This PDF file contains the front matter associated with SPIE Proceedings Volume 13088, including the Title Page, Copyright information, Table of Contents, and Conference Committee information.

Optical remote sensing images have become an important data source for studying the characteristics of the Earth’s surface by their wide coverage, high spatial and temporal resolution, and information richness. However, in practical applications, optical remote sensing images are often affected by cloud occlusion, resulting in loss of surface information and errors. To this end, we propose a generative adversarial network (AM-RSC-GAN) based on the attention mechanism and residual symmetric connectivity structure. AM-RSC-GAN uses a multi-scale feature fusion module and the channel spatial attention mechanism to effectively improve the network’s extraction and fusion of global and local information of the feature image and adopts the residual symmetric connectivity structure to enabling the network to achieve a good performance between effectively removing the cloud layer and retaining the original detail information. The residual symmetric connection structure enables the network to achieve a good balance between effective cloud removal and preservation of original detail information, better preserving the detail information of the original image and achieving faster convergence speed. We experimentally compare with several previous cloud removal models on the open-source RICE dataset and the WHUS2-CR dataset. The results show that AM-RSC-GAN achieves better results in both peak signal-to-noise ratio (PSNR) and structural similarity (SSIM), proving that the model has excellent cloud removal performance.

The advancement of Earth observation technologies has enabled comprehensive Earth monitoring at unprecedented levels. Single remote sensing images often fall short in practical applications due to various limitations. In response, the field of remote sensing has evolved to encompass multiple temporal phases, wavelengths and sensors, resulting in a wealth of multi-source remote sensing data. However, most studies have primarily employed single remote sensing images as their data source, with limited research on deep learning for land cover classification using multi-source mixed samples. There is limited research on the influence of network band selection strategies for different land cover categories. Research into network band selection strategies further enhances the accuracy of land cover extraction from multi-source data. Consequently, this paper presents a methodology for deep learning with multi-source mixed samples, investigating strategies for selecting optimal spectral bands for different land cover categories and ultimately aiming to extract various typical land cover classes from multi-source remote sensing data within the same geographic region. We begin by creating a multi-source remote sensing mixed sample dataset and employ channel enhancement methods to enable the model to predict data with different spectral band combinations. Subsequently, we investigate the network's selection of optimal spectral bands for different land cover categories to enhance land cover extraction accuracy. The results indicate that our approach not only improves accuracy but also enhances the model's generalization capabilities.

Large-format, information-dense satellite-based remote sensing pictures are in conflict with the restricted bandwidth for return transmission to the ground, necessitating the development of appropriate remote sensing image compression techniques. With regard to the data transmission requirements of large-format remote sensing photos, we investigate common lossless compression techniques in this study and suggest a superior approach based on ANS entropy coding. This algorithm, which is superior to Golomb entropy coding in JPEG-LS image compression algorithm in terms of simplicity, high coding efficiency, and extremely low post-compression coding redundancy, can be used to process remote sensing image entropy coding and significantly boost the performance of remote sensing image compression. Research demonstrates that when compared to the original coding scheme, the approach increases the compression ratio and efficiency, decreases the amount of data that must be downlinked, and increases transmission speed. The requirements for real-time processing of large-format remote sensing images can be met by compression ratios that are very close to the limit of lossless compression for remote sensing satellite images with less information and uniform pixel value distribution. These images also have a lower algorithm complexity and require less computing time.

As remote sensing technology continues to advance, the accuracy and quantity of remote sensing images have significantly improved. The generation of a vast amount of available data has facilitated the widespread application of various deep learning methods in the field of remote sensing data processing, such as object detection, semantic segmentation, and change detection. In the aforementioned tasks, Change detection is used to identify alterations occurring on the Earth's surface by utilizing remote sensing (RS) data. In recent years, deep learning based methods have exhibited significantly superior performance compared to traditional change detection techniques. The fundamental strategy enabling these advancements involves extracting appropriate deep learning features from input remote sensing images through various backbone networks such as VGG, ResNet, DenseNet etc . Nevertheless, the features extracted by the aforementioned backbone networks may not fully cater to the specific requirements of remote sensing image change detection tasks. Consequently, our goal is to explore the influence of features extracted by different backbone networks on change detection tasks and introduce a specialized backbone network tailored for change detection. This endeavor aims to produce features that are better suited for the of change detection. The experimental results indicate that our specifically designed feature extraction network for remote sensing image change detection outperforms traditional networks in extracting task-specific features. These features are better suited for subsequent decoder modules, enhancing the generation of image-based change detection results. At the same time, we found that when using general backbone networks for change detection, ResNet achieves the highest metric accuracy, while DenseNet has the lowest memory usage and the fastest training and testing speed. Depending on the specific task, we can choose the appropriate backbone network as needed.

