Computational Efficiency Research Articles

Lightweight bearing defect detection method based on collaborative attention and domain adaptive technology

Abstract With the rapid development of industrial manufacturing, bearing defect detection has become a key link to ensure the safe operation of equipment and improve production efficiency. Aiming at the problems of low accuracy and large parameters of traditional bearing defect detection models, this paper proposes a lightweight bearing defect detection method based on collaborative attention and domain adaptive technology. Firstly, the collaborative attention mechanism uses multiple parallel attention heads to extract the features of different regions in the image respectively. Each attention head can independently extract and weigh the features of specific regions, which can capture the complex features of bearing defects more comprehensively, including shape, texture, edge and other key information. At the same time, the domain adaptation technology aims to make the model adapt to the data distribution of different domains. It reduces the difference between the source domain and the object domain by aligning the feature distribution between the source domain and the object domain during the training process, thereby improving the performance of the model on the object domain. The experimental results show that the proposed method is 2.75% higher than the mAP@0.5 of YOLOv7, and has higher computational efficiency and better generalization ability than the traditional method.

Enhancing transparency in data-driven urban pluvial flood prediction using an explainable CNN model

Mitigating severe losses caused by pluvial floods in urban areas with dense population and property requires effective flood prediction for emergency measures. Physics-based models face issues with low computational efficiency for real‐time flood prediction. An alternative approach for rapid prediction instead of physics-based models is to predict from a data-driven perspective. However, data-driven approaches for urban flood prediction are often perceived as “black box” and might raise concerns. In this study, we propose an explainable deep learning (DL) approach for rapid urban pluvial flood prediction with enhanced transparency using a convolutional neural network (CNN) and the explainable artificial intelligence (AI) framework Shapley additive explanation (SHAP). We process a systematic stepwise feature selection process and establish a CNN model considering topography, drainage networks and rainfall to predict maximum inundation depths. Then, SHAP is applied to provide trustworthy explanations for the decision making in model results. The results show that: 1) Forward selection can offer insights into selecting effective input variables for improved predictions and promote understanding of DL modelling. The spatial pattern of inundation depths predicted by the proposed CNN model shows good agreement with those predicted by the physics-based model, demonstrated by average correlation coefficient (CC) and mean absolute error (MAE) values of 0.982 and 0.021 m, respectively. 2) The CNN model substantially outperforms the physics-based model in computational speed when using the same hardware, achieving speedups of 210 times on GPU and 51 times on CPU in the case study (575167 grid cells, 14.38 km2). Particularly, it can still achieve good performance on a CPU-only standard laptop without high-performance hardware, with only a modest increase in computational time. 3) The SHAP explainable analysis confirms that the CNN model correctly captures the relationships between input variables and water depth, revealing a reasonable decision-making process, enhancing its transparency. The explainable DL approach incorporating SHAP for rapid urban pluvial flood prediction is promising to build trust among stakeholders and provide a trustworthy reference for prompt measures aiming at saving lives and protecting assets during flood emergencies. Additionally, the proposed DL approach can potentially be further expanded to analyze the causes of urban flooding events and serve as a foundation for exploring the transferability of data-driven urban flood prediction, providing benefits for better urban flood risk management.

Boosting tree with bootstrap technique for pre-stress design in cable dome structures

Tensegrity structures, known for their rigidity derived from feasible pre-stresses, present unique challenges in structural engineering. Traditional force-finding methods, though comprehensive, rely heavily on intricate matrix computations, making them computationally intensive and often uncomfortable for considering external loads in practical engineering scenarios. This paper introduces a novel approach to compute pre-stresses in cable dome structures by integrating machine learning and probability theory, collectively termed the boosting tree with bootstrap technique (BTWBT). This method reduces the sample size to as few as 100 per iteration, while improving computational efficiency by randomly generating internal forces. By reframing the force determination as an inverse problem, it ensures that structural displacement converges to zero under feasible pre-stresses. The effectiveness of BTWBT is demonstrated across three distinct cable dome structures: the Geiger dome, Kiewitt dome, and rotating hyperboloid cable dome. Results show that BTWBT achieves the preset displacement requirement (maximum nodal displacement below 0.01 mm) with fewer iterations and reduced computational cost compared to traditional machine learning methods. BTWBT's capability to manage complex structural configurations with minimal data, while incorporating random internal force generation ranges, highlights its potential as a superior alternative for force determination in tensegrity structures.

