Short-term extreme precipitation nowcasting remains challenging because convective rainfall evolves through coupled moisture, dynamical, and microphysical processes that are only partially observed by any single sensor. Radar composite reflectivity (CR) directly describes hydrometeor structures, whereas Global Navigation Satellite System (GNSS)-derived precipitable water vapor (PWV) provides complementary moisture information before and during precipitation development. In complex terrain, however, raw GNSS-PWV fields are affected by station elevation differences. In this study, DEM information is therefore used to support GNSS-PWV elevation reduction before model ingestion, while the neural network fuses DEM-corrected GNSS-PWV and radar CR as two dynamic input streams. This paper proposes M2-CroT, a spatiotemporal fusion framework that combines DEM-corrected GNSS-PWV with radar CR for extreme precipitation nowcasting in complex terrain. The main contributions are: 1) a DEM-assisted elevation reduction method that improves the terrain consistency of GNSS-PWV grids before model ingestion; 2) a dual-branch Vision Transformer (ViT) feature extractor that captures long-range spatial dependencies from sparse moisture fields and dense radar imagery; and 3) a gated cross-attention module that adaptively regulates the contribution of moisture-background and reflectivity features during fusion. Experiments on real-world North China datasets show that the proposed model improves the Critical Success Index (CSI) for extreme precipitation events by 57% over the original TrajGRU radar-only baseline. These results support the value of combining DEM-assisted PWV preprocessing with adaptive multi-modal fusion for high-resolution precipitation nowcasting in complex terrain.
Hyperspectral images (HSIs) usually contain hundreds of highly correlated spectral bands, which may introduce severe redundancy, high computational cost, and the Hughes phenomenon. Band selection (BS) reduces spectral dimensionality by selecting a compact subset of original bands while preserving their physical interpretability. However, existing methods mainly exploit pairwise spectral or spatial dependencies, leaving complex high-order correlations insufficiently characterized. Moreover, manifold preservation and sparse band selection are rarely integrated into a unified framework. To address these issues, this paper proposes a high-order graph-regularized sparse projection method for unsupervised hyperspectral band selection. Based on region-level samples generated by superpixels, the method constructs a high-order graph through multi-step neighborhood propagation, thereby alleviating the limitations of traditional first-order graphs in characterizing complex sample relationships and more comprehensively capturing the intrinsic structure of HSI data. Moreover, self-representation learning, graph regularization, and row-sparse projection are integrated into a unified optimization framework. Graph consistency and Laplacian regularization preserve high-order structural relationships and local manifold information, while row sparsity enables explicit estimation of band importance. Consequently, the proposed method selects a compact band subset with strong discriminative capability and representativeness while reducing spectral redundancy. The source code is available at https://github.com/CPREgroup/hogrsp.
Cloud is essential components of the weather and climate system, influencing it by altering the radiation balance and participating in the hydrological cycle. Cloud Optical Thickness (COT) and Cloud Effective Radius (CER) are critical parameters describing cloud optical and microphysical properties. This study investigates the reflectance characteristics of the Particulate Observing Scanning Polarimeter onboard GaoFen-5(02) satellite (POSP/GF-5(02)) and proposes a reflectance correction method based on the Unified Linearized Vector Radiative Transfer Model (UNL-VRTM). Using the corrected reflectance, COT and CER are retrieved through an optimized dual-channel method (0.865-2.25 μm), with channel selection guided by sensitivity analysis. The retrieval results are further assessment with Moderate Resolution Imaging Spectroradiometer (MODIS) products. Results show that, under the current correction method, the original POSP reflectance used for cloud target is underestimated by 15.61%, 51.59%, and 44.38% at 0.865, 1.61, and 2.25 μm, respectively. Region case studies demonstrate that after reflectance correction, the fractions of retrievals within the predefined accuracy thresholds increase by an average of 61.83% for CER and 47.08% for COT. Further case analysis shows that retrieval accuracy varies with cloud field heterogeneity and surface complexity, with CER accuracy decreasing by 50.56% relative to homogeneous cloud field and COT accuracy decreasing by 13.23% relative to dark ocean surfaces. These results indicate that the proposed reflectance correction framework substantially enhances the accuracy of cloud property retrievals from POSP observations, providing improved capability for cloud remote sensing and supporting further studies of weather and climate processes.
