Catálogo de publicaciones - revistas

<jats:title>Abstract</jats:title><jats:p>Recent studies have demonstrated that deep learning‐based stereo matching methods (DLSMs) can far exceed conventional ones on most benchmark datasets by both improving visual performance and decreasing the mismatching rate. However, applying DLSMs on high‐resolution satellite stereos with broad image coverage and wide terrain variety is still challenging. First, the broad coverage of satellite stereos brings a wide disparity range, while DLSMs are limited to a narrow disparity range in most cases, resulting in incorrect disparity estimation in areas with contradictory disparity ranges. Second, high‐resolution satellite stereos always comprise various terrain types, which is more complicated than carefully prepared datasets. Thus, the performance of DLSMs on satellite stereos is unstable, especially for intractable regions such as texture‐less and occluded regions. Third, generating DSMs requires occlusion‐aware disparity maps, while traditional occlusion detection methods are not always applicable for DLSMs with continuous disparity. To tackle these problems, this paper proposes a novel DLSM‐based DSM generation workflow. The workflow comprises three steps: pre‐processing, disparity estimation and post‐processing. The pre‐processing step introduces low‐resolution terrain to shift unmatched disparity ranges into a fixed scope and crops satellite stereos to regular patches. The disparity estimation step proposes a hybrid feature fusion network (HF<jats:sup>2</jats:sup>Net) to improve the matching performance. In detail, HF<jats:sup>2</jats:sup>Net designs a cross‐scale feature extractor (CSF) and a multi‐scale cost filter. The feature extractor differentiates structural‐context features in complex scenes and thus enhances HF<jats:sup>2</jats:sup>Net's robustness to satellite stereos, especially on intractable regions. The cost filter filters out most matching errors to ensure accurate disparity estimation. The post‐processing step generates initial DSM patches with estimated disparity maps and then refines them for the final large‐scale DSMs. Primary experiments on the public US3D dataset showed better accuracy than state‐of‐the‐art methods, indicating HF<jats:sup>2</jats:sup>Net's superiority. We then created a self‐made Gaofen‐7 dataset to train HF<jats:sup>2</jats:sup>Net and conducted DSM generation experiments on two Gaofen‐7 stereos to further demonstrate the effectiveness and practical capability of the proposed workflow.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Impact craters are geomorphological features widely distributed on the lunar surface. Their morphological parameters are crucial for studying the reasons for their formation, the thickness of the lunar regolith at the impact site and the age of the impact crater. However, current research on the extraction of multiple morphological parameters from a large number of impact craters within extensive geographical regions faces several challenges, including issues related to coordinate offsets in heterogeneous data, insufficient interpretation of impact crater profile morphology and incomplete extraction of morphological parameters. To address the aforementioned challenges, this paper proposes an automatic extraction method of morphological parameters based on the digital elevation model (DEM) and impact crater database. It involves the correction of heterogeneous data coordinate offset, simulation of impact crater profile morphology and various impact crater morphological parameter automatic extraction. And the method is designed to handle large numbers of impact craters in a wide range of areas. This makes it particularly useful for studies involving regional‐scale impact crater analysis. Experiments were carried out in geological units of different ages and we analysed the accuracy of this method. The analysis results show that: first, the proposed method has a relatively effective impact crater centre position offset correction. Second, the impact crater profile shape fitting result is relatively accurate. The <jats:italic>R</jats:italic>‐squared value (<jats:italic>R</jats:italic><jats:sup><jats:italic>2</jats:italic></jats:sup>) is distributed from 0.97 to 1, and the mean absolute percentage error (<jats:italic>MAPE</jats:italic>) is between 0.032% and 0.568%, which reflects high goodness of fit. Finally, the eight morphological parameters automatically extracted using this method, such as depth, depth–diameter ratio, and internal and external slope, are basically consistent with those extracted manually. By comparing the proposed method with a similar approach, the results demonstrate that it is effective and can provide data support for relevant lunar surface research.</jats:p>