Improving heat transfer prediction across catalytic SiC interface through a multiscale framework leveraging machine learning and data fusion techniques

Accurate aerothermal prediction under hypersonic flow conditions is crucial for any thermal protection materials design and engineering. The surface catalytic effect, where dissociated atoms recombine into their molecular forms at the air-solid interface, is an exothermic reaction and plays an important role in high-enthalpy aerodynamic environment prediction. Taking SiC, a widely recognized high-temperature thermal protection material as an example, a multiscale fusion approach for aerothermal prediction is proposed. The approach utilizes reactive molecular dynamics method to construct an interface catalysis model, and the Bayesian maximum entropy to integrate experimental and simulational data into an optimized database. Subsequently, using the radial basis function neural network algorithm, a machine learning-based reaction kinetics model with precise analysis of surface catalysis is trained. The obtained catalytic recombination efficiency is used as a boundary input for computational fluid dynamics simulation, enabling rapid and accurate prediction of hypersonic thermal environment. The result shows that the predicted surface heat flux by this novel approach is consistent with the benchmark studies, but comes with significantly reduced computation time. The stagantion heat flux predictions based on the assumptions of full catalytic wall, finite rate catalytic wall (from the proposed multiscale fusion method), and non-catalytic wall are determined to be 8.65×106 W/m2, 7.51×106 W/m2, and 4.02×106 W/m2, respectively. Comparing these with the wind tunnel benchmark value of 7.48×106 W/m2, the error from the proposed multiscale fusion method is 0.30 %, indicating significant enhancement in prediction accuracy through the proposed upscaling method. The new approach could not only reveal complex exothermic reaction processes at the interface, but also enhance the prediction accuracy at much higher computational efficiency, providing an alternative for multiscale modeling of complex flow and heat transfer at the interface.

Experiment on Lung Disease Classification using YOLOv8

Lung diseases are a leading cause of illness and mortality worldwide. Accurate and timely diagnosis is crucial for improving patient outcomes, but manual interpretation of chest X-rays by radiologists may consume a significant amount of time and is susceptible to errors. This study leverages the state-of-the-art YOLOv8 deep learning model to develop an automated system for classifying lung diseases from chest X-ray images. The proposed approach utilizes a large, diverse dataset of 21,165 chest X-ray images categorized into four classes: COVID-19, viral pneumonia, lung opacity, and normal. The YOLOv8-cls model is fine-tuned using transfer learning and advanced data augmentation techniques. The system achieves a high accuracy of approximately 95% across all disease classes while maintaining real-time performance. Confusion matrix analysis demonstrates strong performance in correctly identifying each condition, with COVID-19, normal, viral pneumonia and lung opacity cases correctly identified 97%, 97%, 99% and 93% of the time respectively. The outcomes validate the adaptability and reliability of deep learning, specifically YOLOv8, for automated lung disease detection, offering the potential to improve clinical workflows and patient care by providing efficient screening tools for practitioners. The study addresses limitations of previous works by utilizing a large consolidated dataset, reporting comprehensive computational efficiency metrics, and extending classification complexity without substantially impacting accuracy or speed. Future research should focus on further optimizations and extending the curated image repositories to include under-represented patient groups to enhance the model's inclusiveness and robustness.