Remote sensing intelligence has long faced a fundamental trade-off between model accuracy and interpretability. Deep learning models offer strong nonlinear representation capacity but remain intrinsically opaque, whereas traditional spectral indices, though physically interpretable, rely on fixed human-designed formulations that cannot adapt to complex or evolving observation scenarios. To address this gap, we propose a data-driven symbolic discovery framework that evolves explicit, task-adaptive spectral models from multispectral observations. The framework comprises two complementary paradigms. The Symbolic Transformer (ST) performs Single-Index Discovery for threshold-based land-cover extraction using the Directional Otsu Separability Fitness. The Symbolic Classifier (SC) performs Synergistic Index-Set Discovery for multiclass discrimination by jointly evolving calibrated class-score functions under multiclass hinge loss. Experiments on hyperspectral and multispectral benchmarks show competitive performance against conventional machine-learning, spectral deep-learning, and symbolic-learning baselines. Across 12 hyperspectral datasets, SC ranked second in average Recall and F1-score among 14 evaluated methods. Compact ST expressions recover established index structures and reveal spectrally plausible relationships. Compact SC expressions likewise support physically informed inspection of band relationships, whereas more complex SC models retain explicit class scores and auditable inter-class margins. Together, the two paradigms support compact spectral-index discovery and auditable multiclass decision modeling. Furthermore, the codes used in this study are available at https://github.com/djw-easy/symbolearn.
An extraction method is devised for accurate relative permittivity εr of soil samples using angular-domain reflection-only spectral free-space measurements. Different from the angular-domain extraction method which uses an error model considering the return loss function Hi(ω) and the product of transmitting and receiving transfer functions Ht(ω) and Hr(ω), the proposed error model considers not only Hi(ω) and Ht(ω)Hr(ω) but also the feedback loss transfer function Hf (ω). In addition, the proposed method generalizes the normal-incidence extraction method using Ht(ω), Hr(ω), and Hf (ω) to any oblique incidence. An improvement in εr determination by considering the Hf (ω) term in our angular-domain extraction algorithm is shown by a numerical analysis and free-space measurements of two soil samples. It is also validated by measurements that relative errors in εr determination decrease if Hf (ω) term is considered in the extraction procedure. The proposed extraction method was also tested for a combination of different incidence angles and for an inaccurate sample thickness.
Hyperspectral (HS) images and synthetic aperture radar (SAR) data provide complementary information for land cover interpretation. However, their cross-modal collaboration remains insufficient in multi-satellite observation systems. Moreover, both remote sensing (RS) modalities are sensitive to environmental interference, hindering critical feature acquisition. To overcome these challenges, a generative modality-specific learning enhanced collaborative fusion network called GenMSLnet is proposed for HS and SAR joint classification. Our core innovation is modality-specific feature learning, directly guided by the RS inherent properties. The framework couples a multi-scale dual-stream generative adversarial network (MDGAN) for modality-specific feature learning with a transformer-based dual-path cross-attention (TDCA) network for complementary fusion. MDGAN comprises dual streams for HSI and SAR data. The former captures spectral-spatial details by multi-scale feature extraction and attention, while the latter employs orientation-aware sampling to capture geometry- and orientation-sensitive backscattering signals. A residual separable U-Net discriminator incorporating depthwise separable channel attention residual blocks is considered to strengthen feature discrimination. TDCA balances RS modality contributions by introducing a novel gated bidirectional dynamic cross-attention mechanism. Experiments demonstrate that GenMSLnet outperforms state-of-the-art techniques quantitatively and qualitatively. Codes will be available at https://github.com/zhanghongzhan-123/GenMSLnet.