<jats:title>Abstract</jats:title><jats:p>The dynamic measurement of position and attitude information of a long‐distance moving object is a common demand in ground testing of aerospace engineering. Due to the movement from far to near and the limitations of camera resolution, it is necessary to use multi‐binocular cameras for segmented observation at different distances. However, achieving accurate and continuous position and attitude estimation is a challenging task. Therefore, this paper proposes a dynamic monitoring technique for long‐distance movement based on a multi‐binocular videogrammetric system. Aiming to solve the problem that the scale in images changes constantly during the moving process, a scale‐adaptive tracking method of circular targets is presented. Bundle adjustment (BA) with joint segments using an adaptive‐weighting least‐squares strategy is developed to enhance the measurement accuracy. The feasibility and reliability of the proposed technique are validated by a ground testing of relative measurement for spacecraft rendezvous and docking. The experimental results indicate that the proposed technique can obtain the actual motion state of the moving object, with a positioning accuracy of 3.2 mm (root mean square error), which can provide a reliable third‐party verification for on‐orbit measurement systems in ground testing. Compared with the results of BA with individual segments and vision measurement software PhotoModeler, the accuracy is improved by 45% and 30%, respectively.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Accurate and efficient registration of unmanned aerial vehicle light detection and ranging (UAV‐lidar) and terrestrial lidar (T‐lidar) data is crucial for forest structure parameter extraction. This study proposes a novel method based on a starburst pattern for the automatic registration of UAV‐lidar and T‐lidar data in forest scenes. It employs density‐based spatial clustering of applications with noise (DBSCAN) for individual tree identification, constructs starburst patterns separately from both lidar sources, and utilises polar coordinate rotation and matching to achieve coarse registration. Fine registration is achieved using the iterative closest point (ICP) algorithm. Experimental results demonstrate that the starburst‐pattern‐based method achieves the desired registration accuracy (average coarse registration error of 0.157 m). Further optimisation with ICP yields slight improvements with an average fine registration error of 0.149 m. Remarkably, the proposed method is insensitive to the individual tree detection number when exceeding 10, and the tree position error has minimal impact on registration accuracy. Furthermore, our proposed method outperforms two existing methods in T‐lidar and UAV‐lidar registration over forest environments.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Building change detection has various applications, such as urban management and disaster assessment. Along with the exponential growth of remote sensing data and computing power, an increasing number of deep‐learning‐based remote sensing building change detection methods have been proposed in recent years. Objectively, the overwhelming majority of existing methods can perfectly deal with the change detection of low‐rise buildings. By contrast, high‐rise buildings often present a large disparity in multitemporal high‐resolution remote sensing images, which degrades the performance of existing methods dramatically. To alleviate this problem, we propose a disparity‐aware Siamese network for detecting building changes in bi‐temporal high‐resolution remote sensing images. The proposed network utilises a cycle‐alignment module to address the disparity problem at both the image and feature levels. A multi‐task learning framework with joint semantic segmentation and change detection loss is used to train the entire deep network, including the cycle‐alignment module in an end‐to‐end manner. Extensive experiments on three publicly open building change detection datasets demonstrate that our method achieves significant improvements on datasets with severe building disparity and state‐of‐the‐art performance on datasets with minimal building disparity simultaneously.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Unmanned aerial vehicle light detection and ranging (UAV‐lidar) and unmanned aerial vehicle (UAV) photogrammetry are currently commonly used surface monitoring technologies. Previous studies have used the two technologies interchangeably and ignored their correlation, or only compared them on a single product. However, there are few quantitative assessments of the differences between these two techniques in monitoring surface deformation and prediction of their application prospects. Therefore, the paper compared the differences between the digital elevation model (DEM) and subsidence basins obtained by the two techniques using Gaussian analysis. The results indicate that the surface DEMs obtained by both the techniques exhibit a high degree of similarity. The statistical analysis of the difference values in the <jats:italic>z</jats:italic> direction between the two DEMs follows a Gaussian distribution with a standard deviation of less than 0.36 m. When comparing the surface subsidence values monitored by the two techniques, it was found that UAV‐lidar was more sensitive to small‐scale deformation, with a difference range of 0.23–0.44 m compared to photogrammetry. The conclusion provides valuable information regarding the utilisation of multisource monitoring data.