Existing NeRF-based methods for reconstructing large-scale scenes face challenges in visual quality and rendering speed due to spectral biases and extensive sampling requirements. Recent 3DGS-based methods for real-time rendering of 3D objects and small scenes outperform NeRF, but several issues persist when extending these techniques to large-scale scenes. These include robust rendering in weak-texture areas, effective densification under memory constraints, finer detail rendering, and achieving natural lighting transitions. To address these, we introduce a geometry-aware 3DGS method for efficient real-time rendering of large scenes. First, we propose a geometry-guided anchor point initialization that reduces noise away from structural surfaces and generates new points in weak texture areas, particularly for large-scale datasets. We also present a structure-aware joint densification strategy combining surface- and curvature-based densification, ensuring Gaussian points are near structural surfaces and increasing density in low-curvature areas. Additionally, we propose a hash grid-assisted, viewpoint-sensitive feature enhancement scheme to improve detail rendering and natural lighting transitions. Our method achieves superior rendering quality compared to state-of-the-art methods while maintaining reasonable memory usage. Extensive experiments across 16 scenes, including 11 from five public datasets (MatrixCity-Aerial, Mill-19, Tanks & Temples, WHU, and UrbanScene3D) and five self-collected scenes from SCUT-CA and plateau regions, demonstrate its generalization capability. Codes are available athttps://github.com/SCUT-BIP-Lab/Geo_gs.

The continuous demand for coastal zone resource exploitation and marine ecosystem protection challenges satellite remote sensing observation of shallow sea apparent optical properties. The existing satellite retrieval method of the LiDAR Attenuation Coefficient (Klidar) relies on water column photons, so it can not achieve accurate water quality measurement in shallow water depths. In this study, the Ice, Cloud, and Elevation Satellite-2 (ICESat-2) data was innovatively fused with Sentinel-3 image to retrieve comprehensiveKlidar. First,Klidaris calculated in offshore areas using ICESat-2 water column photons and in nearshore areas using seafloor photons. ICESat-2-derivedKlidarwas then integrated with Sentinel-3 image, employing Bayesian-optimized CatBoost (BO-CatBoost) to invert the spatial distribution of both offshore and nearshore areas. Finally, the retrieval results’ weights were assigned based on Sentinel-3’s band ratio, yielding comprehensiveKlidarresults for the shallow seas. The proposed method successfully obtained accurate spatial distribution ofKlidarin Passu Keah, Culebra, and Marquesas Keys. The results show that the Coefficient of Determination (R2), Mean Relative Error (MRE), and Mean Absolute Error (MAE) for the three study areas range from 0.7402~0.8557, 0.0405~0.0885m−1, and 0.0033~0.0055m−1, respectively. The Root Mean Square Error (RMSE) ranges from 0.0043~0.0077m−1. It has higher reliability than Sentinel-3's Kd(532). As satellite remote sensing improves, the method will efficiently map the LiDAR attenuation coefficient across vast areas, offering reliable data for resource development, water quality, and ecosystem conservation.

Feature learning of airborne laser scanning (ALS) point clouds is challenged by both the limited annotated samples and imbalanced class distribution. An intuitive way involves pretraining on a well-annotated source dataset and fine-tuning on a limited target dataset. However, cross-domain challenges such as heterogeneous point cloud density, varying terrain features, and inconsistent object categories complicate transfer learning for 3-D land cover classification. In this article, we address these issues by separating the cross-domain ALS point cloud semantic segmentation into two subsequent subtasks, i.e., the cross-domain transfer learning subtask and the intradomain class-incremental learning subtask, and we use a well-annotated photogrammetric point cloud dataset as the source dataset. To mitigate domain discrepancies, the first subtask employs domain adversarial training to learn from base categories that are shared between source and target datasets. Then, the second subtask incrementally learns new categories that are specific within the target dataset using an incremental feature-semantic distillation module and a semantic adversarial learning module while retaining base category knowledge. Experimental results evaluated on three ALS point cloud datasets (ISPRS, DALES, and H3D) with different semantics show state-of-the-art cross-domain performance with few labeled samples. Compared with few-shot learning methods, our method shows promising generalization ability particularly on domain-specific categories, greatly alleviating the dependence on ALS point cloud annotations.