Accurate monitoring of soil erosion over large areas with complex terrain has been facing challenges, which primarily stem from an insufficient capacity of existing satellite remote sensing techniques for monitoring multidimensional geomorphic change in complex terrains. In this study, a framework combining synthetic aperture radar (SAR) and optical satellite imagery was proposed to overcome the above challenges. A multi-modal high-resolution satellite image registration method was proposed to mitigate the geometric distortions of high-resolution optical images induced by terrain undulations within small areas in the complex terrain. The registered optical images were employed to derive horizontal geomorphic changes using the optical pixel offset tracking method, while the surface displacement projected onto the radar line-of-sight (LOS) was retrieved using multi-temporal interferometric SAR (MT-InSAR). The quantification of soil erosion / deposition was then achieved through integrating horizontal geomorphic changes with InSAR-derived LOS deformation using a weighted least-squares inversion, while the corresponding uncertainty was quantified using the covariance matrix. The proposed image registration method was evaluated using registration errors and deformation uncertainty on stable areas, while the detected soil erosion and deposition were verified based on erosion pin measurements, laser scanning, and manually digitized terrain feature line change directions. Results showed that the proposed registration method achieved a sub-pixel registration accuracy, with an RMSE < 5 cm. The derived erosion and deposition were generally comparable with the results obtained by laser scanning and erosion pin monitoring (R2>0.6, p<0.01, NSE>0.5), with an RMSE of 3.28 cm according to erosion pin measurements. Additionally, the migration of terrain feature lines was found to be generally opposite to the direction of surface horizontal deformation, particularly pronounced at gully heads. This further demonstrated the feasibility of our monitoring results. Overall, this study provided a promising framework for soil erosion monitoring over complex terrain areas based on satellite remote sensing, while further work is still required to improve the accuracy of the method, particularly to improve the capacity of InSAR for monitoring large and rapid changes.
In this paper, to comprehensively address under-water degradation challenges covering color deviation, blurring, low contrast, etc., the color representation consistency and artificial potential field contrastive learning-guided unsupervised underwater image enhancement network with self-organizing frequency selection (termed as CRCLSF) is devised. Main contributions are as follows. 1) To overcome inherent enhancement imbalance between color-illumination and texture-structure attributes, by combining self-organizing frequency dynamic selection (SFDS) module with multiple learnable parallel filters and input-dependent dynamic weights, the unsupervised style and content disentanglement enhancement framework is established. 2) By introducing singularity-controllable semantic-type regulatory factor and relative position-aware Sine-type gain coefficient into attractive and repulsive fields, the multi-domain-guided artificial potential field contrastive learning (MAPFCL) module is proposed, thereby facilitating the enhanced images being closer to positive samples and farther from negative ones. 3) By leveraging polygonal convex hull technique to characterize scattered point distributions derived from the enhanced and distortion-free images, the color distribution representation learning (CDRL) module is developed, such that the minor difference of color distribution within Lab space can be measured, and thereby accelerating convergence performance. 4) The entire CR-CLSF framework is ultimately constructed in a modular manner by combining SFDS, MAPFCL and CDRL modules. Comprehensively quantitative and qualitative experiments demonstrate that the devised CRCLSF scheme outperforms state-of-the-art underwater image enhancement approaches.
The backscattering intensity of an ice crystal is typically stronger than near-backscattering intensity due to the coherent backscatter enhancement (CBE) in single scattering, but the effect of the CBE-induced backscattering peak has not been accounted for in forward radiative transfer models of variational LiDAR retrieval algorithms. In this study, the scattering properties of ice crystals are calculated using the improved geometric optics method (IGOM), where the CBE effect is quantified using a paired ray-tracing technique. Then, the effect of CBE on the near-backscattering phase function of ice crystals is applied to a fast-forward model [fast LiDAR and radar multiple-scattering model (FMM)], and the results of the revised model are validated using a Monte Carlo radiative transfer model. Subsequently, a variational method based on the revised fast-forward model is used to retrieve the extinction coefficient of cirrus clouds from Cloud–Aerosol LiDAR and Infrared Pathfinder Satellite Observations (CALIPSO), when the off-nadir viewing angle is 3° to avoid specular reflection from horizontally oriented ice crystals. The results show that the CBE-induced backscattering peak has remarkable impact on the retrieval of extinction coefficient profile, and the consideration of CBE improves the consistency between simulated LiDAR backscatter and collocated passive and active satellite observations (8% reduction in the RMSE for transparent cirrus).
