Using multi-scale and hierarchical deep convolutional features for 3D semantic classification of TLS point clouds

ABSTRACT

Point cloud classification, which provides meaningful semantic labels to the points in a point cloud, is essential for generating three-dimensional (3D) models. Its automation, however, remains challenging due to varying point densities and irregular point distributions. Adapting existing deep-learning approaches for two-dimensional (2D) image classification to point cloud classification is inefficient and results in the loss of information valuable for point cloud classification. In this article, a new approach that classifies point cloud directly in 3D is proposed. The approach uses multi-scale features generated by deep learning. It comprises three steps: (1) extract single-scale deep features using 3D convolutional neural network (CNN); (2) subsample the input point cloud at multiple scales, with the point cloud at each scale being an input to the 3D CNN, and combine deep features at multiple scales to form multi-scale and hierarchical features; and (3) retrieve the probabilities that each point belongs to the intended semantic category using a softmax regression classifier. The proposed approach was tested against two publicly available point cloud datasets to demonstrate its performance and compared to the results produced by other existing approaches. The experiment results achieved 96.89% overall accuracy on the Oakland dataset and 91.89% overall accuracy on the Europe dataset, which are the highest among the considered methods.     

3D building roof reconstruction from airborne LiDAR point clouds: a framework based on a spatial database

Three-dimensional (3D) building models are essential for 3D Geographic Information Systems and play an important role in various urban management applications. Although several light detection and ranging (LiDAR) data-based reconstruction approaches have made significant advances toward the fully automatic generation of 3D building models, the process is still tedious and time-consuming, especially for massive point clouds. This paper introduces a new framework that utilizes a spatial database to achieve high performance via parallel computation for fully automatic 3D building roof reconstruction from airborne LiDAR data. The framework integrates data-driven and model-driven methods to produce building roof models of the primary structure with detailed features. The framework is composed of five major components: (1) a density-based clustering algorithm to segment individual buildings, (2) an improved boundary-tracing algorithm, (3) a hybrid method for segmenting planar patches that selects seed points in parameter space and grows the regions in spatial space, (4) a boundary regularization approach that considers outliers and (5) a method for reconstructing the topological and geometrical information of building roofs using the intersections of planar patches. The entire process is based on a spatial database, which has the following advantages: (a) managing and querying data efficiently, especially for millions of LiDAR points, (b) utilizing the spatial analysis functions provided by the system, reducing tedious and time-consuming computation, and (c) using parallel computing while reconstructing 3D building roof models, improving performance.     

基于卷积神经网络的极光图像分类

极光是由带电粒子经磁层—电离层碰撞大气而产生的。面对形态各异、演变过程复杂的极光图像,对其合理分类为进一步探究日地电磁活动和能量耦合等空间物理问题奠定了基础。针对该问题,引入深度学习的方法,通过卷积神经网络模型自主表征极光特征并实现极光图像分类。该方法对2003年北极黄河站越冬观测的38 044幅和8 001幅典型极光图像分类正确率达93.17%和91.5%;自动识别2004—2009年观测数据的极光形态,4类极光时间分布规律与三波段激发谱能量分布基本一致。实验结果表明,基于卷积神经网络的极光表征方法,能有效实现极光图像的自动分类。     

U-Net深度卷积神经网络在沙脊线提取中的应用

沙丘形态演变过程记录着近地表风况与环境演化的历史,然而对其特征研究一直受限于大范围沙脊线提取效率低和成本高等问题.本文基于深度卷积神经网络搭建U-Net模型,实现批量、高精度沙脊线的提取.将数据增强技术、随机失活神经元、批标准化以及迁移学习技术应用于模型训练和参数更新,使得模型的精度更高.结果 表明:U-Net模型以及...     

Street-Frontage-Net: urban image classification using deep convolutional neural networks

ABSTRACT

Quantifying aspects of urban design on a massive scale is crucial to help develop a deeper understanding of urban designs elements that contribute to the success of a public space. In this study, we further develop the Street-Frontage-Net (SFN), a convolutional neural network (CNN) that can successfully evaluate the quality of street frontage as either being active (frontage containing windows and doors) or blank (frontage containing walls, fences and garages). Small-scale studies have indicated that the more active the frontage, the livelier and safer a street feels. However, collecting the city-level data necessary to evaluate street frontage quality is costly. The SFN model uses a deep CNN to classify the frontage of a street. This study expands on the previous research via five experiments. We find robust results in classifying frontage quality for an out-of-sample test set that achieves an accuracy of up to 92.0%. We also find active frontages in a neighbourhood has a significant link with increased house prices. Lastly, we find that active frontage is associated with more scenicness compared to blank frontage. While further research is needed, the results indicate the great potential for using deep learning methods in geographic information extraction and urban design.     

Simulating urban land use change by integrating a convolutional neural network with vector-based cellular automata

ABSTRACT

Vector-based cellular automata (VCA) models have been applied in land use change simulations at fine scales. However, the neighborhood effects of the driving factors are rarely considered in the exploration of the transition suitability of cells, leading to lower simulation accuracy. This study proposes a convolutional neural network (CNN)-VCA model that adopts the CNN to extract the high-level features of the driving factors within a neighborhood of an irregularly shaped cell and discover the relationships between multiple land use changes and driving factors at the neighborhood level. The proposed model was applied to simulate urban land use changes in Shenzhen, China. Compared with several VCA models using other machine learning methods, the proposed CNN-VCA model obtained the highest simulation accuracy (figure-of-merit = 0.361). The results indicated that the CNN-VCA model can effectively uncover the neighborhood effects of multiple driving factors on the developmental potential of land parcels and obtain more details on the morphological characteristics of land parcels. Moreover, the land use patterns of 2020 and 2025 under an ecological control strategy were simulated to provide decision support for urban planning.     

RecNet: a deep neural network for personalized POI recommendation in location-based social networks

How to exploit various features of users and points of interest (POIs) for accurate POI recommendation is important in location-based social networks (LBSNs). In this paper, a novel POI recommendation framework, named RecNet, is proposed, which is developed based on a deep neural network (DNN) to incorporate various features in LBSNs and learn their joint influence on user behavior. More specifically, co-visiting, geographical and categorical influences in LBSNs are exploited to alleviate the data sparsity issue in POI recommendation and are converted to feature vector representations of POIs and users via feature embedding. Moreover, the embedded POIs and users are fed into a DNN pairwise to adaptively learn high-order interactions between features. Our method is evaluated on two publicly available LBSNs datasets and experimental results show that RecNet outperforms state-of-the-art algorithms for POI recommendation.