax deposition is a major problem during petroleum production causing flow rate reduction and blockage of tubing, pipelines, and surface facilities. Wax deposition occurs when crude oil temperature falls below wax appearance temperature (WAT). WAT is an important parameter for wax deposition modeling and developing, testing, and selecting appropriate wax inhibitors. WAT can be determined in laboratory using several different techniques. Although, each method has its own disadvantages including being time-consuming, expensive, inaccurate, or posing health and safety risks. Experimental WAT data are very scarce in the literature. Reliable and accurate correlations and models for WAT prediction are rare, as well. Hence, development of fast, easy, and reliable models for prediction of WAT is inevitable. The main objective of this research work was to develop and introduce novel intelligent models for precise production of WAT. The artificial intelligent (AI) algorithms used include least-squares support-vector machines (LSSVM), recurrent neural network (RNN), and Adaptive neuro-fuzzy inference system (ANFIS). A high-quality dataset was assembled using experimental WAT data collected from the literature. The dataset consists of 81 experimental which is the largest dataset ever reported. The gathered dataset covers a wide range of density from 0.71 to 0.89 g/cm3, wax content from 0.61 to 25.78 wt%, and Pour point from −36 to 37°C. The selected input parameters include oil density, wax content, and pour point determined using a rigorous feature selection analysis. Statistical analysis showed that pour point has the highest impact on WAT prediction. The modeling results showed that the LSSVM model is the superior model with coefficient of determination (R2) of 0.9893, a root mean squared error (RMSE) of 0.1655, and an average absolute relative deviation (AARD) of 9%. The LSSVM model outperformed the only existing smart model in terms of both performance and accuracy. The proposed smart model can be used for prediction of WAT which is an essential parameter in all wax related research.

Biomass-burning sources control ambient particulate matter, but traffic and industrial sources control volatile organic compound (VOC) emissions and secondary-pollutant formation during extreme pollution events in Delhi

Abstract. Volatile organic compounds (VOCs) and particulate matter (PM) are major constituents of smog. Delhi experiences severe smog during the post-monsoon season, but a quantitative understanding of VOCs and PM sources is still lacking. Here, we conduct a source apportionment study for VOCs and PM using a recent (2022), high-quality dataset of 111 VOCs, PM2.5, and PM10 in a positive matrix factorization (PMF) model. Contrasts between clean monsoon air and polluted post-monsoon air, VOC source fingerprints, and molecular tracers enabled us to differentiate paddy residue burning from other biomass-burning sources, which had previously been impossible. Burning of fresh paddy residue, as well as residential heating and waste burning, contributed the most to observed PM10 levels (25 % and 23 %, respectively) and PM2.5 levels (23 % and 24 %, respectively), followed by heavy-duty vehicles fuelled by compressed natural gas (CNG), with a PM10 contribution of 15 % and a PM2.5 contribution of 11 %. For ambient VOCs, ozone formation potential, and secondary-organic-aerosol (SOA) formation potential, the top sources were petrol four-wheelers (20 %, 25 %, and 30 %, respectively), petrol two-wheelers (14 %, 12 %, and 20 %, respectively), industrial emissions (12 %, 14 %, and 15 %, respectively), solid-fuel-based cooking (10 %, 10 %, and 8 %, respectively), and road construction (8 %, 6 %, and 9 %, respectively). Emission inventories tended to overestimate residential biofuel emissions at least by a factor of 2 relative to the PMF output. The major source of PM pollution was regional biomass burning, while traffic and industries governed VOC emissions and secondary-pollutant formation. Our novel source apportionment method even quantitatively resolved similar biomass and fossil fuel sources, offering insights into both VOC and PM sources affecting extreme pollution events. This approach represents a notable advancement compared to current source apportionment approaches, and it could be of great relevance for future studies in other polluted cities and regions of the world with complex source mixtures.

A Sensor Probe with Active and Passive Humidity Management for In Situ Soil CO2 Monitoring.

Soil CO2 concentration and flux measurements are important in diverse fields, including geoscience, climate science, soil ecology, and agriculture. However, practitioners in these fields face difficulties with existing soil CO2 gas probes, which have had problems with high costs and frequent failures when deployed. Confronted with a recent research project's need for long-term in-soil CO2 monitoring at a large number of sites in harsh environmental conditions, we developed our own CO2 logging system to reduce expense and avoid the expected failures of commercial instruments. Our newly developed soil probes overcome the central challenge of soil gas probes-surviving continuous exposure to soil moisture while remaining open to soil gases-via three approaches: a 3D printed housing (economical for small-scale production) following design principles that correct the usual water permeability flaw of 3D printed materials; passive moisture protection via a hydrophobic, CO2-permeable PTFE membrane; and active moisture protection via a low-power micro-dehumidifier. Our CO2 instrumentation performed well and yielded a high-quality dataset that includes signals related to a prescribed fire as well as seasonal and diel cycles. We expect our technology to support underground CO2 monitoring in fields where it is already practiced and stimulate its expansion into diverse new fields.