Detecting covolcanic ionospheric disturbances using GNSS data and a machine learning algorithm

Many natural and man-made phenomena cause ionospheric responses that can be captured in total electron content data obtained from Global Navigation Satellite System observations. However, the increasing volume of the data poses challenges for its analysis. This study provides an algorithm for detecting covolcanic ionospheric disturbances in total electron content time series based on machine learning. Using the Sarychev Peak eruption (June 11–16, 2009) as an example, we identified and labeled the observational data, proposed data features, and generated datasets. The design and number of features were chosen considering the computational efficiency of the algorithm and its potential applicability in monitoring systems. A machine learning model based on a gradient boosting technique was trained and achieved a Matthews correlation coefficient of 0.86 on the test dataset, indicating the high quality. The proposed algorithm detects 96% of the disturbances in the datasets with a minimal number of false positives (48 out of 200 test files), exhibiting a significant improvement compared with the conventional STA\\LTA algorithm, which detected only 11% of the disturbances. The validation of the algorithm on data corresponding to the Calbuco (2015) and Hunga Tonga-Hunga Ha’apai (2022) volcanoes eruptions showed 94% and 81% of the disturbances detected, respectively.

Orientation method based on lunar hour angle

Abstract Astronomical orientation is a crucial technique for obtaining high-precision azimuths. However, at night, this often relies on the North Star, which can be challenging to locate and identify. An orientation method is proposed based on the lunar hour angle as an alternative. By applying theories from astronomy and geodesy, we conduct an in-depth study of the key points and calculation steps for using a theodolite to measure the moon. By observing the moon’s edge and center and combining this with the observer’s geographic location and observation time, we calculate the lunar hour angle to determine the true north azimuth. To improve computational efficiency, an application has been developed to assist with observation and data processing. Practical applications and tests have validated the feasibility and effectiveness of this lunar hour angle-based orientation method. This method addresses the limitations of traditional North Star orientation, offering an efficient and convenient approach to orientation.

A novel meshless method for numerical simulation of fractured-vuggy reservoirs

This paper proposes a novel meshless numerical simulation method for fractured-vuggy reservoirs. Initially, model discretization involves extracting fracture and vug feature nodes along the contours of the reservoir bodies and in alignment with fracture orientations, tailored to the characteristics of the fractured-vuggy oil deposit. Subsequently, a connection element model is established within the influence domain. Based on the seepage equations of the connection system, a computational method for the parameters of the connection elements (connection transmissibility and connection volume) is defined. In addition, a sequential method is employed to solve for the pressure and saturation, achieving rapid prediction of the reservoir's production dynamics. On this basis, an automatic history matching algorithm is utilized for real-time fitting of oil–water dynamic indicators, inverting the characteristic parameters of the connection element model, and quantitatively characterizing the injection–production connection elements. The research findings indicate that the method is capable of rapidly fitting and predicting the dynamics of oil reservoir production. Furthermore, conceptual model examples have substantiated that it achieves comparable computational efficiency and superior computational accuracy when compared to the traditional finite difference (volume) method under the same nodal conditions. Additionally, the node arrangement for fractures and vugs is comparatively flexible, and it can ensure the integrity of the flow paths with a limited number of nodes to enhance predictive accuracy. Therefore, this method can effectively meet the rapid prediction requirements for fractured-vuggy reservoirs and also provides a novel meshless computational approach for the numerical simulation of such reservoirs.

Transformer Based Pluralistic Image Completion With Reduced Information Loss.

Transformer based methods have achieved great success in image inpainting recently. However, we find that these solutions regard each pixel as a token, thus suffering from an information loss issue from two aspects: 1) They downsample the input image into much lower resolutions for efficiency consideration. 2) They quantize 2563 RGB values to a small number (such as 512) of quantized color values. The indices of quantized pixels are used as tokens for the inputs and prediction targets of the transformer. To mitigate these issues, we propose a new transformer based framework called "PUT". Specifically, to avoid input downsampling while maintaining computation efficiency, we design a patch-based auto-encoder P-VQVAE. The encoder converts the masked image into non-overlapped patch tokens and the decoder recovers the masked regions from the inpainted tokens while keeping the unmasked regions unchanged. To eliminate the information loss caused by input quantization, an Un-quantized Transformer is applied. It directly takes features from the P-VQVAE encoder as input without any quantization and only regards the quantized tokens as prediction targets. Furthermore, to make the inpainting process more controllable, we introduce semantic and structural conditions as extra guidance. Extensive experiments show that our method greatly outperforms existing transformer based methods on image fidelity and achieves much higher diversity and better fidelity than state-of-the-art pluralistic inpainting methods on complex large-scale datasets (e.g., ImageNet).