Remote sensing scene graph generation (RS-SGG) aims to advance remote sensing image interpretation from primitive entity recognition to high-level holistic scene understanding. Due to the large spatial coverage of remote sensing images, objects are often organized into multiple functional subscenes, while predicate semantics are highly dependent on the spatial configurations of the subject and object. However, existing scene graph generation (SGG) methods typically perform unconstrained global interactions over the entire panoramic scene, which introduces substantial noise in multi-subscene scenarios, and they rarely exploit the relative spatial structure underlying subject–object relations. To address these issues, we propose the Spatially Anisotropic Reasoning Network (SARNet) for remote sensing scene graph generation. First, we design an Anisotropic Elliptical Influence Propagation (AEIP) module that models an adaptive spatial influence region for each object using anisotropic Gaussian ellipses derived from oriented bounding boxes, focusing on instance-centered interactions while suppressing noisy long-range relations. Second, we introduce a Relative Spatial Configuration-guided Attention (RSCA) module that incorporates the relative geometric configuration between each relation and its corresponding subject–object pair, and real-locates channel-wise attention to enhance spatially discriminative features for predicate prediction. Extensive experiments on the STAR and AUG benchmarks demonstrate that the proposed method achieves state-of-the-art performances on RS-SGG tasks, highlighting the importance of spatially structured reasoning that jointly captures object-centered functional context and relative spatial configurations in large-area remote sensing images. The code is publicly available at https://github.com/Bamboo0216/SARNet.
Hyperspectral remote sensing imagery provides dense spectral measurements that support material identification and fine-grained classification, but it is expensive to acquire and often limited in spatial resolution. In contrast, RGB imagery is low-cost and easy to capture with rich spatial details, yet its few channels cannot faithfully represent fine spectral signatures. Spectral super-resolution (SSR), which reconstructs hyperspectral images from low-spectral observations, is therefore of practical importance but remains highly ill-posed: multiple high-dimensional spectra may correspond to the same RGB measurement, and narrowband absorption features or abrupt local spectral variations may not be distinguishable in RGB space, causing systematic bias in regions with similar appearance but different materials. Moreover, many existing SSR models cannot adapt their update behavior to input-dependent spectral dynamics, leading to over-smoothing. To address these issues, we propose PLS2-Net, a Physics-Guided Liquid Spectral-Sequence Coarse-to-Fine Network for remote sensing SSR. PLS2-Net first produces a stable and interpretable coarse spectral skeleton using a Multi-scale Interpretable Coarse Spectral Generation (MICSG) module. It then performs input-driven bidirectional spectral-sequence modeling with a Bidirectional Liquid Spectral Dynamics (BiLSD) module, improving the representation of long-range cross-band dependencies and directional spectral evolution. Finally, an SRF-based Physical Consistency (SPC) module together with a Null-space Detail Refiner (NDR) enforces forward consistency and compensates high-frequency details in unobservable directions, enhancing both physical fidelity and detail restoration. Extensive experiments on PaviaU, Chikusei, and Indian Pines demonstrate that PLS2-Net achieves consistently superior reconstruction accuracy over state-of-the-art methods, particularly in spectral-shape preservation and error suppression.
Single-frame infrared small target detection remains challenging due to extremely weak target responses, small spatial extent, and severe background clutter. Existing deep models mainly emphasize target enhancement, but weak targets partially obscured by clutter may still produce structurally incomplete responses in infrared scenes. To address these issues, we propose BCR-Net, a body-boundary collaborative network with adaptive residual compensation for single-frame infrared small target detection. Specifically, BCR-Net adopts a shared encoder and a dual-decoder backbone to model target body and boundary features in parallel. A body-boundary enhancement module is introduced to promote cross-branch interaction, enabling the decoder to preserve target structure while suppressing clutter interference. In addition, we design an adaptive residual compensation stream to extract complementary residual cues from shallow and intermediate features. These compensation features are then adaptively injected into the main stream through a gated adaptive residual fusion module. Furthermore, a dynamic residual-guided supervision strategy is developed to progressively guide the compensation stream from full-target learning to residual-oriented refinement. Extensive experiments on three publicly available and widely used benchmarks, namely IRSTD-1K, NUAA-SIRST, and NUDT-SIRST, demonstrate that the proposed BCR-Net achieves competitive performance against state-of-the-art methods while maintaining a favorable efficiency-performance trade-off with 4.15M parameters and 73.3 FPS. Additional evaluations on SIRST-V2 and WideIRSTD-Full demonstrate the robustness of BCR-Net across diverse and challenging IRSTD scenarios.
