Generated Training Samples Research Articles

Exploring large language models for the generation of synthetic training samples for aspect-based sentiment analysis in low resource settings

Aspect-Based Sentiment Analysis (ABSA) is a fine-grained task in sentiment analysis, aiming to identify sentiment expressed towards specific aspects of an entity. This paper explores the use of Large Language Models (LLMs), specifically GPT-3.5-turbo and Llama-3-70B, for generating annotated data in Aspect-Based Sentiment Analysis (ABSA), aiming to address the scarcity of labeled datasets in the field. Two low-resource scenarios are considered, with 25 and 500 manually annotated examples available. In the 25-example scenario, adding synthetic examples generated through few-shot prompting resulted in F1 scores of 81.33 for Aspect Category Detection (ACD) and 71.71 for Aspect Category Sentiment Analysis (ACSA). For the 500-example scenario, synthetic data augmentation showed a notable gain only for the ACSA task, raising the F1 score from 84.54 to 86.70.

Optical-to-SAR Translation Based on CDA-GAN for High-Quality Training Sample Generation for Ship Detection in SAR Amplitude Images

Optical-to-SAR Translation Based on CDA-GAN for High-Quality Training Sample Generation for Ship Detection in SAR Amplitude Images

Automation of the marked-up data sets formation for machine learning based on simulation modeling

The problematic issue of generating marked-up data sets for training artificial intelligence systems preparing for autonomous operation in the proposed new conditions, when the formation of a feature description of objects and situations with a known target variable is impossible or significantly difficult, is considered. As a general approach to its solution, it is proposed to build and use a software testing area that provides the generation of training and control samples, checking the effectiveness of various machine learning methods on them, forming sets of in-formative features of objects and phenomena, bundles of "vector-implementation of features – classification method (clustering, regression)". The results of the systematization of scientific approaches to the definition of ma-chine learning types, as well as the main machine learning methods used in data mining, are presented. Using the example of solving the problem of assessing the dynamics of changes in the radio-electronic environment, the features of real feature descriptions of objects and phenomena that significantly affect the quality of problem solv-ing are shown. To form training samples that adequately reflect these features, it is proposed to use simulation modeling systems that implement an agent-oriented approach to model construction. An example of setting a task and constructing such a model is presented, which ensures the formation of marked-up data sets corresponding to the simulated environmental conditions.

Localization and saturation of degradation space for weakly-supervised real-world super-resolution

Degradation in the real world is highly complex and random, making real super-resolution datasets expensive. A series of unsupervised methods attempt to model real degradations to address the lack of paired training data in real-world super-resolution tasks. However, these unsupervised methods cannot accurately locate the degradation space and the modeled degradation lacks diversity, making it unable to efficiently saturate the degradation space and leading to suboptimal solutions. In this study, we propose a novel degradation space localization and saturation (DSLS) method for weakly-supervised real-world SR, that can accurately localize the real degradation space to generate diverse degradation samples for training. Specifically, the proposed method uses a few paired data to locate the location of the real degradation space and unaligned data to further precisely locate the boundary of the real degradation space, thereby allowing the generated training samples to possess reasonable real degradation. Furthermore, we propose a degradation modulation mechanism to modulate the degradation strength and direction of the generated training samples, so that the generated training samples can efficiently and accurately saturate the real degradation space, making the trained super-resolution model more robust. Experimental results on real datasets demonstrate that our method can be plug-and-play embedded into the training phase of existing methods, while considerably improving the real super-resolution performance with a slight increase in training cost. We also conduct detailed ablation experiments to analyze and validate our contributions.

