As severe convective weather exerts growing influence on public safety, enhancing forecast accuracy has become critically important. However, the predictive capability remains limited due to insufficient observational coverageenlrg in certain regions or variables, as well as the inadequate representation of the fine-scale physical processes responsible for local convective development. In response to these challenges, this study proposes a physically embedded neural network based on heterogeneous meteorological data, which utilizes satellite multispectral images and atmospheric temperature and humidity profile synergistically retrieved from space-based and ground-based infrared spectral observations, to forecast local convective initiation (CI) within a 6-hour lead time. The core innovation of this study lies in the development of a physically consistent model that explicitly embeds the convective available potential energy equation into the network architecture. By embedding physical information, the model enables the atmospheric thermodynamic feature extraction module to generate physically consistent feature tensors, thereby enhancing the representation of key convective processes. We trained the network using the pretraining and fine-tuning approach, then validated its effectiveness with reanalysis and actual observational data. The results demonstrate that incorporating the retrieved atmospheric profile data leads to a 40% improvement in the 6-hour average critical success index (CSI), increasing from 0.44 to 0.62 relative to forecasts without atmospheric profile input. Furthermore, in validation experiments using reanalysis data and radar observations, the proposed atmospheric profile feature extraction module consistently improves the model’s average forecast CSI by more than 29% compared to models utilizing purely data-driven profile extraction modules.

Mid-wave infrared (MWIR) acousto-optic tunable filter (AOTF) spectral imaging systems suffer from inherent conflicting constraints among optical throughput, AOTF angular acceptance, cold-stop matching, and broadband zero-order leakage suppression. Conventional designs treat these requirements independently, leading to severe signal-to-noise ratio (SNR) degradation and suboptimal system performance. This paper proposes a constraint-driven optomechanical co-design methodology that unifies these coupled parameters into an analytical constraint chain, deriving a crosstalk-free safe design boundary for key system variables. An integrated optical architecture is developed, consisting of an off-axis three-mirror afocal telescope, a TeO₂ AOTF module, rear imaging optics, and a multi-stage conjugate physical stop scheme. Over the 3.7–4.8 μm band, the optimized system achieves an RMS wavefront error better than λ/20 at 4 μm, full-field MTF above 0.5 at the Nyquist frequency, 99.5% cold-stop matching efficiency, and distortion below 0.25%. Full-link simulations show the proposed stop configuration suppresses out-of-field point source transmittance (PST) to 10⁻⁶–10⁻⁷, a 5–6 order-of-magnitude improvement over the baseline. A visible-band TeO₂ AOTF prototype satisfying the same optical constraint framework further supports the feasibility of the proposed zero-order leakage suppression methodology.

Existing retrieval methods based on infrared hyperspectral data still face limitations. On the one hand, physics-based iterative algorithms require multiple radiative transfer calculations, resulting in high computational costs. On the other hand, neural network approaches lack physical constraints, making it difficult to ensure physical consistency and to reasonably evaluate retrieval uncertainties. To address these limitations, a physics-informed network, DeepOE, is introduced, which integrates deep learning with the widely used Optimal Estimation Method (OEM). Synergistic observations from the ground-based Atmospheric Emitted Radiance Interferometer and space-based Geostationary Interferometric Infrared Sounder are employed to conduct atmospheric temperature profile retrieval experiments. DeepOE is capable of simultaneously providing the retrieval results and their associated uncertainties within a very short computation time. With radiosonde observations serving as the reference truth, the results show that for temperature profile retrievals in the 0–12 km range, DeepOE achieves a layer-averaged RMSE of 1.389 K, representing reductions of 0.03 K and 0.05 K compared with OEM and the purely data-driven ResNet model, respectively.

Mid-wave infrared (MWIR) acousto-optic tunable filter (AOTF) spectrometers face two critical challenges in structural-thermal-optical performance (STOP) analysis: internal self-heating during operation and the strong temperature dependence of acousto-optic parameters, which cause ray-tracing deviations. This study established a comprehensive multiphysics coupling framework. The AOTF power distribution under frequency sweeping was experimentally characterized and applied as thermal boundary conditions, while the temperature-dependent refractive index and acoustic velocity were incorporated into the ray-tracing model, enabling integrated opto-mechanical-thermal simulation under realistic operating conditions., Comparative experiments demonstrated that, after excluding alignment and machining errors, the measured MTF degradation under the AOTF sweep operation with 0 °C cooling (0.039 tangential, 0.031 sagittal) agreed closely with the simulated values (0.031 and 0.023), with an absolute error of only 0.008. This close agreement validated the proposed method. Moreover, the minimal MTF degradation observed under this condition validated the thermal control scheme and structural design. The established model is generalizable to AOTF spectrometers operating in other spectral bands.