With the rapid development of artificial neural network (ANN), the field of synthetic aperture radar (SAR) target recognition has witnessed significant progress. However, due to the poor interpretability and ease of being affected by speckle noise, it brings challenges to ANN for SAR target recognition. Spiking Neural Network (SNN) has emerged as the third-generation neural network architecture and presents promising prospects for various applications. This study aims to explore the performance of SNN in SAR target recognition. In our experiments, we achieved comparable performance to conventional neural networks by utilizing directly trained SNN. This indicates the effectiveness of SNN in coping with SAR target recognition tasks. Moreover, we investigated the impact of different spiking encoders on SAR target recognition. Specifically, we compared the performance of SNN using the Poisson encoder and utilizing the first layer of the SNN as an encoder. This comparison provides valuable insights into the optimal coding strategy for SNN-based SAR target recognition. Additionally, we examined the robustness of SNNs in the presence of strong speckle noise. Our findings demonstrate that SNN can maintain good performance under the influence of strong speckle noise. The outcomes of this research shed light on the potential of SNN as a powerful tool for SAR target recognition. Future studies can focus on exploring SNN’s applicability to SAR Interpretation.

Accurate building footprint extraction from optical remote sensing images remains challenging due to the diverse appearance and complex scenarios. Although recent deep learning-based methods have been shown to greatly improve the accuracy of building footprint extraction, vanilla deep networks still suffer from ambiguous predictions of edge pixels. The building edge contains abundant location and shape information, which is important for downstream applications such as building positioning and area measurement. Therefore, the problem of inaccurate edge prediction needs to be resolved urgently. To this end, we propose a novel edge-guided network (EGNet) that makes ample use of the edge prior in an end-to-end manner. First, an edge extraction module (EEM) is proposed to extract the building edge map. Then, an edge-guidance module (EGM) is designed to utilize the edge map to guide each encoder block in extracting edge-related features. Furthermore, a multi-scale context aggregation module (MCAM) is built to enhance the feature representation by aggregating contextual semantics with different receptive fields. EGNet can effectively mine edge semantics and guide the representation learning of boundaries, achieving 75.21% and 91.16% IoU on the Massachusetts and WHU datasets, respectively. Experimental results demonstrate that the proposed EGNet has a certain superiority in both accuracy and efficiency compared with current state-of-the-art (SOTA) methods.

In order to accurately identify large-scale sugarcane plantations, pre-processed Sentinel-2A remote sensing imagery data from three different time periods were used in this study. The vegetation index features NDVI, BI2 and S2REP were calculated and the optimal texture feature, homogeneity, was selected using principal component analysis and the grey level co-occurrence matrix. A random forest algorithm was then used to construct identification models with different combinations of features, including single time period and single feature, single time period and multiple features, multiple time periods and single feature, and multiple time periods and multiple features. The experiment was conducted in Nansha District, Guangzhou, and the results showed that the fusion of NDVI, BI2, homogeneity and temporal features provided the best identification results for sugarcane planting areas, with clear object boundaries and effective control of salt and pepper noise in the images. The overall accuracy reached 0.974 and the kappa coefficient reached 0.931. In the result validation, the identification accuracy for Nansha District, Guangzhou was 81.6%. The experimental errors were also analysed and the method was found to be suitable for determining the area under sugarcane cultivation in the Pearl River Delta region, providing valuable reference information for the agricultural sector in monitoring the growth of sugarcane, estimating yields and forecasting prices.

