Abstract:Road intersections are important parts of road traffic network, the location and type of which are the basic data for various application services such as high-definition map and automatic driving.However, little attention has been paid to classify road intersections compared with the great number of researches on the road boundary extraction from mobile laser scanning point clouds.Here, we propose a road intersection classification method based on dynamic graph neural network.First, we employ geometric and spatial distribution differences of supervoxels to extract road boundaries from ground.Then we calculate the curvature of road boundary points and detect road intersections according to the curvature difference in sliding windows.Finally, we build a dynamic graph neural network to identify the T junction and regular intersections.The experimental results show the proposed method can accurately detect most road intersections.

Abstract:Retrieving surface Soil Moisture (SM) from CYGNSS has attracted great attention in recent years, yet its accuracy and efficiency should be further improved.Here, a pre-classification strategy combined with Artificial Intelligence (AI) algorithm is proposed to predict SM from CYGNSS data.This strategy can improve the accuracy of SM estimation due to the use of AI algorithm and is versatile and easy to use.The field SM data of China in 2018 are used as real ground truth values for modeling and prediction.The results show that the predicted SM is in good agreement with the referenced SM.The correlation coefficient (R) between SM retrieved from CYGNSS and ground truth data is as high as 0.8, and the mean values of Root Mean Square Error (RMSE) and unbiased root mean square error (ubRMSE) are 0.059 cm³/cm³ and 0.050 cm³/cm³, respectively.Meanwhile, the results show that the AI-based pre-classification strategy not only significantly improves the accuracy of SM estimation from CYGNSS, but is applicable to other regression and prediction fields for its good generalization and expansibility.

Abstract:At present, deep learning has been widely applied in object detection, such as vehicle detection.In this paper, the deep learning EfficientDet model was analyzed, and its advantages in vehicle detection were confirmed.A phased adaptive training model was constructed to avoid local optimum in training process, then it was used to detect vehicles from both short and long distance.The detection results showed that compared with detection methods based on Cascade R-CNN and CenterNet, the proposed model was superior in terms of computational complexity, time consumption and detection accuracy.Meanwhile, further analysis figured out the optimal detection distance and angle.Finally, an example is given to verify that the proposed method can be applied to a large range of vehicle detection.

Abstract:At present, the research of road extraction using vehicle-borne point cloud mainly focuses on structured roads with regular road cuts.However, some roads in reality are flanked by grass, and the extraction method of structured roads is no longer applicable.To address this problem, an automatic road surface extraction method integrating the moving window height difference and neighboring point serial number difference is proposed.First, the adjacent point space is used to realize the extraction of scan lines, and the road boundary points are extracted according to the moving window height difference as well as the adjacent point serial number difference.Next, the road boundary points are fitted with RANSAC algorithm.Afterwards, the road surface point cloud is extracted according to the linear programming principle and filter processing.Finally, two sets of road data are used to test the proposed method.The completeness of road surface extraction is 99.79% and 99.52%, the correctness is 99.91% and 99.62%, and the quality is 99.70% and 99.15%, for the two datasets respectively.The experimental results show that the proposed method is applicable for road surface extraction of structured road and unstructured road with grass on the roadside.

Abstract:Architectural models can be constructed from 3D point cloud data due to the advances in spatial information acquisition technology.However, most of the existing model construction methods rely on manual interaction, which is extremely laborious and time consuming, thus are not applicable to the construction of large-scale urban models and related applications.This paper proposes an approach to build the main structure model based on wireframe analysis.First, filter and normalize the original building point cloud, and extract the multi-level boundary polygons of the building to comprehensively describe the building boundary structure; then divide the boundary polygons at each level into multiple rectangular primitives that are simple and easy to control.A robust hierarchical rectangular connection and analysis algorithm is used to generate a complete main structure model of the building.Finally, experiments on three sets of complex structure and severely occluded building point cloud data show that the approach has good performance and can robustly handle the point cloud data of buildings with missing data and different point density.The proposed approach takes into account the structural differences at different heights of the building, and realizes the high-precision construction of the wireframe model for the main building structure.

Abstract:Airborne LiDAR Bathymetry (ALB) system can quickly and efficiently obtain the integrated overwater and underwater data of sea islands, reefs and their adjacent areas.However, due to the fact that most of the measurement areas are shallow near-shore waters with slow terrain changes, the obtained point cloud is low in density and large in thickness, resulting in rare registration characteristics.Few studies have been done on the registration of ALB data due to the difficulty in extracting their homonymous features.To address this problem, we employ three registration methods including Fast Point Feature Histograms (FPFH), Longest Common Subsequence (LCSS) and Generalized Iterative Closest Point (GICP) to register the ALB point cloud data in the South China Sea.The registration performance comparison shows that the LCSS line sequence outperforms the other two methods in registration accuracy and reliability.Moreover, the LCSS can tackle the problems of single information and noise in the corresponding feature matching, improve the robust estimation of corresponding points in the feature curve, and enhance the robustness of airstrip data registration.It can be concluded that the LCSS is an effective solution for ALB data registration.

