Selection Of Training Samples Research Articles

Akte-Liquid: Acoustic-based Liquid Identification with Smartphones

Liquid identification plays an essential role in our daily lives. However, existing RF sensing approaches still require dedicated hardware such as RFID readers and UWB transceivers, which are not readily available to most users. In this article, we propose Akte-Liquid, which leverages the speaker on smartphones to transmit acoustic signals, and the microphone on smartphones to receive reflected signals to identify liquid types and analyze the liquid concentration. Our work arises from the acoustic intrinsic impedance property of liquids, in that different liquids have different intrinsic impedance, causing reflected acoustic signals of liquids to differ. Then, we discover that the amplitude-frequency feature of reflected signals may be utilized to represent the liquid feature. With this insight, we propose new mechanisms to eliminate the interference caused by hardware and multi-path propagation effects to extract the liquid features. In addition, we design a new Siamese network-based structure with a specific training sample selection mechanism to reconstruct the extracted feature to container-irrelevant features. Our experimental evaluations demonstrate that Akte-Liquid is able to distinguish 20 types of liquids at a higher accuracy, and to identify food additives and measure protein concentration in the artificial urine with a 92.3% accuracy under 1 mg/100 mL as well.

Discrepancy between inter- and intra-subject variability in EEG-based motor imagery brain-computer interface: Evidence from multiple perspectives.

Inter- and intra-subject variability are caused by the variability of the psychological and neurophysiological factors over time and across subjects. In the application of in Brain-Computer Interfaces (BCI), the existence of inter- and intra-subject variability reduced the generalization ability of machine learning models seriously, which further limited the use of BCI in real life. Although many transfer learning methods can compensate for the inter- and intra-subject variability to some extent, there is still a lack of clear understanding about the change of feature distribution between the cross-subject and cross-session electroencephalography (EEG) signal. To investigate this issue, an online platform for motor-imagery BCI decoding has been built in this work. The EEG signal from both the multi-subject (Exp1) and multi-session (Exp2) experiments has been analyzed from multiple perspectives. Firstly we found that with the similar variability of classification results, the time-frequency response of the EEG signal within-subject in Exp2 is more consistent than cross-subject results in Exp1. Secondly, the standard deviation of the common spatial pattern (CSP) feature has a significant difference between Exp1 and Exp2. Thirdly, for model training, different strategies for the training sample selection should be applied for the cross-subject and cross-session tasks. All these findings have deepened the understanding of inter- and intra-subject variability. They can also guide practice for the new transfer learning methods development in EEG-based BCI. In addition, these results also proved that BCI inefficiency was not caused by the subject's unable to generate the event-related desynchronization/synchronization (ERD/ERS) signal during the motor imagery.

Deep learning applied to equipment detection on flat roofs in images captured by UAV

Maintenance on flat roofs is a complex activity. Equipment improperly positioned on flat roofs hinders the correct drainage of water and makes maintenance services more difficult. This article presents an experiment with deep learning algorithms involving 330 images acquired in 9 buildings by Unmanned Aerial Vehicle-UAV. This dataset was created by the authors to optimize decision-making for maintenance through automated processes and is being used for the first time in this article. The dataset refers to condenser equipment positioned on flat roofs and was tested in six state-of-the-art object-detection deep learning algorithms: Region-based convolutional neural networks (Faster R-CNN), Focal Loss (Retina-Net), Adaptive Training Sample Selection (ATSS), VarifocalNet (Vfnet), Side-Aware Boundary Localization (SABL) and FoveaBox (Fovea). Nine performance metrics were applied, achieving successful results by Faster R-CNN (Recall=0.93, Fscore=0.93, MAE=0.43) followed by ATSS (Precision=0.95). In a system with many variables, the target is the identification of the best algorithm capable of solving the proposed problem. In conclusion, the types of errors analyzed in detection alert to the diversity of causes related to the inherent characteristics of flat roofs that induce network confusion.

