Search

Zenith tropospheric delay (ZTD), consisting of zenith hydrostatic delay (ZHD) and zenith wet delay (ZWD), is a significant contributor to errors in precise positioning using the global navigation satellite system (GNSS) precise point positioning (PPP) and real-time kinematic techniques. Accurate and timely predictions of ZTD on a global scale are crucial for enhancing GNSS positioning accuracy and expediting convergence. This study proposes an innovative global tropospheric prediction model that leverages long short-term memory (LSTM) neural networks, aiming to achieve both high precision and long-term prediction capability for ZTD. The experimental data utilized were sourced from the Vienna Mapping Functions 3-Optimized zenith total delay (ZTD) dataset. This study delves further into the analysis of ZTD residuals by extracting periodic signals. The ZTD residuals were then utilized to train a modified LSTM neural network model, enabling the prediction of global residuals. The final ZTD predictions were obtained by combining the modified LSTM ZTD residual forecast component with the ZTD periodic component. Our results demonstrate that the average root-mean-square error (RMSE) of the modified LSTM-ZTD model in 2020 was 1.44 cm. In addition, the average RMSE of the forecasted ZTD during spring, summer, autumn, and winter was found to be 1.43 cm, 1.47 cm, 1.56 cm, and 1.36 cm, respectively. Through the integration of the LSTM neural network and the ZTD periodic signal extracted using a physical algorithm, this work has successfully enhanced the accuracy and time span of ZTD forecasts on a global scale.

Brain-computer interfaces (BCIs) are anticipated to improve the efficacy of rehabilitation for people with motor disabilities. However, applying BCI in clinical practice is still a challenge due to the great diversity of patients. In the current study, a novel action observation (AO) based BCI was proposed and tested on stroke patients. Ten non-hemineglect patients and ten hemineglect patients were recruited. Four AO stimuli were designed, each presenting a decomposed action to complete the reach-and-grasp task. EEG data and eye movement data were collected. Eye movement data was utilized to analyze the reasons for individual differences in BCI performance. Task discriminative component analysis was utilized to perform online target detection. The results showed that the designed AO-based BCI could simultaneously induce steady state motion visual evoked potential (SSMVEP) from the occipital region and sensory motor rhythm from the sensorimotor region in stroke patients. The average online detection accuracy among the four AO stimuli reached 67% within 3 s in the non-hemineglect group, while the accuracy only reached 35% in the hemineglect group. Gaze metrics showed that the average total duration of fixations during the stimulus phase in the hemineglect group was only 1.31 s ± 0.532 s which was significantly lower than that in the non-hemineglect group. The results indicated that hemineglect patients have difficulty gazing at the AO stimulus, resulting in poor detection performance and weak desynchronization in the sensorimotor region. Furthermore, the degree of neglect is inversely proportional to the target detection accuracy in hemineglect stroke patients. In addition, the gaze metrics associated with cognitive load were significantly correlated with the accuracy in non-hemineglect patients. It indicated the cognitive load may affect the AO-based BCI. The current study will expedite the clinical application of AO-based BCI.

In order to address the problem of incomplete data set of real seismic impact of cultural relics in collections, we propose a small sample set oriented center-of-mass positioning method to solve the center-of-mass attribute of cultural relics ontology. First, the method combines the cultural relics image data to construct a deep feature fusion cultural relics ontology recognition model, which extracts and learns the cultural relics image features through multi-layer feature extraction, global information perception and ontology fusion recognition, so as to realize the efficient and accurate recognition of the cultural relic’s ontology. Then, the parameter synchronous migration learning method is designed to train and initialize the model, and the data migration strategy of semi-supervised learning is used to retrain the model using the public dataset and the cultural relic image data, to achieve the fine-tuning of the model parameters to improve the accuracy of the cultural relic ontology recognition. Finally, the center-of-mass positioning method is designed to integrate the attention mechanism with the quality of relics and other ontological attributes, and the segmentation of the relic’s ontology region is achieved by rectangular stereo fitting and the calculation of the center-of-mass points. At the same time, the fusion self-attention mechanism is used to adjust the regional center-of-mass weights to achieve weighted positioning of the center-of-mass position of the cultural relics. The experimental results show that the present method achieves the best Dice, Acc and MIoU metrics in comparison with various classical models, with improvements of 8.3%, 5.1% and 3.4% respectively. The overall center-of-mass offset of cultural relics is less than 2%, which can achieve accurate identification and center-of-mass positioning of the cultural relics body, fully improve the seismic impact dataset of cultural relics, and enhance the ability of preventive protection of cultural relics against earthquakes.

