Making mapping system automatically conducting map design and production through intelligent techniques has always been the goal pursued by the cartographic community and the frontier research direction of the International Cartographic Association. Since the 1980s, artificial intelligence has been applied in cartography, gradually solving the automation problems of some processes and improving the production efficiency of map making. However, the level of automation in key steps such as map design is still extremely low, which cannot meet the “customized” and “ubiquitous” mapping demand in the information age. Fortunately, since 2023, artificial intelligence technology represented by large language models such as GPT-4 and Gemini has made breakthroughs and achieved “quasi-general artificial intelligence”, which shows strong language comprehension, reasoning and expression ability. This paper explores the use of large models to improve the intelligence level of map making systems, aiming to establish a new generation of intelligent mapping theory and method system. This paper first analyzes the bottleneck problems of the existing digital mapping system and points out the necessity of establishing a new generation of intelligent mapping technology; then it analyzes the nature and capabilities of large models and demonstrates the sufficiency of establishing such a new generation; then it further analyzes the possibility and methods of combining them, proposes an intelligent mapping framework in the era of large models (e.g. situatedness map representation); finally, it discusses the key technical issues of situatedness map representation: “autonomous consciousness of mapping context”, “autonomous design and production of maps” and “autonomous human-computer interaction in situatedness ”.

Within the context of artificial intelligence generation (AIGC) and large language model (LLM), improving the intelligence level of generating geographic analysis models has gained widespread attention in the field. This paper proposes a geospatial information service hierarchical network model, named 5-HiNet. This model allows for a step-by-step description of heterogeneous geographic analysis models based on the five-layer hierarchical sub-network structure of demand description, abstract model, functional module, service interface, and functional instance, which depicts the realization process of geographic analysis models from the general to the specific. Within the five-layer hierarchical sub-network structure, the 5-HiNet can integrate massive expert knowledge embedded in the geographic analysis models and thus form a well-rounded domain knowledge system. Furthermore, the 5-HiNet can be coupled with the LLM to generate geographic analysis models automatically. A prototype system with a case study is developed in this paper to demonstrate the feasibility of the proposed 5-HiNet, and several research directions and insights for future study are provided.

With the enrichment of spatial data elements and the refinement of scene difference granularity, homogeneous electronic map products seriously affect the effect of spatial data aided decision-making. This paper proposes a conceptual framework for scene-based electronic map design, clarifies the definition and classification rules of application scene, determines four categories: application theme, objective environment, target user and mapping content, and summarizes the inherent requirements of differentiated scenes for electronic maps. According to the above requirements, this paper proposes a scene-based electronic map design method, which completes the electronic map design through specific design means such as scene-suggested symbol system construction and semantic-driven symbol combination expression. List three demand cases, the method of this paper is verified from the whole process of scene extraction, data organization and electronic map product design. The cognitive experiment shows that the scene-based electronic map designed based on the method of this paper is effective in visual processing, visual search and cognitive burden, and the expected consistency of cognitive effect, which provides a theoretical basis and practical path for scene-based electronic map design.

Drainage patterns recognition is essential for analyzing terrain and geomorphology, exploring geological minerals, and transforming river network data across various scales. However, traditional spatial statistical methods based on morphological and geometric features are not robust enough. To overcome this deficiency, graph convolutional methods have emerged as a popular solution. Nevertheless, these methods often focus narrowly on local features, disregarding the crucial global perspective necessary for comprehensive analysis. To address this issue, our study proposes a drainage pattern recognition method supported by graph Transformer. This method incorporates geometric knowledge by constructing river network graph structures using dual graphs. It integrates a GraphSAGE-based local learning module and a Transformer-based global learning module, training the graph Transformer model. Experimental results demonstrate that our method achieves 94% accuracy in accurately recognizing drainage patterns by combining local segment composite features and global river network morphology features. This outperforms the 1^st-ChebNet and GraphSAGE methods, presenting a promising approach for intelligent drainage pattern recognition.

