Search

With the rapid development of geospatial science, artificial intelligence, and high-performance computing, geospatial artificial intelligence has become the major technique for processing and analyzing geospatial big data, and will be widely used in the scientific research and engineering application of earth science, spatial cognition, and smart city. Geospatial intelligence, as an inter-discipline between geospatial science and artificial intelligence, is driven by dual disciplines. At present, it has made important progress in hardware research, system development, data and model sharing, services and applications, and is also facing new challenges. This paper first describes the conceptual evolution of geospatial artificial intelligence and lists the frameworks of some technical systems. The paper then reviews the state-of-the-art in the scientific research and typical application domains. The problems and challenges facing geospatial intelligence are analyzed. Finally, some future prospects and trends are presented.

Cropland is a scarce land resource in hilly and mountainous areas, which has the characteristics of complex topographic conditions and diverse planting structures, leading to the difficulty of rapid and accurate acquisition of cropland information in mountainous areas. Therefore, it is difficult to extract the cropland information in mountainous areas quickly and automatically based on the traditional remote sensing data and remote sensing monitoring methods. Aiming at this problem, this paper takes Xifeng County of Guizhou Province in the southwest mountainous area as the experimental area. According to the heterogeneity of geospatial space, this paper proposes the idea of cropland morphological information extraction by geographical division control and stratification extraction, and constructs a method for extracting cropland morphological information based on geographical division control and stratification extraction under the constraints of geomorphic unit. Firstly, according to the geomorphology-vegetation characteristics, the experimental area is divided into three geographical zones: flatland, hillside area and forest. Then, on the basis of each type of partition, the cropland is divided into different types according to the visual characteristics presented by the cropland, and different deep learning models are designed for hierarchical extraction of different types of cropland. The experimental results show that this method has a good suppression effect on the background noise of complex terrain in mountainous areas, and the extracted cropland plot information is more consistent with the actual distribution pattern of the actual cropland compared with the traditional method, which effectively reduces the rate of missing extraction and wrong extraction.

Traffic plays a vital role in people's life and social economy.Vehicle speed detection is an important part of intelligent transportation system.This paper focus on the vehicle speed detection based on multi-source data from autonomous unmanned aerial vehicle (UAV).Firstly,we build a multi-source data acquisition system on UAV for visible image and thermal infrared image.Secondly,we utilize "You only look once" (YOLO),which is a deep learning framework for vehicle detection.Finally,we track the vehicle based on Kalman filter and calculated the vehicle speed according to the result of vehicle tracking.This paper adopts the UAV platform to increase the flexibility.While the use of multi-source data improves the accuracy of the vehicle detection and tracks the vehicle in different illumination.The result of experiments shows that the strategy is effective and robust,which provides an efficient and flexible monitoring mode for traffic management department.

Accurately identifying roadside objects like trees, cars, and traffic poles from mobile LiDAR point clouds is of great significance for some applications such as intelligent traffic systems, navigation and location services, autonomous driving, and high precision map. In the paper, we proposed a point-group-view network (PGVNet) to classify the roadside objects into trees, cars, traffic poles, and others, which utilize and fuse the advanced global features of multi-view images and the spatial geometry information of point cloud. To reduce redundant information between similar views and highlight salient view features, the PGVNet model employs a hierarchical view-group-shape architecture to split all views into different groups according to their discriminative level, which uses the pre-trained VGG network as the bone network. In view-group-shape architecture, global-level significant features are further generated from group descriptors with their weights. Moreover, an attention-guided fusion network is used to fuse the global features from multi-view images and local geometric features from point clouds. In particular, the global advanced features from multi-view images are quantified and leveraged as the attention mask to further refine the intrinsic correlation and discriminability of the local geometric features from point clouds, which contributions to recognize the roadside objects. We have evaluated the proposed method on five different mobile LiDAR point cloud data. Five test datasets of different urban scenes by different mobile laser scanning systems are used to evaluate the validities of the proposed method. Four accuracy evaluation metrics precision, recall, quality and Fscore of trees, cars and traffic poles on the selected testing datasets achieve (99.19%, 94.27%, 93.58%, 96.63%), (94.20%, 97.56%, 92.02%, 95.68%), (91.48%, 98.61%, 90.39%, 94.87%), respectively. Experimental results and comparisons with state-of-the-art methods demonstrate that the PGVNet model is available to effectively identify roadside objects from the mobile LiDAR point cloud, which can provide data support for elements construction and vectorization in high precision map applications.

