Selection Of Training Points Research Articles

Automatic wide area land cover mapping using Sentinel-1 multitemporal data

This study introduces a methodology for land cover mapping across extensive areas, utilizing multitemporal Sentinel-1 Synthetic Aperture Radar (SAR) data. The objective is to effectively process SAR data to extract spatio-temporal features that encapsulate temporal patterns within various land cover classes. The paper outlines the approach for processing multitemporal SAR data and presents an innovative technique for the selection of training points from an existing Medium Resolution Land Cover (MRLC) map. The methodology was tested across four distinct regions of interest, each spanning 100 × 100 km2, located in Siberia, Italy, Brazil, and Africa. These regions were chosen to evaluate the methodology’s applicability in diverse climate environments. The study reports both qualitative and quantitative results, showcasing the validity of the proposed procedure and the potential of SAR data for land cover mapping. The experimental outcomes demonstrate an average increase of 16% in overall accuracy compared to existing global products. The results suggest that the presented approach holds promise for enhancing land cover mapping accuracy, particularly when applied to extensive areas with varying land cover classes and environmental conditions. The ability to leverage multitemporal SAR data for this purpose opens new possibilities for improving global land cover maps and their applications.

A novel training point selection strategy guided by the maximum reduction of structural state misclassification probability for time-dependent reliability analysis

When adaptively selecting training point in Kriging model based time-dependent reliability analysis, existing methods may lose efficiency due to less considering the Kriging prediction correlation at different time instant as well as the interaction of random input and time variable. Thus, a novel training point selection strategy is proposed by maximizing structural state misclassification probability (SSMP) reduction. Firstly, by taking the Kriging prediction correlation at different time instant into account, the SSMP is derived for random input candidate sample. Secondly, since it is inefficient to actually evaluate SSMP, the expected SSMP is approximately derived with the current Kriging model information by virtually adding the random input and time candidate sample pair into training set. Then, the SSMP reduction, i.e., the difference between the SSMP derived in first step and the expected SSMP in the second one, can be evaluated. By actually adding the random input and time candidate sample pair with the maximum SSMP reduction into training set, the novel training point selection strategy is proposed to update the Kriging model for minimizing SSMP expectedly. Due to considering the Kriging prediction correlation in deriving SSMP, as well as the interaction of random input and time in evaluation the SSMP reduction, the maximum SSMP reduction guided training point selection strategy is superior to the up-to-date methods. The presented example results show that the proposed method needs a least number of evaluating performance function in four compared adaptive Kriging model based methods, and the reduction number of evaluating performance function of the proposed method with respect to the up-to-date methods is more in complex engineering examples than in the simple numerical ones.

Active Learning Based on Transfer Learning Techniques for Text Classification

Text preprocessing is a common task in machine learning applications that involves hand-labeling sets. Although automatic and semi-automatic annotation of text data is a growing field, researchers need to develop models that use resources as efficiently as possible for a learning task. The goal of this work was to learn faster with fewer resources. In this paper, the combination of active and transfer learning was examined with the purpose of developing an effective text categorization method. These two forms of learning have proven their efficiency and capacity to train correct models with substantially less training data. We considered three types of criteria for selecting training points: random selection, uncertainty sampling criterion and active transfer selection. Experimental evaluation was performed on five data sets from different domains. The findings of the experiments suggest that by combining active and transfer learning, the algorithm performs better with fewer labels than random selection of training points.

Fast Posterior Estimation of Cardiac Electrophysiological Model Parameters via Bayesian Active Learning.

Probabilistic estimation of cardiac electrophysiological model parameters serves an important step toward model personalization and uncertain quantification. The expensive computation associated with these model simulations, however, makes direct Markov Chain Monte Carlo (MCMC) sampling of the posterior probability density function (pdf) of model parameters computationally intensive. Approximated posterior pdfs resulting from replacing the simulation model with a computationally efficient surrogate, on the other hand, have seen limited accuracy. In this study, we present a Bayesian active learning method to directly approximate the posterior pdf function of cardiac model parameters, in which we intelligently select training points to query the simulation model in order to learn the posterior pdf using a small number of samples. We integrate a generative model into Bayesian active learning to allow approximating posterior pdf of high-dimensional model parameters at the resolution of the cardiac mesh. We further introduce new acquisition functions to focus the selection of training points on better approximating the shape rather than the modes of the posterior pdf of interest. We evaluated the presented method in estimating tissue excitability in a 3D cardiac electrophysiological model in a range of synthetic and real-data experiments. We demonstrated its improved accuracy in approximating the posterior pdf compared to Bayesian active learning using regular acquisition functions, and substantially reduced computational cost in comparison to existing standard or accelerated MCMC sampling.