Few-Shot Learning for Fine-Grained Emotion Recognition Using Physiological Signals

Fine-grained emotion recognition can model the temporal dynamics of emotions. It is temporally more precise when compared to predicting one emotion for activities (e.g., video clip watching). Previous works require large amounts of continuously annotated data to train an accurate recognition model. However, the experiments to collect large amounts of continuously annotated physiological signals are costly and time-consuming. To overcome this challenge, we propose a few-shot learning algorithm EmoDSN which can rapidly converge on a small amount of training data (typically < 10 samples per class (i.e., < 10 shot)) for fine-grained emotion recognition. EmoDSN recognizes fine-grained valence and arousal (V-A) labels by maximizing the distance metric between signal segments with different V-A labels. We tested EmoDSN on three different datasets, CASE, MERCA and CEAP-360VR, collected in three different environments: desktop, mobile and HMD-based virtual reality, respectively. The results from our experiments show that EmoDSN achieves promising results for both one-dimension binary (high/low V-A, 1D-2C) and two-dimensional 5-class (four quadrants of V-A space + neutral, 2D-5C) classification. We get an averaged accuracy of 76.04%, 76.62% and 57.62% for 1D-2C valence, 1D-2C arousal and 2D-5C respectively by using only 5 shot of training data. We also find that EmoDSN can achieve better recognition results trained with fewer annotated samples if we select training samples from the changing points of emotion and the ending moments of video watching.

Pan-tumor T-lymphocyte detection using deep neural networks: Recommendations for transfer learning in immunohistochemistry

The success of immuno-oncology treatments promises long-term cancer remission for an increasing number of patients. The response to checkpoint inhibitor drugs has shown a correlation with the presence of immune cells in the tumor and tumor microenvironment. An in-depth understanding of the spatial localization of immune cells is therefore critical for understanding the tumor’s immune landscape and predicting drug response. Computer-aided systems are well suited for efficiently quantifying immune cells in their spatial context. Conventional image analysis approaches are often based on color features and therefore require a high level of manual interaction. More robust image analysis methods based on deep learning are expected to decrease this reliance on human interaction and improve the reproducibility of immune cell scoring. However, these methods require sufficient training data and previous work has reported low robustness of these algorithms when they are tested on out-of-distribution data from different pathology labs or samples from different organs. In this work, we used a new image analysis pipeline to explicitly evaluate the robustness of marker-labeled lymphocyte quantification algorithms depending on the number of training samples before and after being transferred to a new tumor indication. For these experiments, we adapted the RetinaNet architecture for the task of T-lymphocyte detection and employed transfer learning to bridge the domain gap between tumor indications and reduce the annotation costs for unseen domains. On our test set, we achieved human-level performance for almost all tumor indications with an average precision of 0.74 in-domain and 0.72–0.74 cross-domain. From our results, we derive recommendations for model development regarding annotation extent, training sample selection, and label extraction for the development of robust algorithms for immune cell scoring. By extending the task of marker-labeled lymphocyte quantification to a multi-class detection task, the pre-requisite for subsequent analyses, e.g., distinguishing lymphocytes in the tumor stroma from tumor-infiltrating lymphocytes, is met.

An Improved Lightweight Yolo-Fastest V2 for Engineering Vehicle Recognition Fusing Location Enhancement and Adaptive Label Assignment

Engineering vehicle recognition based on video surveillance is one of the key technologies to assist illegal land use monitoring. At present, the engineering vehicle recognition mainly adopts the traditional deep learning model with a large number of floating-point operations. So, it cannot be achieved in edge devices with limited computing power and storage in real-time. In addition, some lightweight models have problems with inaccurate bounding box locating, low recognition rate, and unreasonable selection of positive training samples for the small object. To solve the problems, the paper proposes an improved lightweight Yolo-Fastest V2 for engineering vehicle recognition fusing location enhancement and adaptive label assignment. The location-enhanced Feature Pyramid Network (FPN) structure combines deep and shallow feature maps to accurately localize bounding boxes. The grouping k-means clustering strategy and adaptive label assignment algorithm select an appropriate anchor for each object based on its shape and Intersection over Union (IoU). The study was conducted on Raspberry Pi 4B 2018 using two datasets and different models. Experiments show that our method achieves the optimal combination in speed and accuracy. Specifically, the mAP50 is increased by 7.02 % with the speed of 11.24FPS under the engineering vehicle data obtained by video surveillance in a rural area of China.