Simultaneous localization and mapping (SLAM) has broad application prospects in many fields due to its high energy efficiency and low power consumption. However, there are still some problems in the traditional SLAM system: the key frame in the traditional visual odometer does not contain semantic information, the image information obtained by the mobile robot is relatively single, and the key frame in the actual scene always contains a large number of mismatched points and dynamic points. In response to the above problems, this paper proposes a new idea of semantic SLAM mapping technology. First, in order to find the correct and corresponding feature points, abandon the interference of dynamic points and mismatched points simultaneously, a method for judging the feature state of adjacent frames based on the Lucas-Kanade optical flow method is proposed, and this function is regarded as a new feature. The thread is added to the visual odometry part of ORB-SLAM3 to complete the optimization and improvement of part of the traditional SLAM framework. Secondly, in view of the problem that the image frame obtained by the front-end visual odometer of the traditional SLAM system does not contain any semantic information, the target detection algorithm based on YOLOV4 and the Mask R-CNN semantic segmentation algorithm fused with fully connected conditional random field CRF are used to compare ORB-SLAM3. The key frame image processing of the robot effectively improves the perception of the indoor environment of smart devices such as robots.

The rapid development of earth observation projects in China has obtained a large volume of multi-source (multi-type sensors, multi-temporal, multi-scale) remote sensing data. But the capability of intelligent remote sensing image processing lags behind data acquisition. In recent years, people have significantly improved the effectiveness of image feature extraction with deep learning networks. But limited number and variety of sample data is not enough for processing the multi-source remote sensing images. This paper analyzed existing sample datasets and proposed a method for constructing a sample database for intelligent remote sensing image interpretation, including the data model, classification system, data organization, as well as the Internet-based platform for collaborative sample collection and sharing.

Accurate interpretation of remote sensing image (RSI) plays a vital role in the implementation of remote sensing applications. In recent years, deep supervised learning has achieved great success in the field of RSI interpretation by its soaring performance on representation learning. However, this method heavily relies on large-scale and high-quality labeled data, while building a big remote sensing data set is extremely expensive because of the unique spatial-temporal heterogeneity of remote sensing data. This contradiction seriously restricts the performance of deep supervised learning in large areas and complicated remote sensing scenes. How to solve the last mile problem in the field of RSI accurate interpretation becomes urgent. This paper first systematically reviews the main research progress of supervised learning methods in the field of RSI interpretation, and then analyzes its limitations. Afterward, we introduce the concept of self-supervised learning and detail how it works for unsupervised feature learning. Finally, we briefly discuss open problems and future directions of self-supervised learning if it is applied in the field of RSI interpretation, with the aim of providing a new perspective for RSI interpretation with the adoption of huge unlabeled data.

In recent years, point cloud has become an important type of 3D spatial data. How to improve the understanding abilities of point cloud using artificial intelligence for correct semantic labeling and accurate detection of objects is an urgent and difficult problem. This paper hence proposes an end-to-end 3D point cloud deep learning network, which effectively guarantees the efficiencies of point cloud sampling, the accuracy of feature extraction and the optimization of the overall network performance by the up-down sampling strategy of irregular distribution point cloud, multi-layer aggregation and propagation of features and the loss function for uneven samples. The studies on the large-scale 3D point cloud benchmark data show that it achieves excellent performance in semantic labeling for large-scale outdoor scenes of point clouds, better than those of the state-of-art deep learning networks of point cloud, providing a strong support for the high-performance extraction of 3D geospatial information.

Surface subsidence not only affects the sustainable development of social economy, but also threatens the safety of human life. High precision prediction of surface subsidence is of great significance for the prevention of geological disasters. However, the existing prediction methods are difficult to obtain reliable prediction results because of the model parameters or the lack of relevant data. For this problem, a method of surface subsidence prediction based on deep learning is proposed. Firstly, the multiple master-image coherent target small-baseline InSAR (MCTSB-InSAR) is used to obtain the inversion results of large area and high precision ground deformation time series. Secondly, the cyclic neural network is used as the network framework, and the long short-term memory (LSTM) model is used to learn the characteristics of ground settlement. Finally, the grid search method is used to adjust the model parameters, then get the optimal combination scheme of model parameters. The actual observation results show that the average absolute error (0.3 mm) of the prediction model proposed in this paper is reduced by 27.3% at least, and the average prediction accuracy of differential settlement is improved by 8.9% at least. The results of spatial pattern analysis show that the LSTM model is effective for the short-term prediction of large-scale time series deformation.

