A novel iteration scheme for solving consistent linear equations is introduced by exploring the circumcenter of a n-dimensional triangle formed with a starting point and its reflections across two hyperplanes. The proposed method, akin to the block Kaczmarz method, provides optimal linear combination within two subspaces. The convergence rate in expectation is also addressed in detail if the row is selected with probability proportional to p-norm of residuals. Numerical experiments demonstrate its superiority over randomized Kaczmarz and reflection methods, particularly evident as highly coherent rows. Performance are enhanced by proper parameter-tuning of p.

A modified relaxation two-sweep modulus-based matrix splitting iteration method is proposed by using relaxation techniques and the two-sweep scheme for the solution of implicit complementarity problems. The sufficient conditions for convergence of the proposed method are established when the coefficient matrix is a positive definite matrix and an H+-matrix, respectively. Numerical experiments are presented to show the efficiency of the proposed method, which is superior to the modified modulus-based matrix splitting iteration method in terms of the number of iteration steps and the elapsed CPU time.

A class of restarted randomized surrounding methods are presented to accelerate the surrounding algorithms by restarted techniques for solving the linear equations. Theoretical analysis prove that the proposed method converges under the randomized row selection rule and the expectation convergence rate is also addressed. Numerical experiments further demonstrate that the proposed algorithms are efficient and outperform the existing method for over-determined and under-determined linear equation, as well as in the application of image processing.

The existing forest inventories have difficulties in providing data with sufficient spatial and temporal resolution to quantify dynamic properties of forests, which are mainly caused by short-term events such as drought, storm, snow damages, or pest infestations. This study aims to explore in a first step the potential of space-borne LiDAR for forest parameters extraction over Alpine forests in Austria. The space-borne LiDAR data investigated in this study is ICESat2 (Ice, Cloud and Land Elevation Satellite-2) and GEDI (Global Ecosystem Dynamics Investigation). GEDI is a full-waveform, multibeam laser altimeter on the International Space Station, with a footprint diameter of about 25m (Dubayah, 2020). As for ICESat-2, it carries a micropulse, multi-beam photon-counting laser altimeter, with a footprint diameter of about 17m (Neuenschwander and Magruder, 2019). Two footprint-level products are used in this study: GEDI L2A and ICESat-2 ATL03. The GEDI L2A product provides footprint-level elevation and height metrics that extract terrain height, canopy height, and relative height metrics from the received waveform. ICESat-2 ATL03 records horizontal coordinates and ellipsoidal heights of all photon data, and the classification labels are extracted from its higher product, ATL08. The DTM with a resolution of 1m and the DSM derived from ALS (Airborne Laser Scanning) point clouds acts as “ground truth” to assess the accuracy of the terrain and canopy height of the two space-borne LiDAR products, respectively. For ICESat-2 ATL03, only photons with a signal confidence flag ranging from medium confidence (sigal_confidence=3) or high confidence (sigal_confidence=4) are included for evaluation. For GEDI L2A, only waveforms with a valid quality flag (quality_flag=1) are included for evaluation. To evaluate the performance of ICESat2 and GEDI for different forest types and topographic conditions, two study sites in Austria are selected: western part of Tyrol and the Vienna Woods. A preliminary results of terrain and canopy height accuracies shows that the terrain height of the two space-borne LiDAR products fits well with the DTM. Compared to GEDI L2A, ICESat2 ATL03 has a better correlation with DTM values. The canopy height accuracy is not as good as the terrain height accuracy. It has been shown that ICESat-2 tend to underestimate the canopy top height as derived from airborne LiDAR. Overall, GEDI has a better canopy height accuracy than ICESat-2. Furthermore, we have investigated the influence of different beam power, data acquisition time and season. In general, the accuracy of both ICESat-2 and GEDI data acquired in nighttime is higher than that of the daytime data. The statistic also shows that for the ICESat-2 data, the terrain height accuracy of weak beam footprints only slightly worse than that of strong beam footprints. For GEDI, footprints of strong beams always perform better than that of weak beams in terms of terrain and canopy height. Regarding the impact of season, for both ICESat-2 and GEDI, the canopy height is more accurate in summer than the collection in winter. For GEDI, the terrain height can be better measured in winter than in summer.

点云数据的特征处理是机器人、自动驾驶等领域中三维物体识别技术的关键组成部分,针对点云局部特征信息重复提取、点云物体整体几何结构缺乏识别等问题,提出一种基于环查询和通道注意力的点云分类与分割网络.首先将单层环查询和特征通道注意力机制进行结合,减少局部信息冗余并加强局部特征;然后计算法线变化识别出物体边缘、拐角区域的高响应点,并将其法线特征加入全局特征表示中,加强物体整体几何结构的识别.在ModelNet40和ShapeNet Part数据集上与多种点云网络进行比较,实验结果表明,该网络不仅有较高的点云分类与分割精度,同时在训练时间和内存占用等方面也优于其他方法,此外对于不同输入点云数量具有较强鲁棒性.因此该网络是一种有效、可行的点云分类与分割网络.

