ABSTRACT Water salinity is a key characteristic of natural lakes, with its spatial and vertical variations altering water density and affecting aquatic organisms. Traditional lake water salinity monitoring, reliant on in-situ measurements, has limited the comprehensive exploration of both horizontal and vertical water salinity distribution, thereby hindering accurate characterization of volumetric water salinity and total salt content within the entire lake. To address this, our study introduces a novel framework for volumetric salinity estimation, using Pung Co, a deep endorheic lake on the Tibetan Plateau (TP), as a case study. First, we developed a model using machine learning algorithms, with remote sensing data and hydrological and topographical features, to estimate surface water salinity. Secondly, we analyzed field-surveyed vertical water salinity profiles to model the relationship between lake water depth and salinity, revealing the vertical water salinity variation characteristics. Finally, we constructed a gridded water column method to precisely estimate the lake’s total salt content. Results showed that the extreme gradient boosting model (R 2 = 0.85, RMSE = 0.13 g/L, MAPE = 0.91%) effectively estimated surface water salinity. The modeled water salinity was characterized by significant horizontal and vertical variability. Horizontally, the water salinity was higher in the north and west and lower in the south and east, with a lake-wide average of 10.91 g/L. Vertically, the water salinity was strongly influenced by depth, exhibiting a sharp change near the thermocline before stabilizing. When surface water salinity was below 11.20 g/L, it increased with depth. When surface water salinity exceeded 11.20 g/L, it decreased with depth, with both converging toward a stable value of approximately 11.20 g/L. Using our gridded approach, the total dissolved salt content in Pung Co was estimated to be approximately 4.51 × 107 tons. This study establishes a quantitative framework that shifts salinity estimation from a two-dimensional surface assessment to a three-dimensional volumetric estimation, offering significant implications for understanding microbial diversity and the ecological effects of salinity stratification in high-altitude deep lakes.

ABSTRACT As large language models increasingly exhibit hallucinations such as refusal to respond, generation of non-executable code, and poor readability in geospatial code generation tasks, establishing a systematic and quantifiable evaluation framework has become essential for advancing their application in GIS. This paper introduces the GeoCode-Eval framework, the first comprehensive evaluation framework for LLMs in geospatial code generation. Grounded in three dimensions – cognition and memory, understanding and interpretation, and innovation and creation – the framework addresses eight competency levels, including platform and tool cognition, functional knowledge, dataset recognition, information extraction, and various code-related tasks. To support this, the GeoCode-Bench benchmark was developed, consisting of 5,000 multiple-choice questions, 1,500 true/false questions, 1,500 fill-in-the-blank questions, and 1,000 coding tasks. Using six indicators, namely executability, accuracy, readability, location correctness, content correctness, and summary completeness, the study evaluates twelve representative models spanning four categories, including DeepSeek-Coder-V2 and GeoCode-GPT (7B). A combination of analytical methods, including entropy weighting, the coefficient of variation, skewness, and kurtosis, is applied to examine model capability distribution, indicator distribution, code type characteristics, and error type patterns. Results show consistent performance in tool cognition and code summarization, while significant performance gaps persist in code generation, completion, and correction. Common errors include data type and syntax issues. This study provides a quantifiable foundation for the evaluation of capabilities and future optimization of LLMs in geospatial code generation, thereby extending the application boundaries of LLMs in GIS and offering valuable insights into the development of evaluation methodologies for LLM applications in other vertical domains.

ABSTRACT Achieving high-quality rendering with Gaussian Splatting for unconstrained image collections is critical for advancing 3D reconstruction. Currently, the methods using a single multi-layer perceptron and convolutional neural network to predict distractors in input images often suffer from errors and is insufficient in rendering details for appearance modeling. In this study, we propose the U3GS framework for Gaussian Splatting under unconstrained image collections. Specifically, we develop a multi-scale tri-plane feature encoding module that integrates scene features across multiple scales and predicts Gaussian attributes. Furthermore, we model image appearance by combining global and local appearance embeddings, allowing the framework to adapt to illumination variations within the images. Finally, we design an uncertainty estimation module based on visual foundation model to predict distractors in input images, and apply the uncertainty-guided loss to ensure reliable target rendering within the scene. We evaluate our framework on the NeRF On-the-go and Phototourism datasets, demonstrating its effectiveness in distractor removal within highly occluded environments and achieving high-quality rendering across images with diverse appearances.