Influenced by the performance of imaging spectrometer and the distribution of complex ground objects, hyperspectral images have a large number of mixed pixels. Traditional learning-based unmixing methods are shallow models, or lack of comprehensive use of spatial and spectral information. This paper proposes a collaborative deep model with multi-dimensional convolutional network. Using multi-dimensional convolutional network can make full use of multi-dimensional semantic information, which is better to estimate hyperspectral mixed pixel abundance with small samples. The method augments training data, constructs three kinds of convolutional neural networks: spectral dimension, spatial dimension and cube dimension; the method designs fusion layer to concatenate features with three kinds of convolutional neural networks, and to “end-to-end” estimate of mixed pixel abundance; the model uses batch normalization, pooling and dropout to avoid over fitting phenomenon. The experimental results indicate that the introduction of our proposed method can extract spatial-spectral feature information more effectively. Compared with other convolutional network unmixing models, the accuracy of the estimated mixed pixel abundance is significantly improved.

Urbanization has triggered great development and changes in underground space. Exploring the types and positions of underground targets is of vital importance to urban underground security and utilization. GPRs (ground penetrating radar) are widely used in exploring underground space because of its advantages of rapid data collection, convenience, high imaging resolution and non-destructive inspection. However, the heavy manual interpretation costs of object detection from GPR limit the GPR applications in large-scale urban underground objects detection. This paper analyzes and determines seven typical types of urban road underground target that can be detected in GPR images (e.g. rainwater wells, cables, etc.). According to the characteristics of its reflected signals, the underground target in the GPR data of the 400 MHz band acquired by GSSI SIR30 in a typical urban road environment are labeled to construct the training dataset with seven categories and 3033 training samples. With the transfer learning method, the pre-trained Darknet 53 network parameters are fine-tuned, and the end-to-end YOLOV3 detection method is used to automatically extract and locate the underground targets. Finally, the experimental verification was carried out using the GPR data of the 400 MHz band collected by GSSI SIR30 in Caitian Road, Futian District, Shenzhen. Experiments show that the proposed deep learning detection method detects the buried objects from GRP data effectively, in terms of 85% of recall and precision,and the detection speed of 16FPS.

Road information extraction based on vehicle trajectory data is one of the hotspots and difficulties in the field of geographic information. The rapid development of depth learning provides a new idea and method for solving this problem. Aiming at the problem of roadway-level road extraction based on vehicle trajectory data, this paper introduces the generative adversarial nets in the field of deep learning, uses residual network to construct deep network and multi-scale receptive field to perceive different details of trajectory data, and constructs roadway-level road extraction model under the constraint of trajectory direction based on conditional generative adversarial nets. Firstly, the orientation-color mapping rasterization conversion method is proposed to transform the trajectory orientation information into HSV color space. Then, the parameters of the model are learned with the sample data. Finally, the trained model is applied to three experimental areas of Zhengzhou, Chengdu and Nanjing to extract the road data at the roadway level. The experimental results showed that the proposed method can effectively extract the complete road data at the roadway level.

Aiming at the problem of the insufficient generalization ability of traditional road extraction methods when applying to a new dataset, this paper proposes a cross-domain road extraction method that realized by feature-representation-transfer and encoder-decoder network. Firstly, a basic road extraction model based on encoder-decoder network is designed to segment the road from a single data source. Then, based on the structure of road extraction network and the principle of cycle-consistent, a cycle generative adversarial network for feature transfer of cross-domain imagery is used, which maps the feature of target city images to the domain of source data. Finally, the pre-trained road extraction model is used to segment the target domain images after the feature transfer, so that the cross-domain road extraction can be realized. The experimental results show that the proposed method improves the generalization ability of the road extraction network and can extract the road target from cross-domain images accurately and effectively. Compared with the results without feature transfer, the proposed method greatly improves the road extraction metric, and increases the F1-score by more than 50%. The proposed method does not require any annotation of the target domain images, nor does it need to fine-tune the road extraction network, while it only need to train the feature transfer model from the target domain to the source domain. Therefore, it has good application value.