Approximately 50% of the Earth’s surface is covered by clouds. Optical remote sensing satellites encounter challenges in capturing clear ground images due to the difficulty of visible photoelectric magnetic waves penetrating clouds. Therefore, cloud detection is an essential and basic step in the processing and application of optical satellite remote sensing images. The traditional threshold-based cloud detection methods utilize spectral information of images to set thresholds, which is sensitive and less adaptable to varying imaging conditions. To enhance cloud detection performance, this paper proposes a deep learning method that combines residual network module and pyramid structure of dilated convolution. The method employs an encoder-decoder structure where the residual convolution module replaces conventional convolution kernels, reducing parameter quantity while enhancing feature expression capabilities. Furthermore, a pyramid structure of dilated convolution is introduced between the encoder and decoder to improve the acquisition of global information and reduce misjudgment of cloud pixels. In this work, ablation experiments are conducted to validate the reliability of the proposed network model. To evaluate the effectiveness of the proposed method, GaoFen-1 remote sensing image data was used for experimental verification. The results indicated that compared to traditional methods, the proposed approach achieves satisfactory cloud detection results in various surface types, including barren land, forests, grasslands/crops, and wetlands, while having a smaller model size and parameter quantity.

Removing stripe noise is a fundamental task in remote sensing image processing, which is of great significance for improving image quality and subsequent applications. In this paper, an adaptive strip noise removal model is proposed with the spatial characteristics. Firstly, an adaptive weight function is constructed using local absolute differences to adaptively control the constraint intensity of the penalty term at different pixel points in the adaptive strip noise removal model. Secondly, L¹ norm is used to constrain the local smoothness along the direction of the strip, maintaining the obvious smoothness characteristics of the strip noise in its extension direction, while L² norm is used to restrict the image grayscale. Finally, the extended split Bregman iteration method and alternating minimization method is used to optimize the proposed image destriping model. Extensive experiments on both the synthetic and real remote sensing images validate that the proposed model can effectively remove the stripe noise and preserve more fine scale details.

In recent years, high-resolution remote sensing image segmentation has become a key task involving technology in many fields, mainly used to accurately extract target information from remote sensing images, and widely used in land detection, coverage classification, etc. These images are characterized by high resolution and large scale, and the overall segmentation requires high performance of hardware devices, the downsampling method usually used is easy to affect the segmentation quality, cropping and slicing the image will lead to the lack of edge information, and the problems of category homogenization, changes in complex scenes, and noise and occlusion make the segmentation challenging. Therefore, in this paper, we propose An iterative attention context fusion network (IACFNet) based on the attention mechanism as well as information fusion at different scales, which iteratively calculates the attention module weights connecting low and high features and refines the problem of spatial information loss, utilizing multiscale segmentation for information complementation, and utilizing an improved boundary loss function to precisely define boundary instances. Our proposed method obtains a performance improvement of about 2.8% and 1.5% in the mean intersection over union (mIoU) metric compared to the current state-of-the-art methods, respectively.

Aiming at the demand for target detection of typical ground and object backgrounds in hyperspectral earth observation, this project solves the problems of insufficient sample size and low fidelity of target environment scene in intelligent learning, training and testing of a novel optoelectronic detection system, breaks through the key technologies of modeling of hyperspectral imaging simulation mechanism, modeling of hyperspectral reflectance/radiation characteristics of typical target background, computation of hyperspectral atmospheric transmission characteristics, and modeling of hyperspectral optical detectors, etc., and organically integrates remote sensing imaging and spectral measurement technologies. Through simulation, it simultaneously acquires spatial and spectral radiation information data from typical target backgrounds, thus providing foundational data support for the design of target detection algorithms based on hyperspectral data.

In spectral scene simulation applications, it is necessary to recognize and detect desert satellite images and satellite images containing water. However, previous network models such as AlexNet, VggNet, and GoogleNet have gradually fallen behind the trend of the times. Compared with the latest network models, they not only have slow training speeds but also do not achieve high accuracy. Therefore, this study adopts a series of the latest network models to solve this problem and compare and analyze the results.

Ultrasonic nondestructive testing has been widely used in industry due to its various advantages. However. with the increasing demand for non-destructive testing accuracy, low-resolution ultrasonic images can easily lead to misidentification of defects. How to improve the resolution of ultrasonic images has become a key issue restricting development of the non-destructive testing industry. Therefore. this paper proposes a super-resolution model for ultrasonic NDT images consisting of up and down sampling layers and deep residual networks. The model learns the degraded features of the image through the up and down sampling layers, and learns the intrinsic features of the image through the deep residual network, so as to learn the feature information of the image more completely to complete the super-resolution task of the image. Experimental results on ultrasound image datasets also validate the effectiveness of the model, outperforming other models in peak signa-to-noise ratio (PSNR) and structural similarity (SSIM) metrics.