The collection, cleansing, and annotation processes of high-quality remote sensing datasets typically entail substantial costs. Therefore, the remote sensing datasets can be regarded as intellectual properties. However, remote sensing datasets also face threats such as theft, unauthorized usage and redistribution. In order to safeguard the copyright of datasets, we propose an object detection dataset protection method based on backdoor watermarking and region of interest (ROI) encryption. The algorithm embeds object-generation watermark triggers into the original dataset and utilizes an ROI encryption algorithm to encrypt the dataset. During the watermark embedding phase, random samples are selected from the original dataset, and the triggers are embedded into random positions within the samples. During the dataset encryption phase, the ROIs in the annotation files are first initially encrypted. Then, disturbances are added within the encrypted ROIs. Finally, a unique random key is generated for each user based on a hash function, and perform secondary encryption on the initially encrypted annotation files. During the dataset decryption phase, only authorized users can decrypt the encrypted dataset, where the encrypted annotation files are restored to correct ROIs. Thereby obtaining the decrypted legitimate dataset. In the phase of asserting copyright on suspected models, a watermark test set is constructed with the object-generation watermark triggers. This test set is then inputted into the suspected model for prediction. If the watermark prediction success rate exceeds a preset threshold, it is considered that the model has utilized the protected dataset during training. Extensive experiments have demonstrated that this method effectively protects dataset copyrights without compromising dataset quality. The watermarking algorithm exhibits strong robustness against fine-tuning attacks and pruning attacks.

The inverse modeling utilizing signal-to-noise ratio (SNR) data enables the inversion of sea surface height (SSH) and its variations. However, the accuracy and stability of the inversion process hinge on the precision of the initial values and the temporal continuity of the SNR data. Further investigation is required into the performance of inverse modeling based on multi-mode and multi-frequency SNR data for inverting sea-level changes and its application in tidal analysis. This study introduces a dynamic correction for sea surface variations into the inversion results of the Lomb-Scargle periodogram (LSP), which is utilized for initializing parameters in the inverse modeling process. This approach yields stable and uniform high-precision SSH inversion values, which are then employed to conduct tidal harmonic analysis. Three stations with large tidal ranges, namely MAYG, BRST, and SC02, were selected for inverse modeling and inversion experiments using their one-year multi-mode and multi-frequency SNR data. Algorithm validation was conducted through comparative analysis with in-situ SSH measurements from tide gauges. The results indicate that the root-mean-square error (RMSE) of the SSH inversion via inverse modeling is 5.97 cm for MAYG, 8.78 cm for BRST, and 2.38 cm for SC02, demonstrating centimeter-level accuracy in SSH inversion. When compared with the tidal harmonic analysis results of the observed SSH, the annual and monthly fitting residuals exhibit high consistency in terms of mean square error. The mean absolute error (MAE) of the extracted tidal constituent amplitudes is better than 1 cm, and the MAE of the extracted tidal phase lags is within 3°. Both the tidal components and non-tidal water levels extracted from the tidal analysis demonstrate high consistency. Therefore, the inversion of SSH using multi-mode and multi-frequency SNR data can serve as a viable alternative to in-situ SSH measurements for tidal harmonic analysis.