CoF-Net: A Progressive Coarse-to-Fine Framework for Object Detection in Remote-Sensing Imagery

Object detection in remote-sensing images is a crucial task in the fields of Earth observation and computer vision. Despite impressive progress in modern remote-sensing object detectors, there are still three challenges to overcome: 1) complex background interference; 2) dense and cluttered arrangement of instances; and 3) large-scale variations. These challenges lead to two key deficiencies, namely, coarse features and coarse samples, which limit the performance of existing object detectors. To address these issues, in this article, a novel coarse-to-fine framework (CoF-Net) is proposed for object detection in remote-sensing imagery. CoF-Net mainly consists of two parallel branches, namely, coarse-to-fine feature adaptation (CoF-FA) and coarse-to-fine sample assignment (CoF-SA), which aim to progressively enhance feature representation and select stronger training samples, respectively. Specifically, CoF-FA smoothly refines the original coarse features into multispectral nonlocal fine features with discriminative spatial–spectral details and semantic relations. Meanwhile, CoF-SA dynamically considers samples from coarse to fine by progressively introducing geometric and classification constraints for sample assignment during training. Comprehensive experiments on three public datasets demonstrate the effectiveness and superiority of the proposed method.

Improving Twitter Sentiment Analysis Efficiency with SVM-PSO Classification and EFWS Heuristic

In the context of the increasingly interconnected world shaped by microblogging services like Twitter, sentiment analysis holds vital significance. The sway of others' sentiments and opinions over our perceptions is now undeniable, transcending individual viewpoints. Our investigation extends the boundaries of previous endeavours in Twitter sentiment analysis, adopting a strategic utilization of Distant Supervision. However, the existing method deals with challenges in analyzing the immense tweet volume due to computational limitations. To overcome this challenge, novel strategies are led in, aimed at expediting the sentiment analysis process. This involves connecting the subjectivity embedded within tweets to precisely select training samples. Additionally, the conceptual framework of EFWS (Effective Word Score) is introduced, intricately interlinked with tweets. This innovative heuristic, derived from polarity assessments of frequently encountered words, emerges as a catalyst for elevating the efficiency of sentiment classification through conventional machine learning algorithms. Experimental investigations encompass a dataset of 1.6 million tweets, showcasing the superior efficiency and heightened precision of the methods compared to prior approaches. Noteworthy is the overall accuracy of approximately 80%, which ascends to around 85% with the integration of the EFWS heuristic, utilizing a training dataset of 100K tweets – a dataset merely half the magnitude of the baseline model. Remarkably, the proposed model elevates accuracy by 2-3% relative to the baseline, all while accelerating training, achieving double the speed of the baseline. Furthermore, the classification leverages Support Vector Machine (SVM) with tuning via Particle Swarm Optimization (PSO), further enhancing performance.

CBW-MSSANet: A CNN Framework With Compact Band Weighting and Multiscale Spatial Attention for Hyperspectral Image Change Detection

Change detection (CD), aims to detect the changing area of the same scene at different times, which is an important application of remote sensing images. As the key data source of CD, hyperspectral image (HSI) is widely used in CD technology because of its rich spectral-spatial information. However, how to mine the multi-level spatial information of dual-temporal hyperspectral images (HSIs) and focus on the features of the pixels to be classified individually remains a problem in the spatial attention mechanism (SAM). To make full use of the spectral-spatial information of HSIs, in this paper we propose a CNN framework with compact band weighting and multi-scale spatial attention (CBW-MSSANet) for HSI pixel-level CD. The main contributions of this article are as follows: 1) a new method of pseudo-label training sample selection based on k-means (KM) centroid distance is designed; 2) apply the compact band weighting (CBW) module to HSI CD to take full advantage of the spectral information of HSIs; 3) a multi-scale spatial attention (MSSA) module is developed for pixel-level CD, which can mine multi-level spatial information and pay more attention to the features of the pixels to be classified, and combine the spatial information of adjacent pixels to make it more conducive to pixel-level CD. Experimental results on four real HSI datasets demonstrated that the performance of MSSA surpasses the classical single-scale SAM, and CBW-MSSANet is superior to some representative CD methods.