Nearshore bathymetry plays a vital role in marine ecological monitoring, coastal zone management, and nautical chart production. Satellite remote sensing has become a primary approach for bathymetry due to its broad spatial coverage, data abundance, and cost-effectiveness. However, single-temporal methods are often limited by transient noise, such as clouds, waves, and ship wakes, which reduce the accuracy. In response, multi-temporal methods have gained increasing attention. By fusing multi-temporal satellite images, these methods can suppress environmental noises, thereby improving the accuracy and stability of depth retrieval. Nonetheless, most existing multi-temporal methods lack physical constraint mechanisms, making them vulnerable to the influence of anomalous images during fusion, and they often overlook temporal variations in water properties and seafloor substrates. These limitations hinder their stability in complex nearshore environments. To address these challenges, this study proposes a novel Multi-Temporal Spaceborne Bathymetry (MTSB) method that integrates ICESat-2 laser altimetry data with multi-temporal Sentinel-2 images. The proposed method uses laser-derived depth points as physical constraints to guide high-quality image fusion. It then applies a stratified modelling strategy based on column laying and sediment classification to adapt to varying conditions. Experiments were conducted in three representative nearshore regions: Culebra, Puerto Rico; Oahu and Niihau, Hawaii. Results show that the MTSB method outperforms existing methods in accuracy. Specifically, the Root Mean Square Error (RMSE) was 1.32 m in Culebra, 1.48 m in Oahu, and 1.61 m in Niihau. Moreover, MTSB showed superior adaptability to environmental disturbances. This research provides a new perspective for high-precision satellite-derived bathymetry and demonstrates promising potential for filling bathymetric data gaps in remote reefs and other inaccessible marine areas.

LiDAR diffuse attenuation coefficient (${K}_{\text {lidar}}$) describes the laser attenuation degree in water, and Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) is expected to extend the observation of this parameter to satellites. However, the ICESat-2 observation is limited to the satellite trajectories, and it is hard to realize the coverage observation of inland water. Therefore, this study proposes a method fusing active and passive remote sensing data based on machine learning and uses ICESat-2 data and Sentinel-3 images to retrieve$K_{\text {lidar}}$of inland waters. First,$K_{\text {lidar}}$is retrieved from ICESat-2 water column photons. Second, the retrieved$K_{\text {lidar}}$and Sentinel-3 images are used to train machine learning models to generate$K_{\text {lidar}}$covering water bodies. Experiments were carried out in Xiaolangdi Reservoir using multitemporal data. The results show that the performance of CatBoost is better than that of random forest (RF) and XGBoost, and the consistency reaches 68.42%. Compared with in situ data, the root mean square error (RMSE) of the proposed method is$0.0555~\text {m}^{-1}$, which is 61.05% higher than the RMSE of the Sentinel-3 product. By combining laser data with multispectral images, water quality observation close to field measurement can be retrieved from satellite platforms.

Existing airborne laser scanning (ALS) point cloud semantic segmentation approaches are limited by their overreliances on sufficient point-wise annotations that further confine their generalization ability to new scenes. To overcome these problems, a novel three-stage multi-prototype relational network (Thr-MPRNet) is proposed for few-shot ALS point cloud semantic segmentation by transferring knowledge from well-annotated photogrammetric point clouds. In MPRNet, a 3D few-shot learning structure containing a feature learner and a relation learner is built to learn meta-knowledge from multiple point-wise tasks, and a multi-prototype generator is designed to represent the semantic distribution of point clouds that can dynamically adapt to large-scale scenarios. Then, to transfer knowledge across different domains, MPRNet is trained in a unified framework with three task-based learning stages. Prior knowledge is first meta-learned from the source photogrammetric point clouds and then transferred to novel target datasets with a few labeled ALS point clouds. Finally, the MPRNet can be flexibly generalized to the unlabeled target ALS point clouds without further retraining from scratch. In the experiments, the SensatUrban dataset is used as the source photogrammetric point clouds, and two ALS point cloud datasets (ISPRS and DALES) are used to evaluate the few-shot semantic segmentation ability of the proposed method. The experiments demonstrate that Thr-MPRNet obtains promising generalization performance on different target datasets. More importantly, it outperforms supervised networks with 10% labeled samples. In summary, the proposed method achieves state-of-the-art cross-domain semantic segmentation performance and greatly alleviates the dependence on ALS point cloud annotations.

The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) equips with a novel photon-counting LiDAR system, which can generate underwater reflections in nearshore environments. However, due to the water reflection, scattering, and absorption, the distribution of bathymetric photons in the nearshore data varies with depth. The existing bathymetric photon extraction algorithms need more adaptability to seafloor topography. The changing density of bathymetric photons and the fluctuation of underwater topography make the noise removal of nearshore data full of challenges. This study proposed a bathymetric photon extraction algorithm using pre-pruning quadtree isolation (PQI). Firstly, the pre-pruning step judges whether to stop the growth of quadtree in advance during quadtree isolation (QI) to avoid excessive division of noise photons. Secondly, the maximum inter-class variance algorithm (also called the Otsu method) obtains the best threshold of isolation depth and extracts bathymetric photons. The algorithm was tested on the Florida coast. The results show that the PQI algorithm can wholly and accurately extract bathymetric photons with different acquisition times from the data. The F1-score of the extracted results is 93.96%. This study provides an intelligent solution to processing bathymetric data in nearshore environments worldwide.