Including dynamic capacity impact in an improved model for optimal flood control operation in river-type reservoirs

Flood control operations in river-type reservoirs are significantly affected by both the dynamic reservoir capacity and flood propagation within the reservoir area. However, the traditional reservoir flood control optimal operation (RFCOO) model is usually based on the static capacity method, which may bring large errors in scheduling calculations for river-type reservoirs. To address this issue, an improved model, the reservoir dynamic capacity flood control optimal operation (RDCFCOO) model, that includes the influence of the dynamic reservoir capacity and flood propagation, was developed to improve the accuracy of flood regulation calculations in river-type reservoirs. This improved model includes a 1D unsteady flow simulation and a rederived objective function based on the maximum peak clipping criterion, and expresses the state transformation equations using the de Saint-Venant equations. Furthermore, a multi-core parallel DP (PDP) algorithm that incorporates reduction of state dimensionality (RSD) was developed to effectively derive optimal operation schemes. The improved model and algorithm developed herein were validated through an application to the Xiangjiaba Reservoir, China. The results show that: (1) By adding the item (qiw) to the objective function and substituting the de Saint-Venant equations for the water balance equation, the RDCFCOO model developed in this paper can account for the impact of dynamic capacity and flood propagation. As a result, it exhibits a more accurate and detailed flood control process that was closer to the actual conditions of river-type reservoirs compared to the RFCOO model; (2) By removing the invalid discrete state points from the search space and fully utilizing multi-core processors, the proposed PDP with RSD can significantly improve the computational efficiency of the DP in solving RDCFCOO problems. The improved model and algorithm can effectively improve the calculation accuracy of flood control optimal scheduling in river-type reservoirs and provide more information for decision-makers.

Maxwell boundary condition for discrete velocity methods: macroscopic physical constraints and Lagrange multiplier-based implementation

In this paper, we propose an algorithm that imposes macroscopic physical constraints with Lagrange multiplier approach in implementing the Maxwell boundary condition within the framework of the discrete velocity method. For the simulation of rarefied gas flows in the presence of solid walls with complex geometry, the distribution function in the reflection region of the wall surface needs to be constructed in the discrete velocity space, to fulfill the specular reflection in the Maxwell boundary condition. The construction process should not consist of interpolation only, but include certain macroscopic physical constraints at the wall surface, so as to correctly account for gas-surface interaction on a macroscopic level. We demonstrate that for the specular reflection, keeping the symmetry of the first three moments of the distribution function between the incident and reflected region is sufficient for maintaining the conservation of mass, momentum, and energy at the wall surface. Furthermore, to strictly satisfy macroscopic physical constraints, a Lagrange multiplier method is introduced into the construction of the distribution function to correct the pure interpolation solution. In addition, the construction process requires the inversion of a large and sparse matrix (of dimension N × N, where N is the number of points in the velocity space). To improve the computational efficiency, the matrix inversion is converted into that of a much smaller matrix, i.e. (D+2) × (D+2) in the D-dimensional physical space. A series of numerical experiments are conducted to examine the performance of the proposed algorithm under different flow conditions. We demonstrate that the results obtained by the proposed algorithm are more accurate than the pure interpolation solution, comparing with the benchmark data. Moreover, after the validation of our results with previous studies, we find that the method significantly enhances the conservation of total mass and energy, especially for flows in an enclosed domain.

LS-Net: lightweight segmentation network for dermatological epidermal segmentation in optical coherence tomography imaging.