Current large-scale multimodal models are primarily trained on static visible spectrum (RGB) data. However, in practical applications such as remote sensing, data often arrives in streams and undergoes severe cross-modal domain shifts, particularly the transition from visible to infrared (IR) imagery. This dual-incremental scenario (encompassing both task and modality increments) poses significant challenges for continual learning: the model must overcome feature space disorder caused by modality gaps while preventing catastrophic forgetting and preserving the zero-shot capability of the pre-trained backbone. To address these issues, this paper proposes a Domain-Adaptive Mixture-of-Experts (DA-MoE) framework. First, to mitigate the feature distribution discrepancy between RGB and IR modalities, we introduce a Modality-Aware Nested Expert structure. Instead of simply appending parallel experts, this design embeds lightweight adapters within the computational path of frozen base experts, enforcing semantic alignment of new modality features with the established base feature space. Second, to achieve robust modality switching and out-of-distribution (OOD) detection during inference, we propose a Confidence-Calibrated Distribution Selector. This mechanism utilizes self-supervised reconstruction error as a physical prior to transform domain discrepancies into a calibrated probability distribution. By employing continuous soft gating, it dynamically regulates the activation intensity of expert pathways, thereby preventing the instability associated with traditional hard routing at modality boundaries. Experimental results on sequential classification benchmarks containing both visible and infrared imagery demonstrate that the proposed method not only significantly reduces catastrophic forgetting but also effectively enhances the model’s adaptability and zero-shot generalization stability in the infrared domain.

The infrared multispectral imaging system based on an acousto-optic tunable filter (AOTF) possesses strong capabilities in suppressing complex backgrounds and anti-interferences for the detection of aerial targets. It is worth noting that the wavelengths and bandwidths of the detection bands have a significant impact on detection capability. However, the existing spectral parameters selection methods have poor generalization, rely on manual reasoning strategies and repeated training with a large amount of data, and cannot achieve the optimum for the spectral parameters. To address these problems, we propose the wavelength and bandwidth selection models based on multiagent reinforcement deep learning (WBMARL). This is the first attempt to apply the multiagent reinforcement learning to spectral parameters selection tasks, which provides a new paradigm. Based on the assumption that the combinatorial optimization problem of spectral parameters selection is regarded as a Markov game process, the WBMARL, which consists of a deep feature extraction module and two agents, is designed. Through cooperation among the agents, the Nash equilibrium point is reached, and the optimal spectral parameters selection policy is obtained. The experimental results on the infrared multispectral images captured by an AOTF system indicate that the spectral parameters selected by WBMARL are more conducive to the detection of targets in complex backgrounds and anti-interference.

Deep-space infrared target recognition is significantly challenged by similar interference sources and complex backgrounds, which degrade detection accuracy (Acc) and robustness. While detection has shifted from single-detector to multidetector coordinated systems, current algorithms fail to consider both the dependencies between different features and the complementarity of the unique characteristics inherent to each modality, limiting fusion quality. To address this, we propose a secondary attention-based cross-modal fusion (SA-CMF) framework. The first attention stage applies modality-specific attention to enhance the discriminative ability of individual features, fully exploiting each modality's unique characteristics, while the second stage employs cross-attention to capture interdependencies among modal features. Additionally, a dynamic mutual information (MI) adjustment strategy is introduced to suppress redundant information and mitigate cross-modal distribution discrepancies. The experimental results demonstrate that SA-CMF achieves 94.7% Acc, high F1 scores, and robust performance under diverse deep-space conditions. These results validate the framework's effectiveness in improving both feature quality and target recognition, highlighting its potential for complex infrared remote sensing applications.

Polarization multiplexing can increase the throughput of acousto-optic tunable filter (AOTF) spectral imagers, but wavelength-dependent angular separation between ordinary and extraordinary rays causes polarization chromatic aberration at the detector. We propose a passive dispersion-matched compensation method that suppresses this error before image formation. A quarter-wave plate converts the two linear polarization channels into opposite circular states, and a liquid-crystal polarization grating provides opposite Pancharatnam-Berry diffraction for angular redirection. In a 3.7-4.8 μm MWIR AOTF imager, a fixed 1200-μm-period LCPG reduces the relative o-/e-ray misregistration from 5.5 pixels to below 0.9 pixels while providing a 45-62% measured signal enhancement over single-polarization acquisition.