The success achieved by deep learning techniques in image labeling has triggered a growing interest in applying deep learning for three-dimensional point cloud classification.To provide better insights into different deep learning architectures and their applications to ALS point cloud classification, this article presents a comprehensive comparison among three state-of-the-art deep learning networks: PointNet++, SparseCNN, and KPConv, on two different ALS datasets.The performances of these three deep learning networks are compared w.r.t.classification accuracy, computation time, generalization ability as well as the sensitivity to the choices of hyper-parameters.Overall, we observed that PointNet++, SparseCNN, and KPConv all outperform Random Forest on the classification results.Moreover, SparseCNN leads to a slightly better classification result compared to PointNet++ and KPConv, while requiring less computation time and memory.At the same time, it shows a better ability to generalize and is less impacted by the different choices of hyper-parameters.

Airborne laser scanning of the city of Vienna, which was organized by the survey department of the City Administration of Vienna (MA41). The data acquisition was performed in eight flight missions in November 2015. After strip adjustment relative accuracy of the point cloud was in the order of 2cm. The absolute accuracy measured as RMSE is better than 5cm in planimetry and better than 4cm in elevation. The dataset was cleaned and does not involve obvious gross errors, e.g. points high up in the air or points far below ground level. The point density is measured as by last echoes per unit area and is more than 15 points/m2 for 97% of the area. The average point density is 33 points/m2. The LiDAR dataset of the city is organized in a number of 1270m×1020m tiles (including a 10m overlap of all neighboring tiles). Reference labels were generated semi-automatically: a rough filtering of main objects was firstly conducted by the software “Terrasolid”, and the final classification was refined by manual labelling. For the purpose of Vienna city administration, five classes are considered, namely ground, buildings, vegetation, others, water and bridges. All common street objects are categorized as others, such as (e.g.) streetlights, benches, shrubs, cars, construction sites and garbage bins. The different classes are defined by the following numeric integer codes: 2: Ground, 5: Vegetation, 6: Buildings, 8: Others, 9: Water and 17: Bridges. The quality of the labelling was manually checked. In 20 sites of 100m×100m the classification was manually verified, and the average accuracy of reference labels is 95%. Total 9 tiles are published, in which 4 tiles were used for the training and 5 tiles for the evaluation in the paper “A Comparison of Deep Learning Methods for Airborne LiDAR Point Clouds Classification”. Their locations in the city of Vienna can be found in the file of metadata, which also provides WGS84/GRS80 latitude and longitude coordinates of the extent corners of each tile (EPSG: 31256). This can be used to access images of the tiles from Google Maps or other public map service.

Jednou z mnoha aplikací ULS (UAV-borne laser scanning) je inspekce elektrického vedení.Nicméně s výstupem LiDAR sběru dat (mračen bodů) přichází i potřeba data automaticky klasifikovat, neboli sémanticky segmentovat, za účelem navazující analýzy.Metod pro automatickou klasifikaci mračen bodů bylo představeno nemalé množství, mnoho z nich s využitím strojového učení.Motivací tohoto výzkumu je nutná podmínka strojového učení v podobě referenčních (trénovačích) dat pro učení modelukonkrétně dopad chybovosti v klasifikaci referenčních dat na přesnost modelované klasifikace výstupů modelu.K zjištění dopadu chybovosti trénovacích dat na výstup strojového učení pro klasifikaci mračen bodů elektrického vedení jsme použili metodu klasifikačních a regresních stromů (CART) implementovanou v programu Opals.V rámci výzkumu byly testovány datové sady s různou mírou a různým typem chybovosti referenčních dat a jejich vliv na výslednou přesnost byl porovnán s daty, které nebyly použity pro samotné učení modelu.

Based on the mass point cloud data, this paper proposes a hybrid octree mixing point cloud index structure which combines the KD-tree spatial segmentation idea to realize the efficient management of mass point cloud. In this paper, the space of the point cloud is firstly divided by the KD-tree idea. On this basis, the octree is used for further segmentation to establish an octree-like index structure. Then the point cloud dataset is spatially encoded using the improved encoding to achieve better spatial management and neighborhood search. Finally, using five groups of incremented point cloud set as test data, the experimental results and comparison analysis show that the octree-like space can make the overall structure of the data organization more reasonable, effectively improve the access efficiency and reduce the occupancy of memory space. The index structure not only improves the speed of the traditional KD-tree construction index but also improves the problem that the traditional octree is too large for space occupation and the neighborhood search takes too long. It achieves reasonable management of massive point cloud space.