Optical coherence tomography (OCT) can be an important tool for non-invasive dermatological evaluation, providing useful data on epidermal integrity for diagnosing skin diseases. Despite its benefits, OCT's utility is limited by the challenges of accurate, fast epidermal segmentation due to the skin morphological diversity. To address this, we introduce a lightweight segmentation network (LS-Net), a novel deep learning model that combines the robust local feature extraction abilities of Convolution Neural Network and the long-term information processing capabilities of Vision Transformer. LS-Net has a depth-wise convolutional transformer for enhanced spatial contextualization and a squeeze-and-excitation block for feature recalibration, ensuring precise segmentation while maintaining computational efficiency. Our network outperforms existing methods, demonstrating high segmentation accuracy (mean Dice: 0.9624 and mean IoU: 0.9468) with significantly reduced computational demands (floating point operations: 1.131 G). We further validate LS-Net on our acquired dataset, showing its effectiveness in various skin sites (e.g., face, palm) under realistic clinical conditions. This model promises to enhance the diagnostic capabilities of OCT, making it a valuable tool for dermatological practice.

Enhancing Shale Gas EUR Predictions with TPE optimized SMOGN: A Comparative Study of Machine Learning Algorithms in the Marcellus Shale with an Imbalanced Dataset

Oil and gas operators frequently rely on traditional methods to predict Estimated Ultimate Recovery (EUR) but, these methods often fail to accurately predict shale gas EUR. Therefore, machine learning (ML) algorithms were shown as promising alternatives but, the negative effects of imbalanced datasets on their performance still remains underexplored. This study addresses this gap with a Tree-Parzen Estimator (TPE) optimized Synthetic Minority Over-sampling Technique for Regression with Gaussian Noise (SMOGN) to alleviate the detrimental effects of the imbalanced datasets on model performance. Two cases were compared: one employed standard pre-processing while, the other employed TPE optimized SMOGN. Four ML algorithms: Artificial Neural Network (ANN), Deep-ANN, Support Vector Regression (SVR), and eXtreme Gradient Boosting (XGBoost) were trained across both cases with an imbalanced dataset of 460 Marcellus shale wells. Exploratory data analysis revealed that the imbalance was due to the evolution of completion techniques, leading to the underrepresentation of wells completed with more recent, aggressive methods. The proposed framework improved of the models from 0.8243 - 0.8934 to 0.8851 - 0.9186 with more significant gains in the underrepresented wells in the higher EUR regions (), where the improved from 0.2155 - 0.4598 to 0.5615- 0.9472. The SMOGN enhanced SVR had the highest computational efficiency ( to train) and highest performance of 0.9472) in these higher EUR regions, while the SMOGN enhanced Deep-ANN had the highest overall performance ( of 0.9186). This framework outperforms standard pre-processing frameworks. Additionally, it enables operators to predict shale gas EUR more accurately for future infill wells completed with recent techniques, even while the wells are still producing in transient flow, facilitating early cost-saving decisions. To the best knowledge of the authors, this is the first research that proposed TPE optimized SMOGN to improve shale gas predictions.

Investigation of Surge Phenomena in a Centrifugal Compressor by 3D-1D Coupling Approach

Abstract Critical instability, such as surge phenomena, can occur at low flow rates for a compressor system, leading to flow rate fluctuations and potential damage. To study this, traditional numerical analysis requires a comprehensive approach encompassing the compressor and its piping system. This demands extensive computational resources and time due to the complexity of capturing internal flow during surge, including separation and vortex generation. To overcome these limitations, this research proposes a computational strategy combining 3D and 1D analysis for enhanced efficiency and reduced computational demands, which employs 1D analysis for more straightforward flow regions like piping and 3D analysis for the more complex compressor section. This approach is facilitated using AMESim for 1D and STAR-CCM+ for 3D simulations, enabling seamless integration and data exchange between different dimensional analyses. The method improves computational efficiency and reasonably approximates surge phenomena compared to experimental data. The 3D and 1D coupled analysis closely aligns with experimental data, providing a viable and more effective alternative method for surge analysis in compressor systems. This methodology promises to streamline the study of compressor surge phenomena, providing a practical tool for engineers and researchers.