The extraction of precise water level information from satellite altimetry data is crucial for long-term monitoring of lake and reservoir levels. Using Qinghai Lake as a case study, a 20-year dataset is compiled by integrating altimetry data from four different satellites: Envisat, SARAL, Sentinel-3A, and Sentinel-3B. In this study, an innovative algorithm is proposed for the extraction of water levels from multi-source satellite altimetry data. This algorithm integrates adaptive weighting and deviation matching techniques to enhance the accuracy and reliability of water level extraction. Adaptive weighting involves the selection of suitable correction algorithm models based on various environmental conditions and the determination of unique weight parameters for each altimetry data source, thus standardizing the data. The deviation matching method quantifies qualitative data to maximize the precision of water level extraction. Additionally, an artificial intelligence framework is established to automate and integrate the water level extraction process, streamlining the workflow. Experimental results demonstrate that applying adaptive weighting to multi-source altimetry data characteristic values enables reasonable classification and exhibits strong correlations. This approach provides a robust foundation for generating high-precision, long-term water level records. When combined with the deviation matching method, the correlation between daily extracted water levels and actual measurements exceeds 0.9. By setting a correlation coefficient threshold of 0.8, reliable water level extraction for up to a 5-month duration in a single extraction is achievable. To address long-term water level extraction requirements, a methodology is introduced that combines single-day and multi-day extraction, resulting in the construction of 12 years of continuous high-precision water level records. The obtained results exhibit correlation coefficients exceeding 0.9, mean absolute error (MAE) values within the range of 1.5 cm to 2.0 cm, and root mean square error (RMSE) values ranging from 2.0 cm to 2.5 cm. This success underscores the practical value of the data processing algorithm and model in the context of water level extraction and prediction. In conclusion, this research demonstrates the feasibility and utility of combining artificial intelligence with satellite altimetry in constructing long-term, high-precision water level records for small-scale water bodies.

Typhoon is a kind of disastrous weather which seriously affects human production and activities. The effective monitoring of typhoon status is of great significance to avoid and reduce the loss of people's life and property. Water vapor is the main driving force of typhoon development. In this paper, a four-parameter model (TDOPA-4) for estimating typhoon's movement based on the time difference of precipitable water vapor (PWV) arrival (TDOPA) was proposed, which requires PWV time series of several stations near the typhoon's path as inputs. The initial linear velocity, initial direction angle, linear acceleration and angular velocity of the typhoon were estimated, and then the velocity and direction angle of the typhoon at any moment can be further calculated. The typhoon Lekima and Bailu in China in 2019 were selected as cases study for verifying the model. First, the accuracy of ERA5 datasets derived PWV (ERA5-PWV) was tested by using the ground-based global navigations satellite systems derived PWV in China, which was used to estimate the typhoon's movement based on the TDOPA-4 model. Then, the spatial distribution and temporal characteristics of PWV during typhoon's period were analyzed, and the moving trend and relationship between typhoon and water vapor were studied. Finally, the typhoon products provided by the China Meteorological Administration (CMA) and People's Government of Zhejiang Province (PGZP) were as reference, and the velocity and directional angle of Lekima and Bailu estimated from the TDOPA-4 model were evaluated. The results showed that the mean Bias, RMSE and STD were 0.25, 6.39, 6.38 km/h and 4.68°, 21.59° and 20.83° compared to the results from the CMA, respectively. Compared with the results of the PGZP, the mean Bias, RMSE and STD are 2.23, 5.41 and 4.27 km/h, respectively. The above results show that the TDOPA-4 can reflect the water vapor transport of typhoons, and provide a supplementary method for monitoring typhoon activities.