A 30 m annual cropland dataset of China from 1986 to 2021

A 30 m annual cropland dataset of China from 1986 to 2021

ScribbleCDNet: Change detection on high-resolution remote sensing imagery with scribble interaction

Change detection on high-resolution remote sensing imagery using end-to-end deep learning methods has attracted considerable attention in recent years. Nevertheless, the performance of end-to-end models on complicated scenarios still is limited. Interactive deep-learning models have proven to be a valuable technique for enhancing model performance with minimal human interaction. For instance, the clicks-based interactive models have attracted much attention recently, however, their performance on large regions or complex areas still can be further improved, because they cannot provide accurate semantics or shape prior information of the change regions for the interactive models, as we know that the shape and semantic features of changed regions in remote sensing imagery are typically irregular and complex. Scribble-based interactive form, which can accurately represent the shape or semantic features of the changed regions, thus it is quite suitable for change detection tasks in remote sensing imagery. Therefore, we proposed a novel interactive deep learning model called ScribbleCDNet in this manuscript, which pioneered the use of scribble as an interactive form for detecting change in bi-temporal high-resolution remote sensing imageries. Compared with the widely used clicks-based interactive deep learning models, the proposed ScribbleCDNet acquired superior results on four open-sourced change detection datasets. Last but not least, we also developed an interactive change detection tool with a user-friendly graphical interface, and it can aid researchers in conducting change detection or generating training samples conveniently. Moreover, the proposed ScribbleCDNet can also inspire researchers to develop other interactive deep-learning models related to semantic segmentation, landcover classification, or object extraction in high-resolution remote sensing imageries.

Cost-effective disaster-induced land cover analysis: a semi-automatic methodology Using machine learning and satellite imagery

ABSTRACT This study introduces a cost-effective and automated methodology to swiftly and accurately gauge disaster-induced land cover changes, harnessing satellite imagery and machine learning. Using open-source platforms like Google Earth Engine, the approach streamlines training sample generation and assesses diverse machine learning algorithms on pre- and post-disaster Landsat images. The innovation lies in its semi-automatic training sample generation methodology, reducing time and manual effort for accurate land cover mapping and addressing the challenge of crafting training samples in disaster settings. The methodology was tested on the 2010 Chilean earthquake and tsunami, revealing a 45% increase in sediment cover and a 15% reduction in water cover, underlining vast landscape alterations. SVM stood out for its consistent accuracy. To validate the methodological robustness and broaden its applicability, we replicated the procedures using diverse land cover collection, including ESA WorldCover 10 m and Dynamic World. This comprehensive validation encompassed varied catastrophic events, extending the methodology’s relevance beyond its original scope. The consistent findings across different scenarios underscored the method’s robustness and underscored its potential for broader utilization. This streamlined, cost-effective approach suits multiple land cover products, setting the stage for constant Earth feature monitoring. In light of escalating natural disasters, the synergy of machine learning and geospatial technology spotlighted here offers promising avenues for improved disaster preparedness and response.

Enhancing spatial resolution in MÖNCH for electron microscopy via deep learning

Hybrid Pixel Detectors (HPDs) are highly suitable in diffraction-based electron microscopy due to their high frame rates (> 1 kHz), high dynamic range, and good radiation hardness. However, their use in imaging applications has been limited by their relatively large pixel size (≥ 55 μm) and high-energy (>80 keV) electrons scattering over multiple pixels in the sensor layer. To realize the full potential of fast, radiation-hard HPDs across electron microscopy modalities, we developed deep learning techniques to precisely localize the impact point of incident electrons in MÖNCH, a charge integrating HPD with 25 μm pixel size. With neural network models trained using labeled data via simulations and experimental measurements, the best spatial resolution obtained, defined in terms of the root mean squared error, was 0.60 pixels for 200 keV electrons, a three-fold improvement over a simple charge centroid method. This article presents the training sample generation, deep learning model design, training results, and imaging outcomes for a sample containing gold nanoparticles.