Uncertainty sensitivity analysis for vibration properties of composite doubly-curved shallow shells using Kriging method

This paper presents a Kriging based global sensitivity analysis (GSA) method for the frequency response of displacements of composite doubly-curved shallow shells. A unified solution is utilized to develop the dynamic vibration formulation using the First-order Shear Deformation Theory (FSDT) and the Rayleigh-Ritz method. Kriging surrogate model is employed to substitute the frequency response function (FRF) of displacements. Ten parameters including materials and geometrical dimension are considered as input uncertain variables. A variance-based GSA method for dynamic model is employed to quantify the influence of each uncertain parameter. In addition, to avoid the computational burden of Monte Carlo simulation method (MCS), the presented sensitivity indices are computed analytically based on the Kriging mode, which further improves computational efficiency. Based on the convergence studies and comparison with traditional methods, the accuracy and efficiency of the present method are validated. The results shows that the frequency response of displacements exhibits greater sensitivity to changes in width, and thickness is more influential than others in the example from this article. Finally, the presented numerical results demonstrate vibration characteristics of different types of shells and observation points, which can also serve as a reference for further study on uncertainty-propagation analysis.

A hybrid SPH model of the Navier-Stokes and Boussinesq equations for simulating wave propagation and deformation

A two-way hybrid model based on the smoothed particle hydrodynamics (SPH) method is proposed in this paper to simulate wave propagation and deformation effectively. The hybrid model is composed of two submodels: the first submodel is based on the Boussinesq equations (B-SPH) for simulating wave propagation, and the second submodel is based on the Navier-Stokes (N-S) equations (NS-SPH) for simulating wave deformation caused by interactions with structures. The exchange of information between the two submodels is realized through an open-boundary method: the results derived from the B-SPH submodel are transferred to the open boundaries of the NS-SPH submodel through a relaxation function; the information derived from the NS-SPH submodel is matched to the open boundaries of the B-SPH submodel through the continuity equation. To obtain a stable pressure field in the NS-SPH submodel, the particle shifting technique (PST) is implemented. The subsequent validation of the hybrid model includes four typical numerical examples related to solitary and regular waves. The agreement between the numerical results and the experimental data indicates that the proposed hybrid model is suitable for simulating the propagation and deformation of shallow water waves. Compared with the traditional NS-SPH model, which uses N-S equations to describe the whole fluid domain, the hybrid model has greater computational efficiency.

A sound-vibration physical-information fusion constraint-guided deep learning method for rolling bearing fault diagnosis

Although current deep learning models for bearing fault diagnosis have achieved excellent accuracy, the lack of constraint-guided learning of the physical mechanisms of real bearing failures and a physically scientific training paradigm leads to low interpretability and unreliability of intelligent fault diagnosis models. In this study, a sound-vibration physical-information fusion constraint-guided (PFCG) deep learning (DL) method is proposed, aiming at weighted fusion of sound and vibration multi-physical information into a deep learning model, to guide the DL model to learn more realistic physical laws of bearing failure. Firstly, a 15-degree-of-freedom nonlinear dynamics model of multi-stage degraded bearing failure mechanism with sound-vibration response is developed, which considers the evolutionary mechanism of bearing failure from healthy state to different stages, and utilizes a particle filtering algorithm for dynamic calibration of hidden parameters. Moreover, a lightweight DL fault diagnosis model is designed to realize the deep interaction between the physical model and the DL model through the weighted fusion of the cross-entropy loss function, physical consistency loss and uncertainty loss. Moreover, the superior diagnostic performance of the proposed sound and vibration PFCG-DL model is verified by comparing the performance fluctuations and parameter attributes of different DL benchmark models before and after being guided by physical information fusion constraints (PFCG). Eventually, the proposed PFCG-Transformer model achieves a diagnostic accuracy of 99.45% while keeping the number of parameters at only 0.62M, which significantly improves the accuracy and reduces the computational complexity by 81.5% compared to the CAME-Transformer model's 3.24 M number of parameters and 95.00% diagnostic accuracy. In addition, the test time of PFCG-Transformer is reduced to 1.02 s, which is 60.2% less than CAME-Transformer, demonstrating higher computational efficiency and real-time performance. Importantly, in terms of interpretability, the engineering interpretability and credibility of the models are further improved by visualizing the feature learning results of the vibration modal and multimodal fusion models and the sensitivity analyses of the sound-vibration response models with internal and external physical hyperparameters. Therefore, this study proposes a physical information-guided deep learning method with strong interpretability and superior performance, which provides an important reference for further research and application in the field of bearing fault diagnosis.