MODIS water vapor products have become important atmospheric water vapor productsdue to their advantages of high spatial resolution, however, due to uncertain factors such as precipitation, cloud mask, and surface reflection spectrum, the inversion accuracy is limited. To effectively improve the quality of MODIS water vapor products, this article analyzes nonlinear factors such as cloud, land cover type, pixel attitude, time, and position, and a MODIS PWV neural network differential correction model integrating multiple nonlinear factors is constructed for the first time. Firstly, the correlation between MODIS PWV and GNSS PWVat the same site is analyzed, PWV_diff (the difference between the two) is taken as the target value of the neural network model, and the input information for the model is 19 nonlinear factors such as cloud mask confidence, land cover type, periodic day of year, periodic hour of day, sensor zenith angle, solar zenith angle, sensor azimuth angle, and solar azimuth angle. Compared with traditional linear correction model, the root mean square error (RMSE) of corrected MODIS PWV is reduced from 3.271 3 mm to 2.360 2 mm, and the accuracy is improved by 27.85%. The performance of the model isalso evaluated by using high spatiotemporal ERA5 PWV data as a reference value. The experimental results show that the corrected RMSE of the corrected MODIS PWV using the proposed model in this paper is 2.037 4 mm, which improves the accuracy by 57.99% compared to the uncorrected MODIS PWV (RMSE=4.850 3 mm); furthermore, neural network differential correction models are constructed for MODIS PWV products under four different confidence levels provided by cloud mask products, the results show that the RMSEs of MODIS PWV products with cloud confidence levels >99%, >95%, >66%, and <66% are increased by 60.03%, 61.21%, 55.72%, and 54.57% compared to uncorrected MODIS PWV. This indicates that the model constructed in this article has high universality in improving the accuracy of MODIS PWV products under various cloud covers, and is expected to provide high-precision water vapor information for climate change and rainfall forecasting research.

An important issue that limits the effectiveness of the application of time-varying satellite gravity field model is the influence of noise in the spherical harmonic coefficients, which is mainly manifested in the signal recovery of the Earth's surface mass field by the north-south striping (NSS) noise and random noise. In order to effectively suppress the NSS noise, a multi-domain combined gravity recovery and climate experiment (GRACE) de-striping filter is proposed in this paper, which combines the advantages of the Swenson & Whar (S&W) decorrelation filter and the multi-channel singular spectrum analysis in the spatial domain (SSAS) filter, and utilizes the least-squares method and the multichannel singular spectrum analysis to take into account the signal characteristics in the spectral and spatial domains. The results show that, compared with the traditional filters, the combined filter proposed in this paper can adapt to the filtering needs of different latitudes, effectively eliminate the NSS noise effects in the middle and high latitudes and low latitudes, retain more geophysical signals, and improve the possibility of observing small-scale signals.

The daytime star map is characterized by high background noise and low signal-to-noise ratio, which lowers the efficiency of traditional star extraction methods. Moreover, robustness of single frame extraction methods is poor due to the negative factors of daytime star map, including star energy weakness, noise sensitivity and star contour irregularity. Therefore, a star extraction method with modification of apparent motion is proposed for daytime infrared star map processing. With application of camera parameter conversion model, image coordinates of the target star were predicted using the data of station position, observation time and ephemeris. Further, relative displacements of the star in image sequence were precisely calculated using the exposure intervals, which led to precise register of the image superposition on star area, and finally resulted to accurate star extraction. Experimental results show that the success rate of star extraction with the proposed method can reach 100%, which is 163% higher than that of traditional algorithms. Meanwhile, the time consumption was reduced to 23%. Signal-to-noise ratio was also enhanced since the influence of random error was greatly reduced, and the centroid extraction accuracy was improved by 72% compared to single frame methods.

The generalized nonlinear Gauss-Helmert model is a unified expression of explicit and implicit nonlinear function adjustment models that consider the errors of the dependent variable or the whole variable. Aiming at the problem of non-convergence of its Gauss-Newton iterative solution algorithm when the difference between the initial value and the true value is large, the parameter estimation method of the generalized nonlinear Gauss-Helmert model that integrates the homotopy method and nonlinear least squares is proposed. Starting from the nonlinear least-squares adjustment criterion that introduces the homotopy parameter, the system of differential equations for solving the generalized model parameters and the fixed-step prediction formula for tracking the homotopy curve with the Newton's correction formula are derived, and the approximation formula for calculating the residual vector of the implicit function model is given. The complexity of computing the system of differential equations is reduced by introducing the Kronecker product and the matrix straightening operation into the derivation process in order to avoid computing the cubic matrix. The feasibility of the method is verified through three experiments, including distance positioning that only considers the error of the independent variable, pseudo-distance positioning that considers the satellite coordinate error and ranging error, trilateration network that considered the errors in the known coordinates, and circular curve fitting that considers the error of plane coordinates. The experimental results show that the new method converges to a larger range of initial values.