Few-Shot Object Detection (FSOD) in remote sensing imagery faces critical challenges stemming from extreme data scarcity, specifically inadequate feature coverage, severe class imbalance, and pervasive incomplete annotations. To address these interconnected issues, this paper proposes a unified FSOD framework based on a Prototype-Conditioned GAN. The framework integrates three components to enhance robustness from data, feature, and label perspectives. First, Dynamic Augmented Balanced Sampling is introduced to mitigate overfitting by applying diverse, adaptive augmentations to oversampled novel-class instances, improving both numerical balance and visual diversity. Second, to alleviate feature scarcity, a Prototype-Conditioned GAN (P-GAN) is designed to synthesize Region-of-Interest features. Guided by prototype vectors and a cross-attention mechanism, P-GAN helps expand the feature space and alleviate classifier bias toward base classes. Third, to tackle the false negatives caused by missing annotations, a Dynamic Prototype-Aware Label Corrector exploits a teacher-student architecture and prototype similarity to adaptively recalibrate labels and loss weights. Experiments on the DIOR and NWPU VHR-10.v2 benchmarks show that the proposed approach improves detection performance over the compared methods across multiple shot settings by mitigating classifier bias and refining feature representations.
While deep learned hyperspectral image (HSI) compression has achieved remarkable progress, existing methods typically encode latent representations indiscriminately. This entangled paradigm fails to separate global structural priors from local spectral details, thereby bottlenecking the overall spectral fidelity. To address this issue, we propose a Coarse-to-Fine Spectral Prediction Network (CFSP-Net) that strictly decouples HSI reconstruction into two progressive stages: structural estimation and spatial-spectral refinement. Specifically, a lightweight multi-scale pooling network first extracts low-frequency priors to provide stable global semantic guidance. Subsequently, a full-resolution fine prediction network progressively restores high-frequency local textures and complex spectral correlations via a novel spatial-spectral refinement block, which integrates window-based spatial aggregation and hierarchical cumulative modeling. Furthermore, a fine-grained residual quantization strategy and a band-group optimization scheme are introduced to synergistically stabilize high-dimensional training and refine bit allocation. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness of CFSP-Net, especially on structurally complex HSI scenes and low-to-medium bitrate conditions. Notably, on the Chikusei dataset, the proposed method achieves a PSNR of 52.69 dB and a Spectral Angle Mapper (SAM) error of 0.0300 at 0.283 bpppb, improving the PSNR by approximately 0.45 dB over the strongest competing method at a comparable bitrate while maintaining superior spectral fidelity.
Infrared video small object detection remains a challenging problem due to the extremely weak target appearance, low signal-to-noise ratio, cluttered thermal backgrounds, and frequent temporal inconsistency across frames. Most existing detectors follow a frame-wise paradigm, where each frame is processed independently or temporal cues are introduced through additional aggregation modules. Such designs often suffer from insufficient motion perception, background-induced false alarms, and increased computational complexity. In this paper, we propose Cube-IRVSOD, a compact yet effective cubeto- trajectory detection framework for infrared video small object detection. The key idea is to reformulate consecutive infrared frames as a spatiotemporal cube, enabling short-term motion cues to be encoded as intrinsic input structures rather than external temporal dependencies. Specifically, a Cube Data Stream Generator converts sampled consecutive frames into grayscale intensity maps and stacks them into a cube representation, allowing the detector to jointly perceive spatial appearance and temporal evolution. To further suppress background-dominated pseudo motion, we introduce an Event Flow Encoder that derives sparse event maps from inter-frame brightness changes and employs SNN-based temporal encoding to generate motionaware priors. These priors guide sparse feature enhancement over multi-level spatial representations, emphasizing moving target responses while reducing stationary clutter interference. In addition, a trajectory-to-frame inference scheme with confidence voting aggregates predictions from overlapping cube streams, improving temporal stability and localization reliability. Extensive experiments demonstrate that Cube-IRVSOD achieves superior detection accuracy and a favorable accuracy-efficiency trade-off under challenging scenarios.