GAN-assisted Road Segmentation from Satellite Imagery

Geo-information extraction from satellite imagery has become crucial to carry out large-scale ground surveys in a short amount of time. With the increasing number of commercial satellites launched into orbit in recent years, high-resolution RGB color remote sensing imagery has attracted a lot of attention. However, because of the high cost of image acquisition and even more complicated annotation procedures, there are limited high-resolution satellite datasets available. Compared to close-range imagery datasets, existing satellite datasets have a much lower number of images and cover only a few scenarios (cities, background environments, etc. ). They may not be sufficient for training robust learning models that fit all environmental conditions or be representative enough for training regional models that optimize for local scenarios. Instead of collecting and annotating more data, using synthetic images could be another solution to boost the performance of a model. This study proposes a GAN-assisted training scheme for road segmentation from high-resolution RGB color satellite images, which includes three critical components: a) synthetic training sample generation, b) synthetic training sample selection, and c) assisted training strategy. Apart from the GeoPalette and cSinGAN image generators introduced in our prior work, this paper in detail explains how to generate new training pairs using OpenStreetMap (OSM) and introduces a new set of evaluation metrics for selecting synthetic training pairs from a pool of generated samples. We conduct extensive quantitative and qualitative experiments to compare different image generators and training strategies. Our experiments on the downstream road segmentation task show that 1) our proposed metrics are more aligned with the trained model performance compared to commonly used GAN evaluation metrics such as the Fréchet inception distance (FID); and 2) by using synthetic data with the best training strategy, the model performance, mean Intersection over Union (mean IoU), is improved from 60.92% to 64.44%, when 1,000 real training pairs are available for learning, which reaches a similar level of performance as a model that is standard-trained with 4,000 real images (64.59%), i.e. , enabling a 4-fold reduction in real dataset size.

Keeping walls straight: data model and training set size matter for deep learning in building generalization

ABSTRACT Deep learning-backed models have shown their potential of conducting map generalization tasks. However, pioneering studies for raster-based building generalization encountered a common “wabbly-wall effect” that makes the predicted building shapes unrealistic. This effect was identified as a critical challenge in the existing studies. This work proposes a layered data representation model that separately stores a building for generalization and its context buildings in different channels. Incorporating adjustments to training sample generation and prediction tasks, we show how even without using more complex deep learning architectures, the widely used Residual U-Net can already produce straight walls for the generalized buildings and maintains rectangularity and parallelism of the buildings very well for building simplification and aggregation in the scale transition from 1:5,000 to 1:10,000 and 1:5,000 to 1:15,000, respectively. Experiments with visual evaluation and quantitative indicators such as Intersection over Union (IoU), fractality, and roughness index show that using a larger input tensor size is an easy but effective solution to improve prediction. Balancing samples with data augmentation and introducing an attention mechanism to increase network learning capacity can help in certain experiment settings but have obvious tradeoffs. In addition, we find that the defects observed in previous studies may be due to a lack of enough training samples. We thus conclude that the wabbly-wall challenge can be solved, paving the way for further studies of applying raster-based deep learning models on map generalization.

Fast uncertainty quantification of spent nuclear fuel with neural networks

The accurate calculation and uncertainty quantification of the characteristics of spent nuclear fuel (SNF) play a crucial role in ensuring the safety, efficiency, and sustainability of nuclear energy production, waste management, and nuclear safeguards. State of the art physics-based models, while reliable, are computationally intensive and time-consuming. This paper presents a surrogate modeling approach using neural networks (NN) to predict a number of SNF characteristics with reduced computational costs compared to physics-based models. An NN is trained using data generated from CASMO5 lattice calculations. The trained NN accurately predicts decay heat and nuclide concentrations of SNF, as a function of key input parameters, such as enrichment, burnup, cooling time between cycles, mean boron concentration and fuel temperature. The model is validated against physics-based decay heat simulations and measurements of different uranium oxide fuel assemblies from two different pressurized water reactors. In addition, the NN is used to perform sensitivity analysis and uncertainty quantification. The results are in very good alignment to CASMO5, while the computational costs (taking into account the costs of generating training samples) are reduced by a factor of 10 or more. Our findings demonstrate the feasibility of using NNs as surrogate models for fast characterization of SNF, providing a promising avenue for improving computational efficiency in assessing nuclear fuel behavior and associated risks.