ABSTRACT This paper aims to propose a nearshore water depth retrieval method using Chinese Gaofen-3 (GF-3) synthetic aperture radar (SAR) imagery. By applying a deep learning-based multi-layer perceptron (MLP) approach, one-dimensional wave spectra are inverted from 12 quad-polarized GF-3 images. Then, SAR-derived wave spectrum is divided into wind-wave and swell based on the ocean wave overshoot phenomenon. Based on the wave dispersion relation over gentle slopes, water depth is iteratively derived using both wind-wave and swell spectra. The proposed method is validated in the Yangtze River estuary. Compared with General Bathymetric Chart of the Oceans (GEBCO) and Earth Topography 2022 (ETOPO2022) water depth, it is found that accuracy of water depth using swell spectrum is higher than that using wind-sea spectrum, indicating that the error is within ± 2 m. Furthermore, the error analysis presents that swell-based depth estimates exhibit smaller discrepancies compared to wind-sea results, particularly with significant wave heights (SWH) below 2 m and peak wavelengths exceeding 100 m. These findings demonstrate the impact of wave characteristics on inversion accuracy, revealing that swell spectra provide higher reliability under specific hydrodynamic conditions. Additionally, the results underscore the capability of GF-3 SAR for coastal underwater observation.

ABSTRACT The downwelling diffuse attenuation coefficient (K d) is a key optical parameter for ocean color science, yet accurately estimating K d across diverse water types remains a challenge due to variations in water constituents. To address this, the study developed a neural network-based method (K d_NN) to directly estimate the diffuse attenuation coefficient at 490 nm (K d(490)) from remote sensing reflectance (R rs). K d_NN was trained with globally collected K d(490) measurements, representing a wide range of bio-optical conditions. Notably, all measurements were classified into 30 water types and quality-controlled using OC_3S to ensure the quality of the training dataset. The results show that K d_NN delivers robust performance across various water types, reducing the Mean Absolute Percentage Error (MAPE) by approximately 10% for most water types. Spatio-temporal analysis further reveals that K d_NN effectively addresses the underestimation issue observed in NASA’s standard algorithm in coastal waters and captures real oceanic variations regardless of changes in water types. In low- and mid-latitude regions, the Relative Percentage Difference (RPD) between our algorithm and NASA’s remains within ± 5%, whereas in coastal and polar regions the RPD rises to 20% or more, suggesting that the results in these regions need to be revisited. Moreover, an analysis of 20 years of MODIS data demonstrate that K d_NN offers substantial improvements over NASA’s method, which underestimates K d(490) by about 33% in China’s coastal waters. Consequently, K d_NN has the potential to provide high-quality global K d(490) inversion products, supporting marine ecosystem monitoring and management.

ABSTRACT In machine learning-based fractional cover estimation, the fractional cover distribution in training samples critically influences model construction and, consequently the accuracy of the estimations. While some studies have descriptively compared the accuracies of machine learning-based estimations across training sets derived from different sampling methods, a significant gap remains in quantitatively analyzing how the fractional cover distribution in training samples affects accuracy. This study aims to bridge this gap by introducing descriptors for fractional cover distribution in the training set and establishing mathematical relationships between these descriptors and the accuracy of fractional cover estimation. We employed the Dirichlet distribution to characterize the joint fractional cover of multiple land classes and the Beta distribution for single-class cover. Subsequently, two descriptors were developed: the Kullback-Leibler (KL) divergence, measuring the similarity of fractional cover distributions for the target class between the training and test sets, and the geometric angle, representing the fractional cover distributions of the target class in the training set at the same KL divergence. Fractional cover estimation was performed using random forest regression, with accuracy assessed on an independent test set. The relationships between the KL divergence and accuracy, and between the geometric angle and accuracy at the same KL divergence, were modeled using univariate linear models and harmonic models, respectively. The combined effects of these descriptors on accuracy were further analyzed using coupled harmonic analysis and generalized additive models. Our experimental results, using both simulated and real data, demonstrated the effectiveness of these models. Given the strong explanatory power of the KL divergence in the accuracy of fractional cover estimation, we encourage researchers to report detailed statistical information of both training and test sets, enriching the understanding of model performance in fractional cover estimation.

ABSTRACT Asteroid gravity inversion is a method used to infer the internal density distribution of small celestial bodies through external gravitational measurements. In recent years, gravity inversion techniques based on neural networks have seen rapid development, such as GeodesyNet, which utilizes a fully connected SIREN neural network. The traditional SIREN neural network performs well in capturing low-frequency features, such as regular shapes or large-scale density distributions. However, it has limitations in learning high frequency features when processing irregular shapes or abrupt changes in the inner density field. This study introduces an optimized architecture that synergistically integrates SIREN networks with quadratic neurons to augment the SIREN network’s capability in capturing high-frequency gravitational features of asteroids, which seeks to improve the accuracy and stability of inversion processes within the GeodesyNet framework. Eros, Itokawa, 67P/Churyumov-Gerasimenko, and Bennu are selected for testing and evaluation. The results indicate that the density distributions computed by this combined network are more in line with the actual conditions. The maximum relative root mean square error (relRMSE) for homogeneous test models decreased by 6.1% for 67P, and through differential training, the maximum relRMSE for non-homogeneous asteroid surfaces was reduced by 3.0% for Bennu. Moreover, the improvement from GeodesyNetsQ is larger when fewer iterations, and thus fewer data points are used, which will be more helpful to be applied in practical asteroid missions with limited measurements.