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Transformer has achieved outstanding performance in hyperspectral image classification (HSIC) thanks to its effectiveness in modelling the long‐term dependence relation. However, most of the existing algorithms combine convolution with transformer and use convolution for spatial–spectral information fusion, which cannot adequately learn the spatial–spectral fusion features of hyperspectral images (HSIs). To mine the rich spatial and spectral features, a two‐branch global spatial–spectral fusion transformer (GSSFT) model is designed in this paper, in which a spatial–spectral information fusion (SSIF) module is designed to fuse features of spectral and spatial branches. For the spatial branch, the local multiscale swin transformer (LMST) module is devised to obtain local–global spatial information of the samples and the background filtering (BF) module is constructed to weaken the weights of irrelevant pixels. The information learned from the spatial branch and the spectral branch is effectively fused to get final classification results. Extensive experiments are conducted on three HSI datasets, and the results of experiments show that the designed GSSFT method performs well compared with the traditional convolutional neural network and transformer‐based methods.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Vision Transformers (ViTs) are exceptional at vision tasks. However, when applied to remote sensing images (RSIs), existing methods often necessitate extensive modifications of ViTs to rival convolutional neural networks (CNNs). This requirement significantly impedes the application of ViTs in geosciences, particularly for researchers who lack the time for comprehensive model redesign. To address this issue, we introduce the concept of quantitative regularization (QR), designed to enhance the performance of ViTs in RSI classification. QR represents an effective algorithm that adeptly manages domain discrepancies in RSIs and can be integrated with any ViTs in transfer learning. We evaluated the effectiveness of QR using three ViT architectures: vanilla ViT, Swin‐ViT and Next‐ViT, on four datasets: AID30, NWPU45, AFGR50 and UCM21. The results reveal that our Next‐ViT model surpasses 39 other advanced methods published in the past 3 years, maintaining robust performance even with a limited number of training samples. We also discovered that our ViT and Swin‐ViT achieve significantly higher accuracy and robustness compared to other methods using the same backbone. Our findings confirm that ViTs can be as effective as CNNs for RSI classification, regardless of the dataset size. Our approach exclusively employs open‐source ViTs and easily accessible training strategies. Consequently, we believe that our method can significantly lower the barriers for geoscience researchers intending to use ViT for RSI applications.</jats:p>

<jats:title>Abstract</jats:title><jats:p>Graffiti, by their very nature, are ephemeral, sometimes even vanishing before creators finish them. This transience is part of graffiti's allure yet signifies the continuous loss of this often disputed form of cultural heritage. To counteract this, graffiti documentation efforts have steadily increased over the past decade. One of the primary challenges in any documentation endeavour is identifying and recording new creations. Image‐based change detection can greatly help in this process, effectuating more comprehensive documentation, less biased digital safeguarding and improved understanding of graffiti. This paper introduces a novel and largely automated image‐based graffiti change detection method. The methodology uses an incremental structure‐from‐motion approach and synthetic cameras to generate co‐registered graffiti images from different areas. These synthetic images are fed into a hybrid change detection pipeline combining a new pixel‐based change detection method with a feature‐based one. The approach was tested on a large and publicly available reference dataset captured along the Donaukanal (Eng. Danube Canal), one of Vienna's graffiti hotspots. With a precision of 87% and a recall of 77%, the results reveal that the proposed change detection workflow can indicate newly added graffiti in a monitored graffiti‐scape, thus supporting a more comprehensive graffiti documentation.</jats:p>

<jats:title>Abstract</jats:title><jats:p>The problem of sequential estimation of the exterior orientation of imaging sensors and the three‐dimensional environment reconstruction in real time is commonly known as visual simultaneous localisation and mapping (vSLAM). Omnidirectional optical sensors have been increasingly used in vSLAM solutions, mainly for providing a wider view of the scene, allowing the extraction of more features. However, dealing with unmodelled points in the hyperhemispherical field poses challenges, mainly due to the complex lens geometry entailed in the image formation process. To address these challenges, the use of rigorous photogrammetric models that appropriately handle the geometry of fisheye lens cameras can overcome these challenges. Thus, this study presents a real‐time vSLAM approach for omnidirectional systems adapting ORB‐SLAM with a rigorous projection model (equisolid‐angle). The implementation was conducted on the Nvidia Jetson TX2 board, and the approach was evaluated using hyperhemispherical images captured by a dual‐fisheye camera (Ricoh Theta S) embedded into a mobile backpack platform. The trajectory covered a distance of 140 m, with the approach demonstrating accuracy better than 0.12 m at the beginning and achieving metre‐level accuracy at the end of the trajectory. Additionally, we compared the performance of our proposed approach with a generic model for fisheye lens cameras.</jats:p>