FARM: A fully automated rice mapping framework combining Sentinel-1 SAR and Sentinel-2 multi-temporal imagery

Rice farming exemplifies intensive agriculture, demanding significant inputs to achieve optimal yields. Thus, accurate and precise mapping of rice cultivation is vital for effective agricultural management and food security. However, such studies have been limited by the challenges of obtaining optical cloud-free data and dealing with radar data's speckle noise. Identifying crops using a single data source poses many difficulties. Additionally, acquiring sufficient representative training samples that accurately reflect diverse phenological patterns is challenging for large-scale monitoring and rice cultivation classification. To address these challenges, this study proposed a fully automated rice-mapping framework (FARM) that combines the strengths of time-series synthetic aperture radar (SAR) and optical satellite imagery for large-scale rice mapping without manual sample collection. First, an object-based, fully automatic training sample generation strategy is introduced. The phenology constraint rule, based on time-series SAR satellite images and specific rice-flooding features, is used to extract rice sample objects. Second, the extracted rice sample objects, adhering to phenological rules, serve as training samples for paddy rice extraction by integrating multiple random forest (RF) classifiers, referred to as the multi-RF method, where each RF is individually built using images acquired during each phenological phase of the growing season. Furthermore, the study explored the availability of the method in early-season rice identification by transferring the training samples acquired by the FARM to a new year. The proposed FARM approach was then validated under different cropping conditions at three study sites in China and two sites in other countries. The results showed that the FARM framework proved to be more effective than other methods at all five study sites, achieving average overall accuracies (OA) ranging from 89.67%−97.00%. In addition, when transferring the training samples from 2021 to other years (2020/2022), the OAs of site A, site B and site C were high during the heading period, with accuracies of 97.57%, 84.28% and 89.27%, respectively. These results demonstrated that, first, the FARM framework exhibits high efficiency and accuracy in different study areas without the need for extensive fieldwork to collect training samples. Second, the method has good performance in the early-season rice mapping of the new year and can be used to perform timely and accurate rice identification and monitoring tasks. The method shows great potential in obtaining large-area automatic rice mapping results in a timely and accurate manner. The latest release of FARM can be viewed at https://feature-selected.users.earthengine.app/view/farm, and the code is available on GitHub at the link https://github.com/gactyxc/FARM.git.

AGTML: A novel approach to land cover classification by integrating automatic generation of training samples and machine learning algorithms on Google Earth Engine

The timely, accurate, and automatic acquisition of land cover (LC) information is a prerequisite for detecting LC dynamics and performing ecological analyses. Cloud computing platforms, such as the Google Earth Engine, have substantially improved the efficiency and scale of LC classification. However, the lack of sufficient and representative training samples hinders automatic and accurate LC classification. In this study, we propose a new approach that integrates the automatic generation of training samples and machine learning algorithms (AGTML) for LC classification in Heilongjiang Province, China. After optimal focal radii were determined for different LC types using Landsat 8 based on focal statistics and unique phenology. Then target training samples were automatically generated based on the improved distance measure SED (a composite of Spectral angle distance (SAD) and Euclidean distance (ED)). Furthermore, LC classification was performed using four feature combinations and three machine learning algorithms. According to independent validation data, the automatically generated training samples demonstrated good representativeness and stability among all three classifiers, with an overall accuracy (OA) of classification higher than 86%, and showed high consistency in the landscape pattern of classification. RF yielded the highest classification accuracy (92.99% OA). AGTML outperformed GLC-FCS30 in identifying large fragmentation and small patch regions in the landscape types. The AGTML approach was subsequently applied to the Guanzhong Plain using different satellite imagery. Results were consistent and accurate (>96.50% OA), demonstrating that the AGTML approach can be applied to various regions and sensors, and has immense potential for automated LC classification across regional and global scales.