A new 30 meter resolution global shoreline vector and associated global islands database for the development of standardized ecological coastal units

ABSTRACTA new 30-m spatial resolution global shoreline vector (GSV) was developed from annual composites of 2014 Landsat satellite imagery. The semi-automated classification of the imagery was accomplished by manual selection of training points representing water and non-water classes along the entire global coastline. Polygon topology was applied to the GSV, resulting in a new characterisation of the number and size of global islands. Three size classes of islands were mapped: continental mainlands (5), islands greater than 1 km2 (21,818), and islands smaller than 1 km2 (318,868). The GSV represents the shore zone land and water interface boundary, and is a spatially explicit ecological domain separator between terrestrial and marine environments. The development and characteristics of the GSV are presented herein. An approach is also proposed for delineating standardised, high spatial resolution global ecological coastal units (ECUs). For this coastal ecosystem mapping effort, the GSV will be used to separate the nearshore coastal waters from the onshore coastal lands. The work to produce the GSV and the ECUs is commissioned by the Group on Earth Observations (GEO), and is associated with several GEO initiatives including GEO Ecosystems, GEO Marine Biodiversity Observation Network (MBON) and GEO Blue Planet.

Construction of Dynamic Multifidelity Locally Optimized Surrogate Models

The conceptual design phase for the development of new and revolutionary aircraft should include both low- and high-fidelity multidisciplinary engineering analyses to enable reasonable computational cost and accuracy tradeoffs. Multifidelity locally optimized surrogate models can be employed to fuse the multifidelity information to give an accurate representation of the underlying design space. These surrogates can be further enhanced with derivative information and augmented with dynamic training point selection to reduce overall computational cost. In this paper, a framework for multifidelity locally optimized surrogate models is developed, and the superior accuracy for the same computational cost or cheaper cost for the same accuracy compared to standard single-fidelity global surrogate models is demonstrated via analytic test functions. The multifidelity locally optimized surrogate models are also applied to the creation of an aerodynamic database involving the steady turbulent flow around a NACA 00xx airfoil, where the influence of Mach number, angle of attack, and thickness-to-chord ratio variations on the lift and drag coefficient are studied.

Global sensitivity analysis-enhanced surrogate (GSAS) modeling for reliability analysis

An essential issue in surrogate model-based reliability analysis is the selection of training points. Approaches such as efficient global reliability analysis (EGRA) and adaptive Kriging Monte Carlo simulation (AK-MCS) methods have been developed to adaptively select training points that are close to the limit state. Both the learning functions and convergence criteria of selecting training points in EGRA and AK-MCS are defined from the perspective of individual responses at Monte Carlo samples. This causes two problems: (1) some extra training points are selected after the reliability estimate already satisfies the accuracy target; and (2) the selected training points may not be the optimal ones for reliability analysis. This paper proposes a Global Sensitivity Analysis enhanced Surrogate (GSAS) modeling method for reliability analysis. Both the convergence criterion and strategy of selecting new training points are defined from the perspective of reliability estimate instead of individual responses of MCS samples. The new training points are identified according to their contribution to the uncertainty in the reliability estimate based on global sensitivity analysis. The selection of new training points stops when the accuracy of the reliability estimate reaches a specific target. Five examples are used to assess the accuracy and efficiency of the proposed method. The results show that the efficiency and accuracy of the proposed method are better than those of EGRA and AK-MCS.

Unified Framework for Training Point Selection and Error Estimation for Surrogate Models

A unified framework for surrogate model training point selection and error estimation is proposed. Building auxiliary local surrogate models over subdomains of the global surrogate model forms the basis of the proposed framework. A discrepancy function, defined as the absolute difference between response predictions from local and global surrogate models for randomly chosen test candidates, drives the framework, thereby not requiring any additional exact function evaluations. The benefits of this new approach are demonstrated with analytical test functions and the construction of a two-dimensional aerodynamic database. The results show that the proposed training point selection approach improves the convergence monotonicity and produces more accurate surrogate models compared to random and quasi-random training point selection strategies. The introduced root-mean-square discrepancy and maximum absolute discrepancy exhibit close agreement with the actual root-mean-square error and maximum absolute error, respectively, and are therefore proposed as a measure for the approximation accuracy of surrogate models in applications of practical interest. Multivariate interpolation and regression is employed to build local surrogates, whereas kriging and polynomial chaos expansions serve as global surrogate models in demonstrating the applicability of the proposed framework.