An active learning Gaussian modeling based multi-objective evolutionary algorithm using population guided weight vector evolution strategy.

The inverse model based multi-objective evolutionary algorithm (IM-MOEA) generates offspring by establishing probabilistic models and sampling by the model, which is a new computing schema to replace crossover in MOEAs. In this paper, an active learning Gaussian modeling based multi-objective evolutionary algorithm using population guided weight vector evolution strategy (ALGM-MOEA) is proposed. To properly cope with multi-objective problems with different shapes of Pareto front (PF), a novel population guided weight vector evolution strategy is proposed to dynamically adjust search directions according to the distribution of generated PF. Moreover, in order to enhance the search efficiency and prediction accuracy, an active learning based training sample selection method is designed to build Gaussian process based inverse models, which chooses individuals with the maximum amount of information to effectively enhance the prediction accuracy of the inverse model. The experimental results demonstrate the competitiveness of the proposed ALGM-MOEA on benchmark problems with various shapes of Pareto front.

A Rapid Self-Supervised Deep-Learning-Based Method for Post-Earthquake Damage Detection Using UAV Data (Case Study: Sarpol-e Zahab, Iran)

Immediately after an earthquake, rapid disaster management is the main challenge for relevant organizations. While satellite images have been used in the past two decades for building-damage mapping, they have rarely been utilized for the timely damage monitoring required for rescue operations. Unmanned aerial vehicles (UAVs) have recently become very popular due to their agile deployment to sites, super-high spatial resolution, and relatively low operating cost. This paper proposes a novel deep-learning-based method for rapid post-earthquake building damage detection. The method detects damages in four levels and consists of three steps. First, three different feature types—non-deep, deep, and their fusion—are investigated to determine the optimal feature extraction method. A “one-epoch convolutional autoencoder (OECAE)” is used to extract deep features from non-deep features. Then, a rule-based procedure is designed for the automatic selection of the proper training samples required by the classification algorithms in the next step. Finally, seven famous machine learning (ML) algorithms—including support vector machine (SVM), random forest (RF), gradient boosting (GB), extreme gradient boosting (XGB), decision trees (DT), k-nearest neighbors (KNN), and adaBoost (AB)—and a basic deep learning algorithm (i.e., multi-layer perceptron (MLP)) are implemented to obtain building damage maps. The results indicated that auto-training samples are feasible and superior to manual ones, with improved overall accuracy (OA) and kappa coefficient (KC) over 22% and 33%, respectively; SVM (OA = 82% and KC = 74.01%) was the most accurate AI model with a slight advantage over MLP (OA = 82% and KC = 73.98%). Additionally, it was found that the fusion of deep and non-deep features using OECAE could significantly enhance damage-mapping efficiency compared to those using either non-deep features (by an average improvement of 6.75% and 9.78% in OA and KC, respectively) or deep features (improving OA by 7.19% and KC by 10.18% on average) alone.

RGAN-EL: A GAN and ensemble learning-based hybrid approach for imbalanced data classification

Imbalanced sample distribution is usually the main reason for the performance degradation of machine learning algorithms. Based on this, this study proposes a hybrid framework (RGAN-EL) combining generative adversarial networks and ensemble learning method to improve the classification performance of imbalanced data. Firstly, we propose a training sample selection strategy based on roulette wheel selection method to make GAN pay more attention to the class overlapping area when fitting the sample distribution. Secondly, we design two kinds of generator training loss, and propose a noise sample filtering method to improve the quality of generated samples. Then, minority class samples are oversampled using the improved RGAN to obtain a balanced training sample set. Finally, combined with the ensemble learning strategy, the final training and prediction are carried out. We conducted experiments on 41 real imbalanced data sets using two evaluation indexes: F1-score and AUC. Specifically, we compare RGAN-EL with six typical ensemble learning; RGAN is compared with three typical GAN models. The experimental results show that RGAN-EL is significantly better than the other six ensemble learning methods, and RGAN is greatly improved compared with three classical GAN models.