Semantic segmentation of high resolution remote sensing images is a fundamental task in Earth observation, yet its performance is often constrained by the expensive and time-consuming process of pixel-level annotation. While existing semi-supervised learning methods can mitigate this label scarcity, they often overlook the inherent spectral-spatial duality of remote sensing data. This oversight leads to the entanglement of het-erogeneous features and the progressive accumulation of pseudo-label noise, ultimately limiting segmentation accuracy. To address this, we propose a Decoupled Dual-teacher Collaborative Semi-supervised Framework (DDCSF), which generates high-quality supervisory signals from heterogeneous modalities via novel decoupling and alignment mechanisms. DDCSF employs separate spectral and spatial teacher networks to achieve decoupled learning of texture features and geometric structures, respectively. To align these heterogeneous features, we design a Frequency-aware Cross-modal Fusion Module (FCFM). This module leverages the high-frequency components of spatial features to refine target boundaries and the low-frequency components to correct regional semantics, producing collaborative features with both precise details and semantic consistency. Furthermore, a Pixel Confidence Voting Module (PCVM) is introduced to quantify prediction uncertainty from the dual teachers and fusion module, selecting high-confidence pseudo labels to guide student model training. The student model’s ability to represent multimodal inputs is further enhanced by an Adaptive Spectral–Spatial Image Fusion Module (ASSIFM). Experimental results on the ISPRS Potsdam and GID benchmark datasets demonstrate that DDCSF consistently outperforms state-of-the-art methods at various label rates, validating its effectiveness.
The rational function model (RFM) is crucial for the geometric processing of remote sensing satellite imagery. However, estimating the unknown rational polynomial coefficients (RPCs) remains an ill-posed and over-parameterized problem due to severe multicollinearity. Existing regularization methods predominantly rely on the assumption of independent and identically distributed (i.i.d.) observational errors, treating ground control points (GCPs) as independent measurements while neglecting their intrinsic spatial topological relationships. In practice, imaging distortions exhibit strong spatial correlation, indicating that neighboring GCPs share similar geometric deformation patterns. Neglecting this topological prior may limit the accuracy of RFM estimation, especially when the GCPs are insufficient. To address this limitation, this paper proposes a graph-based RFM (G-RFM) estimation method that explicitly encodes spatial dependencies among GCPs within a Bayesian maximum a posteriori (MAP) framework. Specifically, the observation residuals are modeled as a Gaussian Markov random field governed by a graph Laplacian matrix, which constrains estimation errors between spatially adjacent GCPs. To ensure computational efficiency for large-scale applications, an accelerated alternating direction method of multipliers (ADMM) algorithm is developed. Comprehensive experiments demonstrate that the proposed method achieves competitive and often superior performance compared with existing methods in both terrain-dependent and terrain-independent cases.
Building climate-resilient adaptation in rapidly urbanizing African regions requires a deeper understanding of drought dynamics across interconnected urban and rural systems. This study develops a remote sensing–based framework to quantify the relative dominance of urban water–energy demand stress and rural agricultural drought across Africa. The Urban–Rural Drought Stress Ratio (URDSR) is proposed, which is a novel Principal Component Analysis (PCA)–based multi-sensor drought diagnostic that separates urban anthropogenic water–energy demand stress from rural hydroclimatic agricultural drought within a unified framework. Multi-sensor observations from Sentinel-1/2, GLDAS, and VIIRS are harmonized to monthly 0.05° grids. PCA is applied to construct two sectoral indices: an Urban Water Demand Index (Uw) representing infrastructure- and temperature-driven urban water pressure, and the Rural Stress Index (Rs) representing agricultural drought derived from vegetation and hydrological deficits. Their contrast forms URDSR, enabling dynamic assessment of the balance between urban and rural drought stress. The framework is evaluated against conventional drought indicators, including the Standardized Precipitation Index (SPI) and Standardized Precipitation Evapotranspiration Index (SPEI) derived from CHIRPS data, with detailed analysis for Cairo, Lagos, Nairobi, Khartoum, Addis Ababa, and Johannesburg. Validation against benchmark indices (SPI, SPEI, VHI, NDVI, VCI, and TVDI) shows strong correlations (r = 0.70–0.85, p < 0.01), revealing a south–north transition from rural-dominated to increasingly urban drought regimes.