Training dataset generation is a difficult and expensive task for LiDAR point classification, especially in the case of large area classification. We present a method to automatically extent a small set of training data by label propagation processing. The class labels could be correctly extended to their optimal neighbourhood, and the most informative points are selected and added into the training set. With the final extended training dataset, the overall (OA) classification could be increased by about 2%. We also show that this approach is stable regardless of the number of initial training points, and achieve better improvements especially stating with an extremely small initial training set.

Surface point cloud matching is a useful technique for patient positioning during radiation therapy, system registering of surgical navigation system, etc. A common method for 3D point cloud registration is to estimate the registration function based on the 3D keypoint feature correspondences. However, the feature distance of correct correspondence is usually not the closest, but it is hidden in the k-nearest correspondences generally. This makes the rate of false correspondences too high for registration. In this paper, we convert the 3D point clouds rigid registration problem to a new graph matching model which combines the k-nearest feature information and the geometric constraints. The key idea is that the distance between 2 model feature points does not change after rigid transformation. This is a quadratic programming problem which could be transformed into an integer linear programming problem and thus could be efficiently solved. The final output is the 3D keypoint correspondences set which are accurate enough for registration. Our experimental results on point cloud datasets, generated from TCIA CT images datasets, show the advantages of the proposed algorithm.

辐射源信号调制样式的多样化为准确识别辐射源带来了困难,双谱对角切片特征能明显反映辐射源信号特性,深度神经网络学习则能处理信号样本大数据,将双谱特征和深度神经网络学习用于信号调制识别中,能提取辐射源信号本质特征,同时也能提高辐射源调制信号的正确识别率.仿真实验结果表明,相比于其它识别算法,双谱特征能更好地反映信号特性,深度学习模型有更高的信号识别率.

We propose a contextual label-smoothing method to improve the LiDAR classification accuracy in a post-processing step. Under the framework of global graph-structured regularization, we enhance the effectiveness of label smoothing from two aspects. First, each point can collect sufficient label-relevant neighborhood information to verify its label based on an optimal graph. Second, the input label probability set is improved by probabilistic label relaxation to be more consistent with the spatial context. With this optimal graph and reliable label probability set, the final labels are computed by graph-structured regularization. We demonstrate the contextual label-smoothing approach on two separate urban airborne LiDAR datasets with complex urban scenes. Significant improvements in the classification accuracies are achieved without losing small objects (such as façades and cars). The overall accuracy is increased by 7.01% on the Vienna dataset and 6.88% on the Vaihingen dataset. Moreover, most large, wrongly labeled regions are corrected by long-range interactions that are derived from the optimal graph, and misclassified regions that lack neighborhood communications in terms of correct labels are also corrected with the probabilistic label relaxation.

With the rapid development of urban economy, convenient, safe, and efficient urban rail transit has become the preferred method for people to travel. In order to ensure the safety and sustainable development of urban rail transit, the PS-InSAR technology with millimeter deformation measurement accuracy has been widely applied to monitor the deformation of urban rail transit. In this paper, 32 scenes of COSMO-SkyMed descending images and 23 scenes of Envisat ASAR images covering the Shanghai Metro Line 6 acquired from 2008 to 2010 are used to estimate the average deformation rate along line-of-sight (LOS) direction by PS-InSAR method. The experimental results show that there are two main subsidence areas along the Shanghai Metro Line 6, which are located between Wuzhou Avenue Station to Wulian Road Station and West Gaoke Road Station to Gaoqing Road Station. Between Wuzhou Avenue Station and Wulian Road Station, the maximum displacement rate in the vertical direction of COSMO-SkyMed images is −9.92 mm/year, and the maximum displacement rate in the vertical direction of Envisat ASAR images is −8.53 mm/year. From the West Gaoke Road Station to the Gaoqing Road Station, the maximum displacement rate in the vertical direction of COSMO-SkyMed images is −15.53 mm/year, and the maximum displacement rate in the vertical direction of Envisat ASAR images is −17.9 mm/year. The results show that the ground deformation rates obtained by two SAR platforms with different wavelengths, different sensors and different incident angles have good consistence with each other, and also that of spirit leveling.

PSInSAR technology has been widely applied in ground deformation monitoring. Accurate identification of Persistent Scatterers (PS) is key to the success of PSInSAR data processing. In this paper, the theoretic models and specific algorithms of PS point extraction methods are summarized and the characteristics and applicable conditions of each method, such as Coherence Coefficient Threshold method, Amplitude Threshold method, Dispersion of Amplitude method, Dispersion of Intensity method, are analyzed. Based on the merits and demerits of different methods, an improved method for PS point extraction in urban area is proposed, that uses simultaneously backscattering characteristic, amplitude and phase stability to find PS point in all pixels. Shanghai city is chosen as an example area for checking the improvements of the new method. The results show that the PS points extracted by the new method have high quality, high stability and meet the strong scattering characteristics. Based on these high quality PS points, the deformation rate along the line-of-sight (LOS) in the central urban area of Shanghai is obtained by using 35 COSMO-SkyMed X-band SAR images acquired from 2008 to 2010 and it varies from −14.6 mm/year to 4.9 mm/year. There is a large sedimentation funnel in the cross boundary of Hongkou and Yangpu district with a maximum sedimentation rate of more than 14 mm per year. The obtained ground subsidence rates are also compared with the result of spirit leveling and show good consistent. Our new method for PS point extraction is more reasonable, and can improve the accuracy of the obtained deformation results.