Image matching is the process of obtaining corresponding points between two or more overlapping images by a specific algorithm. It is the critical step in the low-altitude photogrammetric data processing. The quality and efficiency of matching directly affect the subsequent data processing and the quality of mapping product generation. Therefore, image matching is one of the hot topics in the field of low-altitude photogrammetry and many relevant algorithms have been proposed. In this paper, the research status and prospect of image matching in low-altitude photogrammetry are described systematically. Firstly, the categories of image matching are summarized and can be generally divided into gray- and feature-based matching. We focus on feature-based image matching, e.g., point, line, and region-based features extraction and the relevant descriptors and similarity measures are described in detail. Besides, the latest image matching algorithms based on deep learning are listed, and the image matching methods involved in data fusion of various sensors on low-altitude platforms are mentioned.

With the rapid development of the reality capture, such as laser scanning and oblique photogrammetry, point cloud has become the third important data source following vector maps and imagery, and also plays an increasingly important role in scientific research and engineering in the fields of earth science, spatial cognition, and smart city, and so on. However, how to acquire valid and accurate three-dimensional geospatial information from point clouds has become the scientific frontier and the urgent demand in the field of surveying and mapping as well as the geoscience applications. To address the challenges mentioned above, point cloud intelligence came into being. This paper summarizes the state-of-the art of point cloud intelligence in acquisition equipment, the intelligent processing, scientific research and the major engineering applications, focusing on its three important areas:the theoretical methods, the key techniques of intelligent processing and the major engineering applications. Finally, the promising development tendency of the point cloud intelligence is summarized.

Due to the similarity of intra-class and dissimilarity of inter-class of high-resolution remote sensing image scene, it is difficult to identify some image scene class. In this paper, a new classification approach for high-resolution remote sensing image scene is proposed based on deep learning and metric learning. Firstly, a clustering center of each class is preset on the output features of deep learning model. Secondly, the Euclidean distance method is used to calculate the average central metric loss. Finally, the final loss function consists of a central metric loss term, a cross entropy loss term, and a weight and bias term. The goal of this method is to improve the classification accuracy by forcing intra-class compactness and inter-class separability. The experimental results show that the proposed method significantly improves the classification accuracy. Compared with state-of-the-art results, the classification accuracy ratios on RSSCN7, UC Merced and NWPU-RESISC45 datasets are increased by 1.46%, 1.09% and 2.51%, respectively.

Automatic extraction of buildings from remote sensing images is significant to city planning, popular estimation, map making and updating.We report several important developments in building extraction. Automatic building recognition from remote sensing data has been a scientific challenge of more than 30 years. Traditional methods based on empirical feature design can hardly realize automation. Advanced deep learning based methods show prospects but have two limitations now. Firstly, large and accurate building datasets are lacking while such dataset is the necessary fuel for deep learning. Secondly, the current researches only concern building's pixel wise semantic segmentation and the further extractions on instance-level and vector-level are urgently required. This paper proposes several solutions. First, we create a large, high-resolution, accurate and open-source building dataset, which consists of aerial and satellite images with both raster and vector labels. Second,we propose a novel structure based on fully neural network which achieved the best accuracy of semantic segmentation compared to most recent studies. Third, we propose a building instance segmentation method which expands the current studies of pixel-level segmentation to building-level segmentation. Experiments proved our dataset's superiority in accuracy and multi-usage and our methods' advancement. It is expected that our researches might push forward the challenging building extraction study.

According to the characteristics of the road features, an Encoder-Decoder deep semantic segmentation network is designed for road extraction of remote sensing images. Firstly, as the features of the road target are rich in local details and simple in semantic features, an Encoder-Decoder network with shallow layers and high resolution is designed to improve the ability of representing detail information. Secondly, as the road area is small proportion in remote sensing images, the cross-entropy loss function is improved, which solves the imbalance between positive and negative samples in training process. Experiments on large road extraction dataset show that, the proposed method gets the recall rate 83.9%, precision 82.5% and F1-score 82.9%, which can extract the road targets in remote sensing images completely and accurately. The Encoder-Decoder network designed in this paper performs well in road extraction task and needs less artificial participation, so it has a good application prospect.

Traditional target detection methods based on sliding window search paradigm and hand-craft based features are difficult to be applied to the multi-class target detection of very-high-resolution remote sensing images. In this paper,we proposed a deformable convolutional networks based multi-class target detection method by introducing deformable convolution layer and deformable RoI (Region-of-Interest) pooling layer. Specially,our method consists of two sub networks:a region proposal network aims to predict candidate regions from several layers with different filter size,and a region classification network for discrimination and regression. The quantitative comparison results on the challenging NWPU VHR-10 data set,large-scale Google Earth images, GF-2 and JL-1 images show that our method is more accurate and robust than existing algorithms.

