HLC-GS: Risk-Map-Guided Height-Layer Consistency Gaussian Splatting for DSM Reconstruction from Optical Satellite Imagery
本文提出HLC-GS方法,通过风险图引导的高度层一致性高斯点渲染来解决从光学卫星影像重建DSM时的高度层混合问题。
本文提出HLC-GS方法,通过风险图引导的高度层一致性高斯点渲染来解决从光学卫星影像重建DSM时的高度层混合问题。
本文提出MomentBA,通过使用局部相似响应的二阶空间矩来推导各向异性对应不确定性,改进了几何优化中的不确定度估计,从而提高单目视觉里程计精度。
Existing feedforward 3D foundation models, constrained by central perspective projection, struggle to accommodate the push-broom imaging geometry of satellites, limiting their applicability in multi-view satellite 3D reconstruction. This work proposes a lightweight adaptation framework that requires no fine-tuning of the backbone network: it selects an optimal view sequence through geometric consistency constraints, parameterizes push-broom rays into geometric tokens using rational function models, and introduces a ray-direction-aware adapter to inject these tokens into a frozen Transformer backbone. This approach achieves, for the first time, an effective integration of physical imaging geometry with deep feedforward architectures, significantly enhancing accuracy and robustness in digital surface model (DSM) generation and demonstrating the critical role of explicit geometric embedding and optimized view selection.
Cloud contamination severely compromises the reliability of optical remote sensing imagery for near-real-time land use and land cover (LULC) mapping. To address this challenge, this work proposes CloudLULC-Net, an end-to-end heterogeneous SAR-optical fusion framework that directly predicts LULC maps from cloud-contaminated Sentinel-2 images together with temporally adjacent Sentinel-1 SAR data, accompanied by the introduction of a large-scale benchmark dataset, CloudLULC-Set. The method innovatively integrates optical reliability modulation, adaptive heterogeneous information aggregation, a unified semantic mapping Transformer, and semantic anchor-guided optimization to effectively mitigate semantic uncertainty under cloud occlusion. Experimental results demonstrate that the model achieves 86.60% overall accuracy, 83.29% F1 score, and 73.51% mIoU on CloudLULC-Set, significantly outperforming both reconstruction-first and existing end-to-end approaches while maintaining robust performance across varying cloud coverage levels.
This study addresses the limitations of low-cost MEMS inertial measurement units (IMUs), whose hardware-constrained accuracy often falls short of high-precision navigation requirements. To overcome this challenge, the work introduces a conditional diffusion-based generative framework—the first to apply diffusion models to IMU signal enhancement—leveraging a U-Net architecture to synthesize high-fidelity virtual IMU signals. The model takes high-accuracy IMU data as a prior and low-cost IMU measurements as conditional inputs. By doing so, it effectively transcends the performance ceiling of low-cost sensors, yielding significantly improved localization and orientation estimates compared to raw measurements. Experimental validation in airborne mapping demonstrates that the enhanced signals produce denser, more consistent point clouds, confirming the method’s efficacy in multi-sensor integrated navigation systems.
本文提出HLC-GS方法,通过风险图引导的高度层一致性高斯点渲染来解决从光学卫星影像重建DSM时的高度层混合问题。
本文提出MomentBA,通过使用局部相似响应的二阶空间矩来推导各向异性对应不确定性,改进了几何优化中的不确定度估计,从而提高单目视觉里程计精度。
Existing feedforward 3D foundation models, constrained by central perspective projection, struggle to accommodate the push-broom imaging geometry of satellites, limiting their applicability in multi-view satellite 3D reconstruction. This work proposes a lightweight adaptation framework that requires no fine-tuning of the backbone network: it selects an optimal view sequence through geometric consistency constraints, parameterizes push-broom rays into geometric tokens using rational function models, and introduces a ray-direction-aware adapter to inject these tokens into a frozen Transformer backbone. This approach achieves, for the first time, an effective integration of physical imaging geometry with deep feedforward architectures, significantly enhancing accuracy and robustness in digital surface model (DSM) generation and demonstrating the critical role of explicit geometric embedding and optimized view selection.
Cloud contamination severely compromises the reliability of optical remote sensing imagery for near-real-time land use and land cover (LULC) mapping. To address this challenge, this work proposes CloudLULC-Net, an end-to-end heterogeneous SAR-optical fusion framework that directly predicts LULC maps from cloud-contaminated Sentinel-2 images together with temporally adjacent Sentinel-1 SAR data, accompanied by the introduction of a large-scale benchmark dataset, CloudLULC-Set. The method innovatively integrates optical reliability modulation, adaptive heterogeneous information aggregation, a unified semantic mapping Transformer, and semantic anchor-guided optimization to effectively mitigate semantic uncertainty under cloud occlusion. Experimental results demonstrate that the model achieves 86.60% overall accuracy, 83.29% F1 score, and 73.51% mIoU on CloudLULC-Set, significantly outperforming both reconstruction-first and existing end-to-end approaches while maintaining robust performance across varying cloud coverage levels.
This study addresses the limitations of low-cost MEMS inertial measurement units (IMUs), whose hardware-constrained accuracy often falls short of high-precision navigation requirements. To overcome this challenge, the work introduces a conditional diffusion-based generative framework—the first to apply diffusion models to IMU signal enhancement—leveraging a U-Net architecture to synthesize high-fidelity virtual IMU signals. The model takes high-accuracy IMU data as a prior and low-cost IMU measurements as conditional inputs. By doing so, it effectively transcends the performance ceiling of low-cost sensors, yielding significantly improved localization and orientation estimates compared to raw measurements. Experimental validation in airborne mapping demonstrates that the enhanced signals produce denser, more consistent point clouds, confirming the method’s efficacy in multi-sensor integrated navigation systems.