There is a large amount of gravel in the Gobi and desert areas in western China. Accurate inversion of soil water content in these areas is of great significance for ecological and meteorological environment monitoring, afforestation and water conservancy project construction in northwest China. However, existing SAR soil moisture inversion models assume that soil is composed of fine particulate matter, without considering the influence of gravel. In this paper, a new method for soil moisture inversion on high gravel surface using polarimetric SAR (PolSAR) data is proposed. First, the backscattering of high gravel surface is modeled as surface scattering caused by ground and volume scattering caused by gravel. For surface scattering, the advanced integral equation model (AIEM) and Oh model are used for soil moisture inversion. For volume scattering, the dense medium radiative transfer (DMRT) model is used to invert soil moisture. Finally, the weighted summation of the inversion results of the two parts is used as the final inversion result. The inversion accuracy of soil moisture inversion model was evaluated by using the field soil moisture measured data and ALOS-2 PolSAR data in Wuhai city, and compared with the commonly used soil moisture inversion methods. The results showed that the accuracy of this method was significantly improved when it was applied to the soil moisture inversion on high gravel surface (traditional soil moisture retrieval method: R²=0.35; new method: R²=0.60).

Landslide cause serious harm to human living environment. The method of manually identifying the landslides is time-consuming and the hidden area is not easy to detect. The use of remote sensing image to identify the landslides can accurately and quickly realize the landslide disaster warning and rescue. With the rapid development of deep learning, semantic segmentation has been widely used in the field of landslide remote sensing image recognition. Aiming at the problems such as error recognition and image edge information loss in the current landslide image segmentation model, this paper proposes a landslide segmentation model MLFIF-Net, which integrates multi-layer feature information fusion. The model uses MobileNetv3-Small as the main trunk network to improve the feature extraction ability of the model. At the same time, a cascade spatial pyramid pool module is constructed to enhance the texture features of landslide images and obtain multi-scale information. An efficient channel attention module is used to focus on image features, and a multi-layer feature information fusion structure is designed to enhance the edge information of images, so as to improve the segmentation effect of the model. The experimental results show that the accuracy of the proposed model on the landslide data set of Bijie city, Guizhou province is 96.77%, the average accuracy of the class is 95.61%, and the average interaction ratio is 87.69%. Compared with SegNet and other six segmentation models, its segmentation accuracy is better, and it can accurately identify the target area and highlight the edge details of the landslide image.

Road markings are important traffic sign information, and onboard LiDAR point clouds provide high precision 3D coordinates and reflectance intensity information for their extraction. Due to factors such as scanning distance and target material, the different object may exhibit similar intensity values, causing interference in the extraction of road markings. Wear and aging during road use can also damage the original structure of the markings, resulting in discontinuities after extraction. In addition, the diversity of road markings and their varying occurrence frequencies in practice can lead to low classification accuracy for categories with fewer samples in the segmentation network extraction results. To address these issues, this paper proposes a two-stage segmentation and classification extraction method that accurately extracts various types of markings and has topological robustness. Firstly, a multi-layer perceptron is used to adaptively learn the relationship between intensity and its influencing factors, and to perform intensity correction on the road point clouds. Secondly, the semantic segmentation network link spatial topology net (LST-Net) is proposed to segment all road markings, which captures line structure information using row-column convolution and attention mechanisms, and is trained with topological punishment to determine the positions of markings. Finally, YOLOv5 is used to detect the markings, and a separate classification network is trained to address the issue of sample imbalance in segmentation. Experiments are conducted on three sets of point clouds from different driving scenarios, and the results show that our approach achieves a marking extraction accuracy of 94.1% and a recall rate of 95.6%, demonstrating strong practicality and effectiveness.