Accelerating Deep Neural Network training for autonomous landing guidance via homotopy

The feasibility of utilizing Deep Neural Networks for generating guidance commands in spacecraft landing scenarios has garnered recent research interest. A primary challenge lies in the inefficiency of training sample generation. This paper proposes a new approach to address it by utilizing nonlinear homotopy methods, which involve embedding the homotopic parameter either into the performance index or the right-hand side of the differential equations. Unlike the existing approaches that overlooked the potential value of the intermediate solutions during homotopy process, the proposed method incorporates these discarded solutions as supplementary training data, which treats the homotopic parameter and state as inputs, and the control as the output. Two examples, namely a minimum-fuel lunar landing and a minimum-time asteroid landing, are provided to illustrate the effectiveness and advantages of the proposed method.

Using GEDI as training data for an ongoing mapping of landscape-scale dynamics of the plant area index

Leaf or plant area index (LAI, PAI) information is frequently used to describe vegetation structure in environmental science. While field measurements are time-consuming and do not scale to landscapes, model-based air- or space-borne remote-sensing methods have been used for many years for area-wide monitoring. As of 2019, NASA’s Global Ecosystem Dynamics Investigation (GEDI) mission delivers a point-based LAI product with 25 m footprints and periodical repetition. This opens up new possibilities in integrating GEDI as frequently generated training samples with high resolution (spectral) sensors. However, the foreseeable duration of the system installed on the ISS is limited. In this study we want to test the potential of GEDI for regional comprehensive LAI estimations throughout the year with a focus on its usability beyond the lifespan of the GEDI mission. We study the landscape of Hesse, Germany, with its pronounced seasonal changes. Assuming a relationship between GEDI’s PAI and Sentinel-1 and -2 data, we used a Random Forest approach together with spatial variable selection to make predictions for new Sentinel scenes. The model was trained with two years of GEDI PAI data and validated against a third year to provide a robust and temporally independent model validation. This ensures the applicability of the validation for years outside the training period, reaching a total RMSE of 1.12. Predictions for the test year showed the expected seasonal and spatial patterns indicated by RMSE values ranging between 0.75 and 1.44, depending on the land cover class. The overall prediction performance shows good agreement with the test data set of the independent year which supports our assumption that the usage of GEDI’s PAI beyond the mission lifespan is feasible for regional studies.

Distantly-Supervised Named Entity Recognition with Adaptive Teacher Learning and Fine-Grained Student Ensemble

Distantly-Supervised Named Entity Recognition (DS-NER) effectively alleviates the data scarcity problem in NER by automatically generating training samples. Unfortunately, the distant supervision may induce noisy labels, thus undermining the robustness of the learned models and restricting the practical application. To relieve this problem, recent works adopt self-training teacher-student frameworks to gradually refine the training labels and improve the generalization ability of NER models. However, we argue that the performance of the current self-training frameworks for DS-NER is severely underestimated by their plain designs, including both inadequate student learning and coarse-grained teacher updating. Therefore, in this paper, we make the first attempt to alleviate these issues by proposing: (1) adaptive teacher learning comprised of joint training of two teacher-student networks and considering both consistent and inconsistent predictions between two teachers, thus promoting comprehensive student learning. (2) fine-grained student ensemble that updates each fragment of the teacher model with a temporal moving average of the corresponding fragment of the student, which enhances consistent predictions on each model fragment against noise. To verify the effectiveness of our proposed method, we conduct experiments on four DS-NER datasets. The experimental results demonstrate that our method significantly surpasses previous SOTA methods. The code is available at https://github.com/zenhjunpro/ATSEN.