The InterVision Framework: An Enhanced Fine-Tuning Deep Learning Strategy for Auto-Segmentation in Head and Neck.

Adaptive radiotherapy (ART) workflows are increasingly adopted to achieve dose escalation and tissue sparing under dynamic anatomical conditions. However, recontouring and time constraints hinder the implementation of real-time ART workflows. Various auto-segmentation methods, including deformable image registration, atlas-based segmentation, and deep learning-based segmentation (DLS), have been developed to address these challenges. Despite the potential of DLS methods, clinical implementation remains difficult due to the need for large, high-quality datasets to ensure model generalizability. This study introduces an InterVision framework for segmentation. The InterVision framework can interpolate or create intermediate visuals between existing images to generate specific patient characteristics. The InterVision model is trained in two steps: (1) generating a general model using the dataset, and (2) tuning the general model using the dataset generated from the InterVision framework. The InterVision framework generates intermediate images between existing patient image slides using deformable vectors, effectively capturing unique patient characteristics. By creating a more comprehensive dataset that reflects these individual characteristics, the InterVision model demonstrates the ability to produce more accurate contours compared to general models. Models are evaluated using the volumetric dice similarity coefficient (VDSC) and the Hausdorff distance 95% (HD95%) for 18 structures in 20 test patients. As a result, the Dice score was 0.81 ± 0.05 for the general model, 0.82 ± 0.04 for the general fine-tuning model, and 0.85 ± 0.03 for the InterVision model. The Hausdorff distance was 3.06 ± 1.13 for the general model, 2.81 ± 0.77 for the general fine-tuning model, and 2.52 ± 0.50 for the InterVision model. The InterVision model showed the best performance compared to the general model. The InterVision framework presents a versatile approach adaptable to various tasks where prior information is accessible, such as in ART settings. This capability is particularly valuable for accurately predicting complex organs and targets that pose challenges for traditional deep learning algorithms.

Challenges in Contemporary Spine Surgery: A Comprehensive Review of Surgical, Technological, and Patient-Specific Issues.

Spine surgery has significantly progressed due to innovations in surgical techniques, technology, and a deeper understanding of spinal pathology. However, numerous challenges persist, complicating successful outcomes. Anatomical intricacies at transitional junctions demand precise surgical expertise to avoid complications. Technical challenges, such as underestimation of the density of fixed vertebrae, individual vertebral characteristics, and the angle of pedicle inclination, pose additional risks during surgery. Patient anatomical variability and prior surgeries add layers of difficulty, often necessitating thorough pre- and intraoperative planning. Technological challenges involve the integration of artificial intelligence (AI) and advanced visualization systems. AI offers predictive capabilities but is limited by the need for large, high-quality datasets and the "black box" nature of machine learning models, which complicates clinical decision making. Visualization technologies like augmented reality and robotic surgery enhance precision but come with operational and cost-related hurdles. Patient-specific challenges include managing postoperative complications such as adjacent segment disease, hardware failure, and neurological deficits. Effective patient outcome measurement is critical, yet existing metrics often fail to capture the full scope of patient experiences. Proper patient selection for procedures is essential to minimize risks and improve outcomes, but criteria can be inconsistent and complex. There is the need for continued technological innovation, improved patient-specific outcome measures, and enhanced surgical education through simulation-based training. Integrating AI in preoperative planning and developing comprehensive databases for spinal pathologies can aid in creating more accurate, generalizable models. A holistic approach that combines technological advancements with personalized patient care and ongoing education is essential for addressing these challenges and improving spine surgery outcomes.

FMAW-YOLOv5s: A deep learning method for detection of methane plumes using optical images

Natural gas hydrates stored in the subsurface seabed of continental margins are one of the most important carbon reservoirs on Earth. Research on natural gas hydrates is of great significance to global warming and ecological protection. Methane plumes caused by crustal dynamics are usually considered as a sign of existence of natural gas hydrates. Detection of methane plumes thus becomes the first step of cold seep research. This paper conducts comprehensive research on detection of methane plumes based on deep learning methods and optical images. First, we proposed a method of creating high quality and balanced datasets for methane plumes detection tasks using open-source videos. We then proposed a FMAW-YOLOv5s method for methane plumes detection. The FMAW-YOLOv5s method improves the traditional YOLOv5s in design of backbone network, neck network and loss function. The FMAW-YOLOv5s method can realize accurate and fast detection of methane plumes with a precision of 96.9% and FPS of 141.7. The lightweight feature of FMAW-YOLOv5s also enables the deployment in edge computing devices such as AUVs and ROVs. This research can not only promote the study of cold seep activities, but also provide meaningful insights for detection of other underwater events such as gas pipelines leakage.