Photon-counting light detection and ranging (LiDAR) Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) enables the drafting of global elevation maps. However, vegetation cover, terrain undulation, and residual noise in signal photons substantially reduce the accuracy of ground photon extraction. Existing ground photon extraction algorithms do not consider the factors influencing photon extraction, and the threshold setting lacks a theoretical basis. This study proposed a photon-extraction algorithm with scenario adaptability. First, the cloth simulation (CS) was adapted with a terrain index (TI) to extract ground photons; based on this, the cloth breakage concept was proposed to remove residual noise. We tested the algorithm in Denali National Park and compared its results with those of other extraction algorithms. The results showed that the TI was robust and consistent with the actual terrain; the adaptive CS achieved the best accuracy and precision under different canopy heights and terrains. The mean absolute error (MAE) and root mean square error (RMSE) of extracted photons were 0.95 and 3.41 m, respectively. This study provides a solution to estimate ground elevation using photon-counting LiDAR data.

The Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) is the world’s first satellite-borne photon-counting laser altimeter with unprecedented detection performance. Noise removal is an important process applied to raw data and determines the quality of the end product. Assuming that the sparse spatial distribution of noise photons makes them more easily isolated than signal photons, we propose a noise-removal algorithm without input parameters based on quadtree isolation. MATLAS was used to evaluate the performance of our algorithm. We compare our algorithm to the improved density-based spatial clustering of applications with noise (DBSCAN) algorithm. Experimental results show that our algorithm accurately extracts signal photons from raw data and is superior to the improved DBSCAN in accuracy and time efficiency. This novel algorithm makes it possible to efficiently remove noise from photon-counting light detection and ranging (LiDAR) data.

Planetary mapping products play a significant role in landing site selection, surface operations, and scientific investigations. This article describes the techniques and processes used to create geometrically controlled image mosaics for the candidate landing regions of China’s first Mars exploration mission (i.e., Tianwen-1) using Mars Express (MEX) High-Resolution Stereo Camera (HRSC) images. To deal with the extremely complex situations due to large data volumes (hundreds of images), various imaging conditions (e.g., 5°–80° incidence angles), and long time spans (>10 years), we developed corresponding methods and in-house software for the photogrammetric processing of MEX HRSC images. According to the characteristics of planetary images and multiline pushbroom imaging, we optimized the algorithms of establishing control network, eliminating blunders, setting weights, and processing abnormal images. The generated MEX HRSC image mosaics with a resolution of 12.5 m delivered a high relative accuracy (< 1 pixel). Compared with the existing Mars global image mosaics, the generated image mosaics exhibit significantly better spatial resolution. The techniques and methods that we developed to solve the complicated photogrammetric processing problems can also be applied to produce global or large-scale Mars mapping products using existing and new returned orbital images.

The digital orthophoto map of planetary bodies plays a significant role in landing site selection, traverse path planning, and scientific research. However, existing orthorectification methods in the planetary mapping community exhibit low-computational efficiency for linear pushbroom images. To solve this problem, this letter presents a novel orthorectification method based on an improved back-projection algorithm. The back-projection algorithm is based on the geometric constraints of the central perspective plane (CPP) and is further improved to process linear pushbroom images with distortions. Specifically, we segment the linear array in the focal plane into multiple line segments and find the exact CPP using simple analytical geometric calculations. The proposed method was fully tested and evaluated with well-known planetary mapping software packages, namely, the integrated system for imagers and spectrometers (ISIS). The experimental results demonstrated that compared with ISIS, the proposed orthorectification method can increase the computational efficiency by more than fivefold and deliver consistent geometric accuracy. The proposed orthorectification method greatly enhances geometric processing capabilities for massive planetary remote sensing images.

Non-negative matrix factorization (NMF) learns to approximate a non-negative matrix by the product of two lower-rank non-negative matrices. Since NMF usually learns sparse representation,it has been widely used in pattern recognition and data mining. However, NMF cannot deal with the datasets that contain offsets. To remedy this problem, Laurberg and Hansen proposed affine NMF (ANMF) by jointly learning the offset vector, but the proposed multiplicative update rule neither guarantees non-negativity constraints over factor matrices nor converges sufficiently rapid. In this paper, we adopt the well-known hierarchical alternating least squares (HALS) algorithm to solve ANMF. Since the update of offset vector is in the same frame of updates of factor matrices, HALS is quite suitable for solving ANMF and the experimental results on simulated datasets validate its efficiency.