ABSTRACT This paper introduces a novel globally consistent dense RGB-D SLAM system designed to address the challenges of online dense reconstruction under fast camera motion. To achieve robust pose tracking and globally consistent map building, we present a random optimization pose estimation framework that integrates both geometric and photometric information. Furthermore, we propose a voxel revisitation-based loop closure detection scheme that relies solely on depth maps. Innovatively, we introduce a local curvature feature, to address loop closure detection when photometric information is unavailable, efficiently extracting 3D geometric information directly from 2D images. In addition, we leverage voxel revisitation observations to accelerate the search for potential loop closures. This is achieved by maintaining a dynamic active map and detecting voxel outflow or revisits, thereby enhancing loop closure detection efficiency. This method is the first voxel revisitation-based loop closure detection scheme that relies entirely on depth information. Combined with a random optimization pose estimation framework utilizing both geometric and photometric features, it adapts well to fast camera motion, enabling accurate and globally consistent dense map construction. Extensive evaluations on public datasets demonstrate the effectiveness of our method under fast camera motion and highlight its significant advantages over existing methods.

ABSTRACT Countries are now shifting from one-off lunar exploration missions to the establishment of a lunar base for long-term sustainable surveys. The selection of a suitable site plays a crucial role in sustaining the base’s operation and facilitating scientific research. Previous site selection methods have been limited in their ability to consider multiple objectives and factors simultaneously, often relying heavily on expert knowledge. To address this challenge, this paper proposes a comprehensive site selection method that integrates multiple engineering and scientific objectives/factors. Through site selection factor comprehending, gridded data processing and site evaluation based on Technique for Order of Preference by Similarity to Ideal Solution (TOPSIS), the proposed method balances subjective and objective considerations and outputs quantitative descriptions of the most suitable candidate sites. This paper applies the proposed method to the range of 75°S − 90°S. The results reveal four large areas with continuous high suitability near the South Pole, while a small area located at the rim of Cabeus B crater shows the highest suitability. Qualitative and quantitative comparisons are conducted between the areas selected by the proposed method, areas recommended by previous studies, and areas announced by Artemis III. The candidate areas identified in this paper outperform others in terms of a wider range of site selection factors, making them valuable references for future lunar base site determination. Additionally, the proposed site selection method holds relevance for similar site selection tasks where multiple objectives are set. It offers a more automated and quantifiable approach, minimizing manual intervention and providing robust results for decision-making processes.

ABSTRACT Accurate assessment of the physicochemical parameters of rice is critical for increasing agricultural production and ensuring food security. The leaf area index (LAI), leaf chlorophyll content (expressed as the soil plant analysis development (SPAD)), and canopy chlorophyll content (CCC) of the rice at single and multiple periods were estimated using abundance correction indicators by integrating the normalized difference red edge index (NDRE) and refined abundance information, with the goal of developing a high-precision and unified estimation model (AC-NDRE) of the physicochemical parameters of rice that can be adapted to various temporal and spatial scales. The results showed that while not significant, the accuracy of predicting the LAI, SPAD, and CCC of rice using the NDREgreen with the soil background removed was better than that of the NDRE model. The NDREgreen yielded the highest coefficients of determination (R 2) of 0.69, 0.71, and 0.70, root mean square errors (RMSEs) of 2.35, 2.19, and 97.61, relative RMSEs (RRMSEs) of 31.87%, 5.61%, and 33.06%, respectively. Furthermore, severe instability was observed in the accuracy of the NDREgreen model on both the spatial and temporal scales. The AC-NDRE-I, which is based on strong light, and the AC-NDRE-II, which is based on moderate light, exhibited evident advantages in estimating the LAI/CCC and SPAD, respectively. The optimal LAI, SPAD, and CCC estimation accuracies based on the AC-NDRE were R 2 values of 0.83, 0.74, and 0.82, RMSEs of 1.73, 2.06, and 76.41, RRMSEs of 23.49%, 5.29%, and 25.88%, respectively. The AC-NDRE approach achieved a stable performance under complicated circumstances. The conclusions of this study indicated that the AC-NDRE-based method for estimating the LAI, SPAD, and CCC of rice could effectively address the issues of a limited model estimation accuracy caused by soil background, NDRE saturation during the middle to late growth stages of rice, and shaded leaves.