An efficient method for time-dependent reliability problems with high-dimensional outputs based on adaptive dimension reduction strategy and surrogate model

Time-dependent reliability analysis with high dimensional outputs is a challenge because of ‘curse of dimensionality’ and accurate reliability estimation over entire time interval is computationally expensive. In this paper, an efficient time-dependent reliability analysis method is proposed for systems with high dimensional outputs based on adaptive dimension reduction strategy and Kriging. The adaptive Kriging model is constructed in low-dimensional space after performing principal component analysis (PCA) on original time-dependent output. A new learning function and corresponding stopping criterion are developed as the guideline for selecting training samples at each iteration. The proposed learning function focuses on prediction accuracy over the entire time interval and the stopping criterion is directly linked to failure probability. Subsequently, fewer number of function evaluations is required compared with existing competitive works. Moreover, the key advantage of the proposed method is that it is a single-loop process, i.e., the selection of training samples, dimension reduction and Kriging modeling have been combined simultaneously at each iteration with adaptive manner. However, the existing competitive methods, generally, have two independent stages, i.e., stage one is collecting training samples and stage two is performing PCA and Kriging modelling for time-dependent reliability analysis. Thus, these methods cannot fully utilize the information provided by PCA to find optimal training samples. It is noteworthy that the dimension reduction is also an adaptive process in the proposed method, i.e., dimension reduction changes with the training samples at each iteration. The applicability, accuracy and efficiency of the proposed method are validated through three numerical examples and one engineering example.

Rapid Estimation of Decameter FPAR from Sentinel-2 Imagery on the Google Earth Engine

As a direct indicator of vegetation photosynthesis, the fraction of absorbed photosynthetically active radiation (FPAR) serves as a critical input in a series of land surface models. While existing satellite FPAR products are generally at coarse resolutions ranging from 250 m to 1 km, operational FPAR products at fine resolution are urgently needed in studying land surface processes at the plot scale. However, existing methods for estimating fine-resolution FPAR were mainly designed for Landsat data, and few studies have attempted to develop algorithms for Sentinel-2 data. In particular, the operational estimation of decameter FPAR has a higher requirement for the algorithms in terms of generalizability, efficiency, accuracy, and adaptability to Sentinel-2 data. In this paper, we developed a retrieval chain on the Google Earth Engine (GEE) platform to estimate FPAR by learning the relationship between MODIS FPAR and Sentinel-2 surface reflectance. Scale-consistent multilinear models were used to model the relationship between MODIS FPAR and Sentinel-2 surface reflectance, and the model coefficients were regressed from the selected training samples. To account for the spectral and spatial characteristics of the Sentinel-2 data, we designed criteria for selecting training samples and compared different band combinations. Three strategies for band combination were used: (1) green, red, and near infrared (NIR) bands at 10 m resolution (i.e., three bands); (2) green, red, NIR, and red edge (RE) 1, RE2, and RE3 bands at 20 m resolution (i.e., five bands); and (3) green, red, NIR, RE1, RE2, RE3, shortwave infrared1 (SWIR1) and SWIR2 bands at 20 m resolution (i.e., eight bands). Meanwhile, the official Sentinel Application Platform (SNAP) method has also been implemented to estimate the Sentinel FPAR at 10 m and 20 m resolutions for comparison. Both methods were applied to the western Guanzhong area, Shaanxi Province, China, for FPAR estimation of all cloud-free Sentinel-2 images in 2021. The results show that the scaling-based method using five bands at 20 m resolution was the most accurate compared to the in situ measurements (RMSE = 0.076 and R² = 0.626), which outperformed the SNAP method at 10 m and 20 m resolutions and the scaling-based method using other strategies. The results of the scaling-based method using all three strategies were highly consistent with the MODIS FPAR product, while the SNAP method systematically underestimated FPAR values compared to the MODIS FPAR products. The proposed method is more ready-to-use and more efficient than SNAP software. Considering that the service of the MODIS sensor is overdue, the proposed method can be extended to alternatives to MODIS products, such as VIIRS and Sentinel-3 data.