The Acousto-Optic Tunable Filter (AOTF) imaging spectrometer, as an advanced spectral imaging technology, has gained widespread global adoption due to its advantages including stare imaging capability, flexible band switching, and absence of moving components. However, during AOTF installation and alignment, shifts in the central wavelength of diffracted light can be induced by alignment errors, consequently reducing diffraction efficiency, while the imaging quality is simultaneously significantly impacted by such errors. In conventional design processes, tolerance allocation has typically been performed separately, where an empirical tolerance distribution scheme would be developed to independently verify whether alignment errors meet design requirements for both imaging and spectral performance. Such approaches often result in excessively stringent tolerances with correspondingly high manufacturing costs. Through the establishment of a comprehensive model and simulation of the AOTF imaging spectrometer, sensitivity analyses of various alignment errors were conducted across both spectral and spatial dimensions. High-sensitivity and low-sensitivity parameter were identified through this analysis, enabling the development of assembly tolerances that simultaneously satisfy both requirements. The proposed tolerance allocation scheme was validated by Monte Carlo simulations, with results demonstrating that optical design outcomes meet specified performance metrics in both spectral and spatial dimensions under the allocated tolerances. Compared with conventional tolerance allocation approaches, tolerance requirements were significantly relaxed while production costs were reduced by the proposed method. Ultimately, an integrated tolerance design methodology was developed that holistically considers both the AOTF device itself and its optical-mechanical system, effectively addressing the historical issue of redundancy in AOTF spectrometer system tolerance design.

Conventional 3D measurement techniques face challenges when applied to large-scale structures, including limited depth of field, narrow field of view, poor real-time performance, and low efficiency. Omnidirectional vision sensor enables simultaneous multi-directional measurements and offer significant advantages for deployable structures. To achieve clear imaging in both near-field and far-field regions, this paper proposes a composite omnidirectional vision sensor consisting of two sub-sensors. Each sub-sensor utilizes a hexagonal pyramid mirror to reflect images captured by a single camera and reconstructs 3D data based on depth estimation. The mechanical structure and field of view are analyzed. Camera intrinsic parameters are calibrated using a planar circular calibration target, while relative poses between virtual cameras are determined by the geometry of the mirrors. A global coordinate transformation is used to merge data from both sub-sensors. Experimental results demonstrate that the proposed sensor achieves a measurement error of less than 0.7 mm within 50 meters. Dynamic reconstruction experiments on vibrating targets confirm the feasibility and effectiveness of the method.

Topographic variation introduces substantial uncertainty into hyperspectral surface reflectance (SR) retrieval. This study proposes a combined correction and uncertainty quantification framework that integrates ISOFIT-based atmospheric correction (ATCOR) with a semi-empirical modified Minnaert (MM) topographic correction, while further introducing a novel uncertainty evaluation approach. The framework combines optimal estimation (OE) with a guide to the expression of uncertainty in measurement (GUM)-based uncertainty propagation to explicitly account for terrain parameter uncertainties, enabling efficient and physically consistent characterization of SR uncertainty. Application to the environmental mapping and analysis program (EnMAP) hyperspectral data demonstrates that the method achieves comparable correction accuracy to established algorithms, while providing per-pixel, per-band uncertainty estimates. Validation is performed through cross-sensor comparison with SI-traceable Landsat 8 SR products across diverse land cover types and terrain conditions. The results confirm the reliability of the proposed uncertainty estimates and highlight their value for assessing cross-sensor reflectance consistency. Analysis further reveals that reflectance uncertainty generally follows a Gaussian distribution and that slope uncertainty is the dominant driver of reflectance uncertainty. Overall, this work delivers a practical framework for uncertainty-aware SR retrieval in mountainous regions and establishes a pathway toward uncertainty-based interoperability of hyperspectral products across multiple sensors.

Conventional 3-D measurement techniques face challenges when applied to large-scale structures, including limited depth of field, narrow field of view, poor real-time performance, and low efficiency. Omnidirectional vision sensor enables simultaneous multidirectional measurements and offer significant advantages for deployable structures. To achieve clear imaging in both near-field and far-field regions, this article proposes a composite omnidirectional vision sensor consisting of two subsensors. Each subsensor utilizes a hexagonal pyramid mirror to reflect images captured by a single camera and reconstructs 3-D data based on depth estimation. The mechanical structure and field of view are analyzed. Camera intrinsic parameters are calibrated using a planar circular calibration target, while relative poses between virtual cameras are determined by the geometry of the mirrors. A global coordinate transformation is used to merge data from both subsensors. Experimental results demonstrate that the proposed sensor achieves a measurement error of less than 0.7 mm within 50 m. Dynamic reconstruction experiments on vibrating targets confirm the feasibility and effectiveness of the method.