Abstract:Water level information is an important parameter for studying changes in the water cycle, even the climate and ecological environments.It is of great significance to monitor the water level changes in near real-time and with high precision.However, traditional water level gauge is a relative water level measurement approach in a small range and has a high cost.The Global Navigation Satellite System Interferometric Reflectometry (GNSS-IR) uses the Signal-to-Noise Ratio (SNR) data from the GNSS receiver installed on the coast to estimate the water level change, which provides a new monitoring approach for water level measurement with high accuracy and near real-time.Here, 30 days of GPS, BDS and GLONASS SNR data from a multi-GNSS station at Badong station in the upper reaches of the Yangtze River are used to estimate the water level changes, which are then compared with the in-suit water level station observations.The results show that GNSS-IR obtains centimeter-level water level results with the minimum RMSE of 6.43 cm.The estimation results of GPS and GLONASS system at L1 frequency and BDS system at B1 frequency are better, and the accuracy of GLONASS L2 frequency is lower than that of other frequencies.The joint multi-GNSS system improves the time resolution of GNSS-IR estimation of water level changes.

Abstract:Soil moisture is a key parameter of the water cycle and energy budget in terrestrial ecosystems.Land data assimilation system can provide spatio-temporally continuous soil moisture data, however, its low spatial resolution limits the further application.Here, the soil moisture output in 0-10 cm soil layer from China Meteorological Administration Land Data Assimilation System (CLDAS-V2.0) was downscaled from 6 km to 1 km in North China by three single models (gradient boosting machine, deep feedforward neural network and random forest) and a Stacking ensemble learning method.The downscaled results for period of April to October in 2019 show that the four downscaling methods can reflect the temporal and spatial variation of soil moisture in North China and somehow alleviate the overestimation of CLDAS products.Both the spatial distribution details and accuracies are improved compared with original CLDAS soil moisture data.Furthermore, the Stacking ensemble learning method outperforms the other three in downscaling performance, including its highest correlation coefficient with observed data (R=0.756 8) and lowest error (RMSE=0.050 5 m³/m³, Bias=-0.005 2 m³/m³).Meanwhile, the downscaled results by Stacking ensemble learning are also highly correlated with the dynamic changes of soil moisture, with lowest RMSE and bias compared with station observations, followed by random forest and deep feedforward neural network.

Abstract:In recent years, ozone has become the primary air pollution in major cities of China.Therefore, tropospheric ozone products are very important for monitoring near surface ozone concentration.However, existing ozone products cannot meet the requirements of high spatial & temporal resolution monitoring.Here, we use a spatial & temporal fitting method to repair the total ozone data from OMI, and then inverse the tropospheric ozone distribution by a residual approach.The results show that the spatial & temporal fitting method is qualitatively and quantitatively superior to both Kriging interpolation and inverse distance weighted interpolation, indicated by its smaller RMSE and MAE.The tropospheric ozone profile obtained by the proposed method is consistent with official ozone products of OMI/MLS according to their correlation coefficient (R) up to 0.82.

Abstract:Here, we try to realize the spatialization of the added values of the secondary and tertiary industries in Fujian province, by adding thermal infrared remote sensing data to land use data and night light remote sensing data, and considering surface temperature as well.The results show that compared with previous method of using land use data plus night light remote sensing data, the adding of thermal infrared remote sensing data and consideration of surface temperature improves the GDP spatialization model in coefficients of determination (R², 0.966 vs. 0.743, 0.870 vs. 0.776 for the secondary and tertiary industry, respectively) and simulation accuracy (MRE, 20.45% vs. 72.60%, 19.82% vs. 60.10%, for the secondary and tertiary industry, respectively).Further, we take Xiamen as an example to show the potential of thermal infrared remote sensing data in GDP spatialization model.It is found that the proposed method can greatly improve the spatialization of the added values of the secondary and tertiary industries, indicated by its high consistency with reality.The results of this paper can provide reference for regional economic development planning.

Abstract:Human action recognition has always been a hot topic in computer vision research and widely applied in virtual reality, short video, etc.Meanwhile, the fast development of deep learning in recent years has also inspired the action recognition algorithms.Compared with traditional methods, the action recognition algorithms based on deep learning have advantages of strong robustness and high accuracy.Here, we make a survey on the action recognition algorithms based on deep learning proposed in recent years, and focus on those developed from two-stream network and 3D convolutional network, then summarize their performances and positive results, and finally make prospects in this field.

Abstract:In order to improve the existing time series algorithms for precipitation forecasting, this paper proposes a time series precipitation forecasting model (DeepAMogLSTM) based on improved dual-stage attention mechanism.The algorithm can be divided into two parts.In the input attention stage, a three-layer attention mechanism is designed to pay multiple attention to the input sequence; while in the time attention stage, the hidden state most relevant to the target value is selected to calculate the long-term correlation of the time sequence.In this manner, input features can be stably selected and input into the prediction structure.The algorithm also introduces Mogrifier LSTM (Long Short-Term Memory) to enhance the feature representation ability.The model uses preprocessed automatic station data from 2016 to 2019 and ECMWF weather field model data for integrated forecast, and corrects the model forecasts using observation data of the same period.The experimental results show that the evaluation indexes of the model are improved in the 2-hour precipitation nowcasting, in which the maximum square root error is 1.877 mm, the maximum average absolute error is 0.727 mm, and the goodness of fit (R²) is 0.783.At the same time, the modeled precipitation fits actual precipitation in spatial distribution, which is better than the correction effect of other models.

Abstract:In order to optimize and standardize the application and approval process of marine scientific data access and realize the online collaborative office of the application and approval process, the workflow of the application and approval of marine scientific data access has been recounted in detail and its functional module based on B/S framework has been designed in this paper.Application and approval online can effectively promote the management and application of marine scientific data, and improve the service level and work efficiency of marine science data management department.