Positioning failure error identification of industrial robots based on particle swarm optimization and Kriging surrogate modeling

AbstractTo tackle the low identification accuracy of robot positioning, this paper proposes a positioning failure error identification method for industrial robots, combining the Kriging surrogate model and the improved particle swarm optimization algorithm. First, based on the Denavit‐Hartenberg and the small displacement screw methods, the kinematic model of a six‐degree‐of‐freedom industrial robot with joint errors is constructed. Then, the Kriging surrogate model of the kinematic error is constructed, which is trained by generated training samples. Finally, the joint error identification of industrial robots is solved by an improved particle swarm optimization algorithm based on exponential inertia weight and sub‐population cooperation under the multi‐attitude positioning condition. The simulation results show that the proposed optimization algorithm can significantly improve the convergence accuracy and accurately identify the actual errors at each joint of the industrial robot.

Physics-Driven Deep Learning Methods for Fast Quantitative Magnetic Resonance Imaging: Performance improvements through integration with deep neural networks

Quantitative magnetic resonance imaging (qMRI) aims to obtain quantitative biophysical parameters based on physical models derived from MR spin magnetization evolution. This requires the acquisition of multiple MR images, resulting in very long scan times. Recently, deep learning (DL) has shown significant potential for reconstructing undersampled MR data. When applying DL to fast qMRI, physical priors can be integrated into the DL framework to improve the performance of qMRI further. This article introduces the physical models of qMRI and four ways to integrate these models into a deep neural network (NN), namely, training sample generation, contrast image prediction, loss function design, and network architecture design. The core concepts and methods for each category are presented. The associated signal processing issues regarding the generalization and reliability of physics-driven DL-based fast qMRI methods are also discussed.

LSTM-based real-time signal quality assessment for blood volume pulse analysis.

Remote photoplethysmogram (rPPG) is a low-cost method to extract blood volume pulse (BVP). Some crucial vital signs, such as heart rate (HR) and respiratory rate (RR) etc. can be achieved from BVP for clinical medicine and healthcare application. As compared to the conventional PPG methods, rPPG is more promising because of its non-contacted measurement. However, both BVP detection methods, especially rPPG, are susceptible to motion and illumination artifacts, which lead to inaccurate estimation of vital signs. Signal quality assessment (SQA) is a method to measure the quality of BVP signals and ensure the credibility of estimated physiological parameters. But the existing SQA methods are not suitable for real-time processing. In this paper, we proposed an end-to-end BVP signal quality evaluation method based on a long short-term memory network (LSTM-SQA). Two LSTM-SQA models were trained using the BVP signals obtained with PPG and rPPG techniques so that the quality of BVP signals derived from these two methods can be evaluated, respectively. As there is no publicly available rPPG dataset with quality annotations, we designed a training sample generation method with blind source separation, by which two kinds of training datasets respective to PPG and rPPG were built. Each dataset consists of 38400 high and low-quality BVP segments. The achieved models were verified on three public datasets (IIP-HCI dataset, UBFC-Phys dataset, and LGI-PPGI dataset). The experimental results show that the proposed LSTM-SQA models can effectively predict the quality of the BVP signal in real-time.

Unsupervised Ship Detection in SAR Images Using Superpixels and CSPNet

Ship detection in synthetic aperture radar (SAR) images is critical to the ocean surveillance and rescue. Although many deep learning SAR ship detection methods have been proposed, the performance of these methods depends on the size and quality of the training samples. To resolve these issues, this letter presents an unsupervised ship detection method in SAR images using superpixel segmentation and cross stage partial network (CSPNet). First, the SAR image is over-segmented into superpixels based on our previously proposed superpixel generation algorithm. Then, the complex signal kurtosis (CSK) and a local superpixel contrast are integrated as a statistical indicator for the automatic identification of ship superpixels and background superpixels, thus leading to generation of training samples. Finally, the segmented superpixels are input to the CSPNet, which can learn a representative feature set with high discrimination ability between ships and backgrounds. Our method can achieve pixel-level detection map rather than the bounding box result. Experiments based on the Gaofen-3 and TerraSAR SAR data demonstrate that our method can achieve above 90% actual detection rate.