Forecasting two-dimensional channel flow using machine learning

Over the past decade, the integration of artificial neural networks (ANNs) has garnered significant interest, capitalizing on their ability to discern intricate patterns within data. Focused on enhancing computational efficiency, this article explores the application of ANNs in forecasting fluid-dynamics simulations, particularly for the benchmark problem of fluid flow in a two-dimensional (2D) channel. Leveraging a multilayer perceptron trained on finite volume method numerical data, for both interpolation and extrapolation estimations and various grid resolutions, our findings demonstrate the ANN's prowess as a swift and accurate surrogate for traditional numerical methods. Overall, the results of this work mark a pioneering step toward leveraging machine learning for modeling complex relationships in fluids phenomena, promising transformative advancements in computational fluid dynamics.

Removing thick clouds in landsat 8 OLI images with cloudy time-series auxiliary data

Abstract Thick cloud occlusion severely limits the subsequent application of optical remote-sensing images. Existing mainstream cloud removal methods often use cloud-free homologous temporal images to provide auxiliary information. However, due to the limitation of cloud occlusion and satellite revisit cycle, the time interval between the cloud-free auxiliary image and the target cloudy image is often too large, and the possible land cover changes during the period increase the uncertainty of effective information on the auxiliary images. In this case, more accurate reconstructions may be obtained by using temporally closer and partially cloud-contaminated images. In this paper, a deep network, convLSTM, is developed for cloud removal (i.e., CR-convLSTM), which makes full use of the complementary effective information on partially cloud-contaminated time-series images. Experiments based on simulated clouds indicated that CR-convLSTM can obtain more accurate predictions than the classical cloud removal method MNSPI. CR-convLSTM is an effective cloud removal method with high computational efficiency while ensuring predicting accuracy.

PAGE: Prototype-Based Model-Level Explanations for Graph Neural Networks.

Aside from graph neural networks (GNNs) attracting significant attention as a powerful framework revolutionizing graph representation learning, there has been an increasing demand for explaining GNN models. Although various explanation methods for GNNs have been developed, most studies have focused on instance-level explanations, which produce explanations tailored to a given graph instance. In our study, we propose Prototype-bAsed GNN-Explainer ([Formula: see text]), a novel model-level GNN explanation method that explains what the underlying GNN model has learned for graph classification by discovering human-interpretable prototype graphs. Our method produces explanations for a given class, thus being capable of offering more concise and comprehensive explanations than those of instance-level explanations. First, [Formula: see text] selects embeddings of class-discriminative input graphs on the graph-level embedding space after clustering them. Then, [Formula: see text] discovers a common subgraph pattern by iteratively searching for high matching node tuples using node-level embeddings via a prototype scoring function, thereby yielding a prototype graph as our explanation. Using six graph classification datasets, we demonstrate that [Formula: see text] qualitatively and quantitatively outperforms the state-of-the-art model-level explanation method. We also carry out systematic experimental studies by demonstrating the relationship between [Formula: see text] and instance-level explanation methods, the robustness of [Formula: see text] to input data scarce environments, and the computational efficiency of the proposed prototype scoring function in [Formula: see text].