Hyperspectral (HS) pansharpening refers to fusing low spatial resolution HS (LRHS) images with the corresponding panchromatic (PAN) images to create high spatial resolution HS (HRHS) images. Most of the existing HS pansharpening methods overlook the spatial and spectral imbalance of the ground objects of different types in the observed scenes. To address the dilemma, in this article we develop a novel tree-structured neural network (Tree-SNet) to form an adaptive spatial-spectral processing for HS pansharpening. The Tree-SNet method maps a convolutional neural network (CNN) onto a hierarchical tree structure, where routing nodes automatically tune the data distributed to tree paths, which is adaptive to the local characteristics of the data, while spatial enhancement (SpatE) and spectral enhancement (SpecE) modules are dynamically performed in the tree paths to further strengthen the adaptive processing. The proposed Tree-SNet is evaluated on several datasets, and the experimental results verify its superiority.

Occlusion handling is crucial for improving the performance of convolutional neural networks (CNNs) in real-world remote sensing images, which are often captured in complex and unconstrained environments. In particular, occlusion scene classification has received significant attention due to its usefulness in various remote sensing tasks. However, existing methods are limited by their dependence on close-world learning, which assumes that all test cases are included in the training set. This is problematic because occlusion is too complex to be thoroughly annotated. To address this issue, we propose a novel contrastive learning-based CNN that can classify out-of-distribution occluded scenes without the need for occlusion annotation. Our approach uses a two-branch subnetwork to learn representations of unoccluded anchor images and occlusion-augmented images. We then employ cascade supervised contrastive learning to make the network's representations invariant to occlusion. Unlike standard contrastive learning, our method leverages category information to avoid incompact intraclass distribution and uses a cascade strategy to hierarchically learn occlusion-invariant representations. Finally, we use a multilayer perceptron to classify the learned representations and assess representation quality. We evaluate our method on the UAVDT dataset and two simulated datasets, and the results demonstrate that our approach accurately characterizes occluded objects and achieves more precise classification in occlusion scenarios.

Due to the popularity of 5G connectivity and The Internet of Things sensors, deep learning algorithms are being extended to edge devices. Compared with AI(Artificial Intelligence) cloud platforms, the deployment of deep neural networks on edge devices must focus on low power consumption, low latency, stability and reliability. In recent years, the development of lightweight deep neural network architecture has provided a basis for the deployment of deep neural networks on edge devices. However, the shortcomings of deep neural networks, such as overconfidence, vulnerability to adversarial attack, and easy over fitting when samples are insufficient, still limit their applications in many fields. One of the ways to compensate for these defects is to use deep ensemble. An ordered aggregation-based ensemble selection algorithm is proposed, which uses soft-margin as the importance assessment metric to take full advantage of the diversity and complementarity of lightweight deep neural networks obtained from different initialization training, so as to improve the overall performance of multiple edge devices. The experimental results show that this algorithm has a significant improvement in generalization performance compared with random ensemble and ordered aggregation algorithms based on accuracy or diversity, and provides a new complementary idea for the deployment of lightweight deep neural networks on edge devices.

Many patients with genetic syndromes have special facial features, which boast significant potential value for clinical diagnosis. Deep learning and computer vision technology can be employed to diagnose genetic diseases by analyzing facial features of patients. As a matter of fact, the application of deep learning technology in the area of genetic diseases is confined owing to the difficulties of patient data acquisition. This study develops BioFace, a deep learning framework that can recognize multiple genetic diseases facial attributes based on limited datasets. BioFace is a deep neural network architecture designed premised on Resnet. To increase the weight of effective features and weaken the weight of invalid or unobvious features during extraction of facial features, we add Squeeze-and-Excitation (SE) blocks in the network. In combination with this network architecture, we designed a cross-loss training method based on transfer learning. This method can transfer the ability learned from the task of face identification to the task of recognition of genetic diseases facial attribute, and improve the inter-class distance of different genetic diseases and the intra-class distance of similar genetic diseases simultaneously. These render it possible for deep learning to be applied to recognition of multiple genetic diseases facial attribute with very small amount of data. In this research, we tested 10 syndromes with our framework and the Top-1 accuracy was 93.5%, which is the state-of-the-art in multiple genetic syndromes recognition research. In practical clinical applications, our framework and methods can be extended to the disease identification of more small datasets, potentially offering valuable assistance for the auxiliary clinical application of genetic diagnosis and other related genetic research.