VGI and Satellite Imagery Integration for Crisis Mapping of Flood Events

Timely mapping of flooded areas is critical to several emergency management tasks including response and recovery activities. In fact, flood crisis maps embed key information for an effective response to the natural disaster by delineating its spatial extent and impact. Crisis mapping is usually carried out by leveraging data provided by satellite or airborne optical and radar sensors. However, the processing of these kinds of data demands experienced visual interpretation in order to achieve reliable results. Furthermore, the availability of in situ observations is crucial for the production and validation of crisis maps. In this context, a frontier challenge consists in the use of Volunteered Geographic Information (VGI) as a complementary in situ data source. This paper proposes a procedure for flood mapping that integrates VGI and optical satellite imagery while requiring limited user intervention. The procedure relies on the classification of multispectral images by exploiting VGI for the semi-automatic selection of training samples. The workflow has been tested with photographs and videos shared on social media (Twitter, Flickr, and YouTube) during two flood events and classification consistency with reference products shows promising results (with Overall Accuracy ranging from 87% to 93%). Considering the limitations of social media-sourced photos, the use of QField is proposed as a dedicated application to collect metadata needed for the image classification. The research results show that the integration of high-quality VGI data and semi-automatic data processing can be beneficial for crisis map production and validation, supporting crisis management with up-to-date maps.

A Qualitative and Quantitative Analysis Strategy for Continuous Turbulent Gas Mixture Monitoring

Electronic noses are one of the predominant technological choices for gas mixture detection, but their application in real-world atmospheric environments still leaves several issues to be resolved. The key bottleneck is the effect of turbulence caused by the diffusion of gases in the atmosphere on the quantitative and qualitative analytical performance of the electronic nose. In light of this, this paper presents a quantitative and qualitative analysis strategy for gas mixture monitoring. This strategy adopts baseline manipulation of the raw sensor data to reduce drift interference, and then performs feature extraction on the multidimensional response signals of the MOS gas sensor array using principal component analysis (PCA). In order to improve gas mixture recognition accuracy, the whale optimization algorithm (WOA) is used to optimize the network structure of the long short-term memory (LSTM) model for turbulent gas mixture composition recognition. The least squares support vector machine (LSSVM) algorithm is adopted to implement turbulent gas mixture concentration prediction. This paper focuses on two aspects of hyper-parameter optimization for the development of an LSSVM with particle swarm optimization (PSO) and for improved training sample selection for the LSSVM which should subsequently increase the accuracy of concentration estimation. The effectiveness of the proposed strategy is evaluated with a dataset from a chemical sensor array exposed to turbulent gas mixtures. Experimental results revealed that the proposed strategy for turbulent gas mixtures has satisfactory outcomes for both qualitative gas composition recognition and quantitative gas concentration prediction.

BoxPaste: An Effective Data Augmentation Method for SAR Ship Detection

Data augmentation is a crucial technique for convolutional neural network (CNN)-based object detection. Thus, this work proposes BoxPaste, a simple but powerful data augmentation method appropriate for ship detection in Synthetic Aperture Radar (SAR) imagery. BoxPaste crops ship objects from one SAR image using bounding box annotations and pastes them on another SAR image to artificially increase the object density in each training image. Furthermore, we dive deep into the characteristics of the SAR ship detection task and draw a principle for designing a SAR ship detector—light models may perform better. Our proposed data augmentation method and modified ship detector attain a 95.5% Average Precision (AP) and 96.6% recall on the SAR Ship Detection Dataset (SSDD), 4.7% and 5.5% higher than the fully convolutional one-stage (FCOS) object detection baseline method. Furthermore, we also combine our data augmentation scheme with two current detectors, RetinaNet and adaptive training sample selection (ATSS), to validate its effectiveness. The experimental results demonstrate that our newly proposed SAR-ATSS architecture achieves 96.3% AP, employing ResNet-50 as the backbone. The experimental results show that the method can significantly improve detection performance.

Universal lymph node detection in T2 MRI using neural networks.