Conventional imaging detection relies on electrical modulation, but it grapples with limitations in speed and adaptability within complex environments. Intelligent imaging detection, capable of manipulating micro-nano-lightfields generated by imaging objective lenses, is essential. The all-optical deep diffractive neural network (D²NN) offers a potential solution, enabling swift spatial information transformation and recognition. However, existing D²NN versions are vulnerable to optical disturbances due to their reliance on computer-based backpropagation. This paper introduces a robust diffractive neural network (SRNN) mathematical model designed to mitigate the impact of structural errors and lightwave frequency shifting without necessitating hardware modifications. The SRNN model incorporates Gaussian noise at various levels during training to emulate the error distribution of the phase mask, augmenting the network's resilience to hardware errors. Weight-Noise-Injection training assists the network in locating an error-insensitive region, thereby enhancing its robustness.

The inspection and recognition of readings on pointer instruments are important tasks in power inspection. Due to the diversity and the complex dial information of pointer instruments, the adaptability and low accuracy of existing methods are not enough. In this paper, we proposed a pointer instrument reading recognition algorithm based on local fast matching. Firstly, YOLOv8 is used to locate and segment the instrument in the inspection image. The positions and values of key scale are marked in the template image. Based on the key local information, a mapping relationship is established between the instrument location region in the inspection image and the template image. Then, the Hough transform is used to detect the position of the pointer, map it to the template image, and finally calculate the instrument reading based on the relative positions of the key graduations on the template image and the mapped pointer position. Experimental results demonstrate that the proposed method achieves a relative error within 5% for pointer instrument reading recognition in indoor substations.

With the maturity of big data analysis technology, the emergence of cloud learning platforms oriented to distributed computing and other computing has promoted the reform and innovation of modern education technology. The modern teaching environment has changed. Course teaching gradually changed from traditional closed teaching to open integration based on cloud platform. The development method of cloud platform teaching system mainly includes the establishment and maintenance of the backend database, the intermediate business is the processing of the album and the display and interaction of the front-end interface data. The system consists of three roles: students, teachers, and administrators, and realises a number of functions including student registration, information query, course learning, online examination, and interactive message. Road through the use of the conclusion is to show that the system has a good user interface, can effectively reduce the teacher's workload, effectively improve the efficiency of student learning, but also to help students learn independently, to strengthen the teacher-student communication, to cultivate better learning nitrogen perimeter. The system development tools for MyEclipse integrated development environment, JAVA language as the main programming, SQL Sever as the background database, JSP written user interface, through the background analysis and then synthesise the data displayed to the front-end, can be deployed in the AliCloud platform, the purchase of servers, open tomcat and the server port, you can move through the IP ask. Most of the interactive learning in traditional teaching is concentrated in the offline classroom because of the geographical location limitation, and the offline communication can only be carried out by the same classmates, but in the online classroom of the cloud platform, it is not limited to the geographical location of the class, and it is possible to study anytime and anywhere on the network, and it is also possible to leave a message and comment under the course, and it is possible to learn and discuss with each other across the classes and even across the schools, and it is possible to supervise and learn and progress with everyone together. Learning and progressing together. For most of the system, this system can provide students with cloud learning process, but also provide students and teachers to register, login function, which includes a variety of tools module, similar to the background module of the course management, the foreground module of the course online learning, learning notes, etc., the cloud education system can provide timely feedback on students' learning, so that the students' learning efficiency and gain greatly improved, that is, for the time and some of the Unsolvable difficult problems are greatly improved and solved. Cloud platform teaching system is an online system that can provide great help for everyone's learning and life online learning.

This study examines the factors influencing and predicts the cultural communication on social media, using the Palace Museum's Weibo account as a case study. Machine learning techniques are employed to analyze data including microblog content and user interactions. Factors considered include content diversity, sentiment polarity, topic popularity, and user engagement behaviors. Predictive models achieve 85% accuracy in determining the influence of content diversity, sentiment polarity, and topic popularity. User engagement behaviors, such as reposts and comments, positively correlate with dissemination effectiveness. Findings emphasize the importance of diverse content and positive sentiment in enhancing dissemination, as well as the significance of user interactions. Practical implications for cultural institutions in leveraging social media platforms are highlighted, emphasizing optimization of content diversity, sentiment expression, and encouraging user engagement.