The increasing availability of hyperspectral image (HSI) data has motivated unsupervised domain adaptation (UDA) for multi-scene scenarios. Nevertheless, fixed network architectures adopted by most existing methods limit their transferability when uncertainty is amplified by scene diversity and leads to complex distribution shifts. To address these problems, we propose a heterogeneity-aware dynamic state modeling network (HDSM) for UDA in HSI. In HDSM, a heterogeneity-aware uncertainty estimator (HAUE) is designed under a Bayesian-inspired inference perspective to estimate heterogeneity-induced uncertainty and prediction variability across scenes. Based on the estimated reliability, a scene-aware dynamic Mamba (SADM) module further incorporates uncertainty into the internal state transition process of the state space model, allowing the state modeling process to adapt dynamically to different scenes. Additionally, a multi-branch loss is designed to effectively exploit unlabeled target data. Extensive experiments on two publicly available datasets and one self-constructed multi-scene dataset show that HDSM surpasses state-of-the-art methods, improving classification accuracy by 3.27%, 2.99%, and 11.01%, particularly excelling on complex multi-scene domains.

UV–visible limb-scattered observations provide key information on the vertical distributions of trace gases in the middle atmosphere. As Earth-observing satellites progressively acquire large off-nadir limb-imaging capability, imaging spectroscopy is expected to further enhance the spatiotemporal information content of limb observations. However, existing models typically focus on single line-of-sight spectral radiance calculations and do not explicitly incorporate the instrument imaging chain, which hinders end-to-end assessment and broader application of limb-imaging spectrometers. In this study, starting from atmospheric composition and including selected instrument parameters, we implement UV–visible Earth limb-imaging spectral simulations and perform validation with quantitative error analysis.The proposed approach builds a forward-simulation framework for UV–visible limb-imaging spectroscopy based on the SASKTRAN-HR radiative transfer engine, generating physics-based limb images and hyperspectral three-dimensional data cubes over 300–800 nm and 6–97 km. A three-dimensional atmospheric scene is constructed using CAMS reanalysis data (deriving air number density from pressure and temperature and incorporating ozone and sulfate aerosols, among others), while the upper-atmospheric background state is extended using the CIRA-86 model. The instrument imaging chain is further coupled, including field of view (FOV), spectral response function (SRF), and point spread function (PSF), to represent pixel-level spectral–spatial coupling effects.Validation is conducted in the spectral domain. Simulated radiances are convolved to the effective spectral resolution of OSIRIS, and a height-by-wavelength evaluation is performed against a single OSIRIS limb scan (scan No. 54300035). Over 300–800 nm and 6–97 km, the simulations exhibit an overall systematic underestimation, with a mean absolute relative error (MARE) of 31.5% (median 25.1%). The dominant error source is attributed to discrepancies between the constructed atmospheric scene and the actual atmospheric state. Within the 20–55 km altitude range commonly used for trace-gas profile retrievals, the MARE is 10.9% (median 8.5%). Errors increase substantially above 80 km (MARE = 63.1%), likely related to stray light and reduced signal-to-noise ratio due to weak scattering signals. In addition, the simulated results are converted into pseudo-color imagery using the CIE 1931 color matching functions to enable a qualitative consistency check of limb radiance gradients and chromaticity variations.This imaging-spectroscopy simulation framework provides a testbed for limb-imaging instrument design, development and evaluation of trace-gas retrieval algorithms, and satellite validation activities for current and future limb-imaging missions (e.g., ALTIUS).

Online abnormal event detection of non-cooperative space targets based on space-based optical observations is an urgent requirement for ensuring space security, particularly for sudden state changes such as attitude and orbit maneuver events. However, existing methods are dominated by ground-based monitoring and offline analysis, while space-based online methods are still under development and struggle to handle tasks under changing observation scales. To address this issue, we propose a Knowledge-Guided Factor Graph Neural Network (KG-FGNN). A heterogeneous graph is constructed to model high-order relationships among target properties, behaviors, observation conditions, and image features. Then, a dual-confidence dynamic weighting mechanism is designed to adaptively adjust the confidence of nodes and edges according to the observation scale. Finally, bidirectional temporal residuals are introduced to explicitly decouple sudden events from normal evolution using forward and backward prediction errors in an image sequence. Experimental results on simulated datasets demonstrate that KG-FGNN achieves higher detection accuracy for events compared to existing methods, while maintaining stable performance under cross-scale observation scenarios. This study provides an effective technical approach for online event detection in space-based optical image sequences and contributes to enhancing space situational awareness.