Identification of lymph nodes (LNs) that are suspicious for metastasis in T2 Magnetic Resonance Imaging (MRI) is critical for assessment of lymphadenopathy. Prior work on LN detection has been limited to specific anatomical regions of the body (pelvis, rectum). Therefore, an approach to universally detect both benign and metastatic nodes in T2 MRI studies of the body is highly desirable. We developed a Computer Aided Detection (CAD) pipeline to universally identify LN in T2 MRI. First, we trained various neural networks for detecting LN: Faster RCNN with and without Hard Negative Example Mining (HNEM), FCOS, FoveaBox, VFNet, and Detection Transformer (DETR). Next, we show that VFNet with Adaptive Training Sample Selection (ATSS) outperformed Faster RCNN with HNEM. Finally, we ensembled models that surpassed a 45% mAP threshold. Experiments on 122 test studies revealed that VFNet achieved a 51.1% mAP and 78.7% recall at 4 false positives (FP) per volume, while the one-stage model ensemble achieved a mAP of 52.3% and sensitivity of 78.7% at 4FP. We found that VFNet and the one-stage model ensemble can be interchangeably used in the CAD pipeline. Our CAD pipeline universally detected both benign and metastatic nodes in T2 MRI studies, resulting in a sensitivity improvement of [Formula: see text]14% over the current LN detection approaches (sensitivity of 78.7% at 4 FP vs. 64.6% at 5 FP per volume).

Life-Cycle Production Optimization of the CO2-Water-Alternating-Gas Injection Process Using Least-Squares Support-Vector Regression Proxy

SummaryIn this study, we present a framework for efficient estimation of the optimal carbon dioxide (CO2)-water-alternating-gas (WAG) parameters for robust production-optimization problems by replacing a high-fidelity model with a least-squares support-vector regression (LS-SVR) model. We provide insight and information on the proper selection of feature space and training samples of the LS-SVR proxy model for the CO2-WAG life cycle production optimization problem. Given a set of training points generated from high-fidelity model-based simulation results, an LS-SVR-based proxy model is built to approximate a reservoir-simulation model. The estimated optimal design parameters are then found by maximizing net present value (NPV) using the LS-SVR proxy as the forward model within an iterative-sampling-refinement (ISR) optimization algorithm that is designed specifically to promote the accuracy of the proxy model for robust production optimization. As an optimization tool, the sequential quadratic programming (SQP) method is used. CO2-WAG design variables are CO2 injection and water injection rates for each injection well at each cycle, production bottomhole pressure (BHP) for each production well at each WAG half-cycle, and inflow control valve (ICV) for each well at each WAG half-cycle and at each valve. We study different scenarios where we fix some of the design variables to investigate the importance of design variables on the life cycle production optimization of the CO2-WAG problem. We compare the performance of the proposed method using the LS-SVR runs with the popular stochastic simplex approximate gradient (StoSAG) and reservoir simulation runs for a synthetic example considering a three-layer, channelized reservoir with four injectors and nine producers. Results show that with the properly selected feature space and training points, the proposed LS-SVR-based ISR optimization framework is at least 1.5–8 times computationally more efficient, depending on the cases considered, than the StoSAG using a high-fidelity numerical simulator. However, we observe that the size and sampling of the training data, as well as the selection of well controls and their bound constraints for the well controls, seem to be influential on the performance of the LS-SVR-based optimization method. This is the first LS-SVR application to the CO2-WAG optimal well-control problem. The proposed LS-SVR-based ISR optimization framework has the potential to be used as an efficient tool for the CO2-WAG optimization problem.

AugFCOS: Augmented fully convolutional one-stage object detection network

As a pioneering work of introducing the idea of full convolutional network into the field of object detection, the fully convolutional one-stage object detection network (FCOS) has the advantage of excellent performance with low memory overhead. However, there are certain problems with FCOS that merit more research: the centerness quality assessment loss does not decrease during the late training stage, and its adaptive training sample selection (ATSS) relies heavily on the hyperparameter. To solve the aforementioned problems, we propose a novel object detection network, named augmented fully convolutional one-stage object detection network (AugFCOS). First of all, we propose an improved dynamic optimization loss (DOL) to mitigate the impact of the original centerness loss not decreasing. Then, a Robust Training Sample Selection (RTSS) is proposed to get rid of the dependence of hyper-parameter in ATSS of FCOS. Finally, a novel mixed attention feature pyramid network (MAFPN) is presented to enhance the multi-scale representation ability of feature pyramid network (FPN) and further improve the ability of multi-scale detection. The experimental results on MS COCO demonstrate the effectiveness of our proposed AugFCOS, where AugFCOS achieves approximate 2.0% to 2.9% increase when compared with ATSS and FCOS.