The common statistical methods for supervised classification usually require a large amount of training data to achieve reasonable results, which is time consuming and inefficient. In many methods, only the features of each point are used, regardless of their spatial distribution within a certain neighborhood. This paper proposes a tensor-based sparse representation classification (TSRC) method for airborne LiDAR (Light Detection and Ranging) points. To keep features arranged in their spatial arrangement, each LiDAR point is represented as a 4th-order tensor. Then, TSRC is performed for point classification based on the 4th-order tensors. Firstly, a structured and discriminative dictionary set is learned by using only a few training samples. Subsequently, for classifying a new point, the sparse tensor is calculated based on the tensor OMP (Orthogonal Matching Pursuit) algorithm. The test tensor data is approximated by sub-dictionary set and its corresponding subset of sparse tensor for each class. The point label is determined by the minimal reconstruction residuals. Experiments are carried out on eight real LiDAR point clouds whose result shows that objects can be distinguished by TSRC successfully. The overall accuracy of all the datasets is beyond 80% by TSRC. TSRC also shows a good improvement on LiDAR points classification when compared with other common classifiers.

The common statistical methods for supervised classification usually require a large amount of training data to achieve reasonable results, which is time consuming and inefficient. This paper proposes a tensor sparse representation classification (SRC) method for airborne LiDAR points. The LiDAR points are represented as tensors to keep attributes in its spatial space. Then only a few of training data is used for dictionary learning, and the sparse tensor is calculated based on tensor OMP algorithm. The point label is determined by the minimal reconstruction residuals. Experiments are carried out on real LiDAR points whose result shows that objects can be distinguished by this algorithm successfully.

Feature selection and description is a key factor in classification of Earth observation data. In this paper a classification method based on tensor decomposition is proposed. First, multiple features are extracted from raw LiDAR point cloud, and raster LiDAR images are derived by accumulating features or the “raw” data attributes. Then, the feature rasters of LiDAR data are stored as a tensor, and tensor decomposition is used to select component features. This tensor representation could keep the initial spatial structure and insure the consideration of the neighborhood. Based on a small number of component features a k nearest neighborhood classification is applied.

Early detection and early warning are of great importance in giant landslide monitoring because of the unexpectedness and concealed nature of large-scale landslides. In China, the western mountainous areas are prone to landslides and feature many giant complex landslides, especially following the Wenchuan Earthquake in 2008. This work concentrates on a new technique, known as the “hybrid-SAR technique”, that combines both phase-based and amplitude-based methods to detect and monitor large-scale landslides in Li County, Sichuan Province, southwestern China. This work aims to develop a robust methodological approach to promptly identify diverse landslides with different deformation magnitudes, sliding modes and slope geometries, even when the available satellite data are limited. The phase-based and amplitude-based techniques are used to obtain the landslide displacements from six TerraSAR-X Stripmap descending scenes acquired from November 2014 to March 2015. Furthermore, the application circumstances and influence factors of hybrid-SAR are evaluated according to four aspects: (1) quality of terrain visibility to the radar sensor; (2) landslide deformation magnitude and different sliding mode; (3) impact of dense vegetation cover; and (4) sliding direction sensitivity. The results achieved from hybrid-SAR are consistent with in situ measurements. This new hybrid-SAR technique for complex giant landslide research successfully identified representative movement areas, e.g., an extremely slow earthflow and a creeping region with a displacement rate of 1 cm per month and a typical rotational slide with a displacement rate of 2–3 cm per month downwards and towards the riverbank. Hybrid-SAR allows for a comprehensive and preliminary identification of areas with significant movement and provides reliable data support for the forecasting and monitoring of landslides.

Similarity search is a fundamental process in many hyperspectral remote sensing applications. In this article, we investigated the locality-sensitive hashing (LSH) algorithm for approximate similarity search on hyperspectral remote sensing data and proposed a new method that uses a regulated random hyperplane projection hash function family to index data and an adjacent graph-probing method for similarity queries. This method improves the performance of LSH over non-uniformly distributed hyperspectral data sets. Comparative experiments with three benchmark hyperspectral data sets showed that the proposed method is at least two times faster than the basic LSH and two other improved methods whilst achieving the same search accuracy. Moreover, results of the proposed method are proven equivalent to the accurate similarity search for real applications.