Hyperspectral image classification must model long-range spectral structure while handling mixed pixels and uncertain class boundaries. Existing selective state-space classifiers use input-dependent parameters, and explicit local reliability information can guide their parameter generators. We introduce NUP-Mamba, which derives neutrosophic-inspired truth and indeterminacy descriptors, together with complementary falsity, from spectral tokens for selective-parameter generation. These descriptors condition the input-injection term $B$, discretization step $\Delta$, and output gate of a bidirectional Mamba encoder, while the transition matrix $A$ remains stable and input independent. Local reliability thereby controls how evidence enters and leaves the recurrence while preserving its transition dynamics. Matched controls and a capacity stress test on HanChuan support this descriptor-to-parameter pathway. Cross-dataset evaluations show that the benefit depends on the scene and module configuration. Uncertainty maps and parameter diagnostics associate high indeterminacy with elevated error risk and changes in generated $B$ and $\Delta$. The results support uncertainty-conditioned parameter generation as a complementary mechanism for spectral-spatial state-space classifiers.
Shallow water bathymetry is of great significance for coastal zone resource development and ecological conservation. Owing to the complex marine environment and scattered distribution of islands and reefs, traditional in-situ acoustic or optical bathymetric methods fail to meet the demands of largescale or remote coastal topographic mapping. The advanced spaceborne photon-counting lidar system aboard ICESat-2 is capable of shallow water bathymetry with global coastal zone coverage. However, ICESat-2 bathymetric signals are highly susceptible to noise swamping due to solar radiation and water scattering, especially for weak beams. In this study, a coastal bathymetry method based on multi-scale robust approximation of topography (CB-MRAT) is proposed to process strong and weak beam data under the interference of various environmental noises. Validation experiments were conducted on islands and reefs across different marine regions worldwide, and the results demonstrate that the proposed method can effectively and robustly extract signal photons to achieve accurate bathymetric measurements. Specifically, the mean F1-score for signal photon extraction reaches 97.7%, and the mean root mean square error (RMSE) of bathymetry is 0.43 m, which maintains a high degree of consistency with in-situ bathymetric measurements. Compared with the ATL24 built-in algorithm and the photon denoising algorithm based on adaptive elevation difference thresholding (AEDTA), CB-MRAT enables accurate and complete detection of signal photons and yields superior bathymetric accuracy for datasets with varying pulse energies and photon densities, with the mean RMSE reduced by 65.9% and 76.5%, respectively. Notably, CB-MRAT achieves comparable bathymetric performance for both strong and weak beam data, which attests to its excellent adaptability. This study presents a robust approach for highprecision and automated ICESat-2 bathymetry, thereby providing technical support for shallow water bathymetric mapping of remote islands and reefs and even global coastal zones.
Great Wall relic extraction from high-resolution remote sensing (RS) imagery is important for cultural heritage discovery, mapping, and monitoring. However, this task remains challenging because mountainous relics often appear as narrow linear structures with weak visual contrast and are easily affected by vegetation cover, terrain shadows, and complex backgrounds. To address these challenges, this paper proposes TFCL-Net for mountainous Great Wall relic extraction using optical RS imagery and terrain priors. Specifically, the terrain-aware fusion (TAF) module integrates DEM and ridge information at different network depths, allowing the model to better exploit terrain priors for relic localization and representation. The connectivity-aware learning (CAL) module improves the continuity of narrow wall predictions and reduces local discontinuities. In addition, a structure-conditioned image synthesis (SCIS) strategy generates synthetic images with clearer relic traces, and an adaptive filtering step removes unreliable samples. To support model training and evaluation, we construct a multisource Beijing Great Wall relic dataset that includes high-resolution optical RS imagery, DEM data, ridge prior maps, and polygon-level ground truth annotations. A three-level evaluation system, including pixel-level, centerline-level, and relaxed-level metrics, is adopted to assess both segmentation accuracy and structural quality. Experimental results across repeated runs show that TFCL-Net achieves 73.33% IoU and 84.61% F1-score, as well as 94.13% centerline quality and 89.42% relaxed IoU. These results indicate that the proposed framework can support Great Wall relic discovery, mapping, and monitoring in mountainous environments.
Leaf area index (LAI) retrieval is fundamental for large-scale monitoring of forest growth, evolution and health. The current LAI retrieval algorithms mainly differentiate between coniferous forests and broad-leaved forests, while the increasing coverage of coniferous and broad-leaved mixed forests worldwide has brought about more sources of bias in the forest LAI products. It is difficult but significant for the classical LAI inversion algorithm to address the complex spatial heterogeneity of mixed forests. In this study, a coniferous and broad-leaved mixed forest dataset was established for evaluation experiments, consisting of 3750 simulated mixed-forest plots with specific structures generated by the 3D radiative transfer model LESS, 21 field measured mixed forest plots and 76 mixed forest plots collected from the NEON dataset. The stochastic radiative transfer (SRT) model used in MODIS LAI products was utilized to evaluate the LAI retrieval bias in the application of coniferous and broad leaved mixed forests. A lookup table (LUT) approach integrating both conifer and broadleaf species was proposed to improve LAI retrieval performance in mixed forests. Compared with the conventional broadleaf and coniferous LUTs, the integrated LUT demonstrated moderately improved applicability in both simulated mixed-forest plots (R²integrated/R²broadleaf/R²coniferous = 0.905/0.894/0.731, RMSEintegrated/RMSEbroadleaf/RMSEconiferous = 0.128/0.132/0.074 m2·m-2, RIintegrated/RIbroadleaf/RIconiferous = 100.0%/96.8%/35.0%) and real mixed-forest plots (R²integrated/R²broadleaf/R²coniferous = 0.511/0.434/0.333, RMSEintegrated/RMSEbroadleaf/RMSEconiferous = 1.652/1.733/1.106 m2·m-2, RIintegrated/ RIbroadleaf/ RIconiferous = 80.4%/72.2%/22.7%). The simulation experiments revealed that the effect of mixing ratio, tree density, and vertical distribution of tree species on LAI retrieval in mixed forests indicated regular biases controlled by the characteristics of coniferous and broad-leaved tree species. The bias mechanisms were further explained in real sample plots by an attribution analysis based on Random Forest and SHapley Additive exPlanations (SHAP). This study systematically evaluates the bias of the LAI retrieval algorithm based on the SRT model in coniferous and broad-leaved mixed forests, which provides theoretical support for the improvement of the LAI retrieval algorithm for coniferous and broad-leaved mixed forests and more accurate monitoring of complex forest surface information.
Low-light remote sensing (RS) images often cover large scenes with spatially continuous structures and minorblueground objects at varying scales, making it difficult to recover scene-wide structures while preserving fine details under nonuniform illumination and noise. Closely related RS methods generally organize frequency modeling around predefined amplitude–illumination and phase–structure roles. In contrast, we propose Conv-Fourier-Metaformer (CFformer), a self-attention-free MetaFormer that adaptively coordinates local spatial cues and global frequency context at the feature level. To recover scene-wide structures while preserving local details, its SF-Mixer combines parallel spatial and frequency modeling with input-adaptive channel recalibration and frequency-position modulation. To represent ground objects at varying scales, MSG-FFN employs multiple receptive fields for scale-aware feature modulation. Because shallow encoder features retain useful details but may also contain noise-contaminated responses, AFM adaptively regulates their injection through skip connections to preserve details while limiting the transfer of noise-contaminated responses. Extensive experiments on multiple low-light RS datasets and general low-light benchmarks demonstrate that CFformer achieves the best overall performance among the compared methods while better preserving structural consistency and fine-grained ground-object details.