To a certain extent, foreign objects on road surfaces can cause damage to vehicles and increase the risk of accidents, which is even more immeasurably harmful in such special scenarios as airports. Effective detection of foreign objects on road surfaces is of great significance to airport flight safety and road safety maintenance. Due to the traditional watershed segmentation method in the recognition of foreign objects, there are obvious over-segmentation, noise complexity and so on. Therefore, this paper proposes an improved watershed segmentation method based on Lab-color space using image filtering techniques and morphological operations to initially separate the image foreground and background, combined with Adaptive Gaussian thresholding, labeling method for detection. This effectively solves the problem of over-segmentation, can better suppress image noise, and accurately detects foreign objects on road surfaces.

To effectively remove noise from the point clouds obtained for road surface defect detection while preserving the feature information, a bilateral filtering algorithm based on curvature features is proposed for denoising road surface defect point clouds. First, divide the noise points in the point cloud into several regions based on their distribution characteristics and curvature features. Then use statistical filtering to remove noise points in the obvious noise area and the flat area. Finally, an improved bilateral filtering algorithm is applied to smooth the noise points in the complex detail area.

In the line structured light three-dimensional measurement system, high-precision laser stripe centerline extraction is the key to improving measurement accuracy. The laser line extraction technology based on neural network can automatically detect and extract laser lines with high accuracy and robustness. In this paper, a BP neural network containing one hidden layers is constructed, and the centerline obtained through feature extraction of the network can effectively overcome the influence of noise interference. Compared with the gray-gravity method and the Steger algorithm, the method in this paper has the advantage of high accuracy and can meet the requirements for extracting the center of complex light strips.

In outdoor complex lighting environments, laser stripe extraction becomes challenging due to the possibility of strong exposures in the external environment as well as reflective interference on the road surface. In order to overcome these problems, this paper proposes a multi-feature structured light extraction algorithm based on seed neighborhoods, which utilizes multiple seed points and comprehensively considers the laser stripe features, especially the morphological features of the optical stripes in their surrounding neighborhoods, so as to obtain the approximate distribution location of the real laser stripe. Finally, the laser streak extraction is performed by the gradient difference of the optical streak boundary as well as its own brightness. The experimental results show that the method is still effective in extracting laser lines even in high exposure environments where there are interferences such as reflections on the road surface.

Monte Carlo algorithm’s performance change is studied when collision method and path length method are used as the energy beam tracking method for calculating radiative heat transfer under the same physical property condition.The results show that when the surface emissivity, scattering albedo and dielectric absorptivity are fixed at 0.5 and the average optical thickness of a single mesh is less than 0.1, the surface performance factor of the collision method is stable, but the spatial performance factor decreases with the increase of the average optical thickness, and these changes do not change much with the increase of the number of beams.Under the same conditions, the surface performance factor and the space performance factor decrease with the increase of the average optical thickness.Under the same conditions, the surface performance factor and the space performance factor of the path length method decrease with the increase of the average optical thickness.

This study investigates the impact of angular mesh dispersion, scattering albedo, and surface emissivity on the computational accuracy of the Monte Carlo method for solving the radiation heat transfer problem with participating medium in a three-dimensional cylindrical coordinate system under isothermal radiation equilibrium conditions. The relationship between the surface emissivity and the minimum number of beams is quantitatively analyzed by collating the data, and the maximum allowable optical thickness under a certain number of bundles is obtained.

The size of the gap in the switch machine can reflect the working status of the switch machine and is related to the driving safety and driving efficiency of railway vehicles.In the traditional switch machine inspection and maintenance process, manual measurement is mainly used to complete the inspection of the gap size. This article discusses a gap monitoring system based on image recognition.

Like fingerprints, facial information and other biological characteristics, human voice also carries the physiological characteristics of living things. It is unique, stable personal information that cannot be stolen or lost. The speaker's voiceprint can contain changes that remain unchanged, so these features make the voiceprint features deeper, more elusive, and more difficult to forge, making the authentication stronger and safer. Based on the basic characteristics of human voiceprints, this paper designs a voiceprint recognition system based on neural networks. In order to enable the network to better use original data to obtain output results, time and frequency domain masking methods are used for data enhancement. In the network part, the encoder-decoder method uses the transformer architecture to achieve end-to-end data processing, and uses the triplet loss function to evaluate and optimize the parameters within the neural network to improve the prediction accuracy of the model. Modeling experiments were conducted on the LibriSpeech and CN-Celeb datasets, respectively. The system realizes human voiceprint recognition end-to-end based on deep learning and has been tested to meet the design needs.