Evaluation of diverse classification approaches for land use/cover mapping in a Mediterranean region utilizing Hyperion data

Information on Earth's land surface cover is commonly obtained through digital image analysis of data acquired from remote sensing sensors. In this study, we evaluated the use of diverse classification techniques in discriminating land use/cover types in a typical Mediterranean setting using Hyperion imagery. For this purpose, the spectral angle mapper (SAM), the object-based and the non-linear spectral unmixing based on artificial neural networks (ANNs) techniques were applied. A further objective had been to investigate the effect of two approaches for training sites selection in the SAM classification, namely of the pixel purity index (PPI) and of the direct selection of training points from the Hyperion imagery assisted by a QuickBird imagery and field-based training sites. Object-based classification outperformed the other techniques with an overall accuracy of 83%. Sub-pixel classification based on the ANN showed an overall accuracy of 52%, very close to that of SAM (48%). SAM applied using the training sites selected directly from the Hyperion imagery supported by the QuickBird image and the field visits returned an increase accuracy by 16%. Yet, all techniques appeared to suffer from the relatively low spatial resolution of the Hyperion imagery, which affected the spectral separation among the land use/cover classes.

GIS-Based Slope Stability Analysis, Chuquicamata Open Pit Copper Mine, Chile

The risk of slope failure in the Chuquicamata open-pit mine was analyzed using Geographic Information System (GIS) software and modeling techniques. Models incorporated various component layers at a relatively large map scale (1:5000): alteration, geotechnical unit, proximity to major faults (VIF), GSI (geological strength index), slope (from digital elevation model), proximity to watertable (difference grid between topography and modeled watertable), and composite structural density grid (VIF, smaller faults, and fracture frequency); not all layers were used in all models. Three modeling techniques were used: fuzzy logic, in which parameters in each component layer were ranked by mine geotechnical experts according to their influence in promoting slope failure, and two data-driven techniques, weights-of-evidence and logistic regression, in which statistical correlation of training points (known failures) with parameters were used to derive a relative probability of failure. Because most slope failures are controlled by structure, VIF and smaller faults were divided by orientation into subsets with dip direction parallel, opposite, and normal to slope aspect; these orientations promote circular and planar, toppling, and wedge-type failures, respectively. Density grids of these subsets show high-risk areas for individual failure types. The models demonstrate sensitivity of the analysis to (1) selection of component layers, (2) selection of training points, (3) classification and ranking of categorical parameters, and (4) data problems in certain layers. Predicted high-risk zones in the final models show a high degree of correspondence with recent, post-model failures. Such models can be used to anticipate future pit design concerns. The results presented here illustrate how vast amounts of data, in multiple geo-referenced layers, can be analyzed and modeled using GIS techniques for predictive studies at relatively large map scales. Such modeling techniques could provide a powerful tool for predictive modeling in a vast array of large-map-scale applications requiring similar data integration and evaluation.

Application of neural networks for crossflow milk ultrafiltration simulation

The ability of neural network approach was investigated for the dynamic simulation of crossflow milk ultrafiltration. It aims to model the permeate flux and total hydraulic resistance as a function of pH, fat per cent and operation time. Feed forward perceptron networks with a single hidden layer were used to simulate the time-dependent rate of ultrafiltration from a few experimental data. The effect of the number of training points, the number of hidden neurons and training data arrangements on the accuracy of simulation are studied in this work. The results showed that the quality of simulation could be improved using appropriate selection of training points and small network. The best network was able to accurately capture the non-linear dynamics of milk ultrafiltration, so that the agreement between the actual data and simulated values was excellent with maximum and average errors less than 3.61% and 1.06%, respectively.

Dynamic ultrafiltration of proteins – A neural network approach

A neural network approach for the prediction of the rate of ultrafiltration of proteins has been developed. The approach has been used to predict the rate of ultrafiltration of bovine serum albumin as a function of pH and ionic strength. This is a very non-linear problem that has previously been best described through sophisticated descriptions of protein–protein interactions within the layer close to the membrane surface. Networks with a single hidden layer have been used to predict the dynamic rate of filtration from very few data points. Emphasis has been placed on using a small number of training data points and small networks. Variation of the number of training points and use of different training point selection schemes have shown that it is the quality of training points rather than the quantity that leads to the best predictions. The network training process may be optimised by using physical insights to select appropriate input variables. Testing of the neural network approach showed that it could give excellent agreement with experimental results, with average errors less than 2.7%.