Medium spatial resolution surface reflectance image series from the combination of Landsat-8 Operational Land Imager and Sentinel-2 Multispectral Imager observations have great importance to the land surface monitoring tasks, for which great efforts have been paid for blending the two data. However, most of the efforts focus on the image series with spatial resolution of 30 m, which cannot meet the data demand of some applications. Therefore, it is necessary to fuse Landsat-8 and Sentinel-2 images to provide 10-m image series. Currently, there are three means to achieve that, including the area-to-point regression kriging fusion approach (ATPRK), spatiotemporal fusion methods, and deep-learning-based fusion models. However, the ATPRK and spatiotemporal fusion methods suffer from the limited fusion performance, while the deep-learning-based fusion models are hardware dependent, i.e., requiring the graphics processing units, which may not be satisfied sometimes. To address these issues, in this article, we develop a new pixel-wise local normalization-based fusion method (LN-FM) for fusing Sentinel-2 and Landsat-8 images. The newly proposed LN-FM is compared to the ATPRK and three representative spatiotemporal fusion methods in experiments, which use imagery collected from both a rural area and an urban area. The experimental results demonstrate that the newly developed LN-FM exhibits excellent qualitative and quantitative performance, as well as remarkable spatial, spectral, and pixel distribution fidelity. Furthermore, this approach is fast, which may improve its applicability

In recent years, convolutional neural network (CNN) algorithms promote the development of stereo matching and make great progress, but some mismatches still occur in textureless, occluded and reflective regions. In feature extraction and cost aggregation, CNNs will greatly improve the accuracy of stereo matching by utilizing global context information and high-quality feature representations. In this paper, we design a novel end-to-end stereo matching algorithm named Multi-Attention Network (MAN). To obtain the global context information in detail at the pixel-level, we propose a Multi-Scale Attention Module (MSAM), combining a spatial pyramid module with an attention mechanism, when we extract the image features. In addition, we introduce a feature refinement module (FRM) and a 3D attention aggregation module (3D AAM) during cost aggregation so that the network can extract informative features with high representational ability and high-quality channel attention vectors. Finally, we obtain the final disparity through bilinear interpolation and disparity regression. We evaluate our method on the Scene Flow, KITTI 2012 and KITTI 2015 stereo datasets. The experimental results show that our method achieves state-of-the-art performance and that every component of our network is effective.

Hyperspectral (HS) pansharpening aims at fusing a low-resolution HS (LRHS) image with a panchromatic image to obtain a full-resolution HS image. Most of the existing HS pansharpening approaches are usually based on traditional multispectral pansharpening techniques, which are not especially tailored for two inherent challenges of the HS pansharpening, i.e., much wider spectral range gap between the two kinds of images and having to recover details in many continuous spectral bands simultaneously. In this article, we develop new spectral-fidelity convolutional neural networks (called HSpeNets) for HS pansharpening to keep the fidelity of a pansharpened image to its true spectra as much as possible. Our methods particularly focus on the decomposability of HS details, accordingly synthesizing these details progressively, and meanwhile introduce a spectral-fidelity loss. We give theoretical justifications and provide detailed experimental results, showing the superiorities of the proposed HSpeNets with regard to other state-of-the-art pansharpening approaches.

Today’s Internet is vulnerable to numerous attacks, including source spoofing, distributed denial of service, prefix hijacking, and route forgery. Network-layer accountability is considered as an effective deterrence tool which can be used to address these attacks. Much research has been devoted to improving network-layer accountability of today’s Internet. In this paper, we first investigate the state-of-the-art network-layer accountability research and summarize a general definition of network-layer accountability. Next, we propose a network-layer accountability framework and present a taxonomy of network-layer accountability protocols according to accountability granularity. Furthermore, we compare these protocols and discuss their pros and cons mainly from accountability function, deployability, and security. Finally, some open research questions are emphasized for directing future designs.