Virtual geographic cognition experiment is an experimental framework to understand the perception,cognition,emotion and behavior of human to the environment.Since it has been proposed,experimental geography and other related field has been benefited from its theory and methodology.The coming big data era has brought with a huge massive of human cognitive,emotional and behavioral data,potentially bringing opportunities to establish new research paradigm in virtual geographic experiment,to help researchers look deeper into the man-land relationship.In this paper,we first proposed the framework of virtual geographic cognition experiment based on data-intensive scientific research paradigm,where we introduced how to treat human activity data and urban context data by integrating the theory in environmental psychology,and artificial intelligence.Second,we demonstrated the framework by introducing a case study,which has explored a series of visual factor in urban scene that would have an impact on depressing emotion of individual.

Man-made object detection is important for object detection from remote sensing images.In this paper we propose a variational approach for man-made object detection which formulates the man-made object detection problem as a problem of variational energy optimization.In this method,an image is firstly segmented into superpixels,and the saliency map by combining image features such as texture,color and gradient is computed.In second step,we construct an energy function with saliency,area,edge,texture and intensity variance constrains.The energy function is solved via variational method to obtain the foreground,which is the detected man-made objects.The proposed approach on several remote sensing images is evaluated and compared with the C-V model,MRF model and deep learning based semantic segmentation.Experimental results show that the proposed approach can effectively detect man-made objects on remote sensing images with low false alarm and false negatives rates.The comparison and analysis with deep learning based method are also presented.

This paper proposed an novel algorithm for exploring deep belief network (DBN) architectures to extract and recognize roadside facilities (trees,cars and traffic poles) from mobile laser scanning (MLS) point cloud.The proposed methods firstly partitioned the raw MLS point cloud into blocks and then removed the ground and building points.In order to partition the off-ground objects into individual objects,off-ground points were organized into an Octree structure and clustered into candidate objects based on connected component.To improve segmentation performance on clusters containing overlapped objects,a refining processing using a voxel-based normalized cut was then implemented.In addition,multi-view features descriptor was generated for each independent roadside facilities based on binary images.Finally,a deep belief network (DBN) was trained to extract trees,cars and traffic pole objects.Experiments are undertaken to evaluate the validities of the proposed method with two datasets acquired by Lynx Mobile Mapper System.The precision of trees,cars and traffic poles objects extraction results respectively was 97.31%,97.79% and 92.78%.The recall was 98.30%,98.75% and 96.77% respectively.The quality is 95.70%,93.81% and 90.00%.And the F1 measure was 97.80%,96.81% and 94.73%.

Traditional building extraction from very high resolution remote sensing optical imagery is limited by low precision and incomplete boundary. In this paper, a high-resolution remote sensing image building extraction method based on deep learning is proposed. Firstly, Principal Component Analysis is used to pre-train network structure in an unsupervised way and obtain the characteristics of remote sensing image. Secondly, an adaptive pooling model is proposed to reduce the feature information loss in the pooling process. The texture features are extracted by non-subsampled contour wave transformation and introduced to the network to improve the building extraction. Finally, the obtained image features are inputted into the softmax classifier for classification and building extraction results. A typical experiment areas selected. The comparison with typical building extraction method, the experimental results shows that the proposed method can extract the buildings with higher accuracy, especially the clearer and more complete boundary.

Snow cover in Qinghai-Tibetan plateau (QT plateau) is very important to global climate change. Because of the complex topography and high altitude, the recognition accuracies of existing snow cover products in QT plateau are significantly lower than flat areas. This paper proposed a new method of snow cover recognition for QT plateau based on deep learning. The multispectral remote sensing data from Chinese meteorological satellite FY-3A and the multiple geographic elements information are put together as the data sources, the insitu snow depth measurements and existing snow cover products are used for selecting the labeled samples. A stacked denoising auto-encoders (SDAE) network was built and trained for feature extraction and classification, this network can be used as a classifier for distinguishing the snow cover from cloud and other snow-free surface features. The recognition results are verified by snow depth data of meteorological station observations, verification results show that the recognition accuracy of this method is significantly higher than the snow product FY-3A/MULSS, which is using the same remote sensing data source FY-3A, and slightly higher than the widely used snow products MOD10A1 and MYD10A1,and the cloud coverage rate of this method is the lowest. According to the validation results, this method can effectively improve the accuracy of snow cover recognition, and reduce the interference of clouds.