In urban environments, targets are often located in non-line-of-sight (NLOS) regions of the radar, which makes conventional direction of arrival (DOA) estimation methods ineffective. Reconfigurable intelligent surface (RIS), with its capability of regulating electromagnetic wave propagation, can establish a virtual line of sight (LOS) path between radar and targets, offering a viable solution for NLOS target detection. However, when the signal is reflected by the RIS and received by the radar, the echo becomes a superposition of the signals reflected by each element of the RIS, in which the phase information of the target steering vector is jointly accumulated, making it difficult to extract angle information directly. To solve this problem, an RIS-aided monostatic radar DOA estimation method based on steering vector decoupling is proposed. First, RIS phase shift vectors are designed to perform spatial scanning, thereby acquiring echo data from different beam directions. Matrix operations are then employed to decouple the mixing effect introduced by the RIS and to recover the outer-product matrix of the target steering vector. Subsequently, the Root-MUSIC algorithm is applied to achieve high resolution angle estimation. Simulation results demonstrate that the proposed method significantly improves the angle estimation accuracy in typical NLOS scenarios.

To address the stringent autonomous navigation requirements in cislunar space, this study proposes a hybrid constellation navigation framework utilizing inter-satellite links. It integrates near-rectilinear halo orbits (NRHOs), distant retrograde orbits (DROs), and BeiDou system (BDS) geostationary orbit/inclined geosynchronous orbit satellites into a dual-layer network, employing Ka-band dual one-way ranging and an extended Kalman filter for real-time state estimation. Monte Carlo simulations under high-fidelity dynamics reveal that: Without BDS augmentation, NRHO and DRO satellites achieve 100 m-level 3D accuracy with 5 d arcs, improving to 10 m level with 10 d arcs, and reaching sub-meter (DRO) and meter-level (NRHO) with 15 d arcs. Symmetric NRHO configurations show poor short-arc performance due to geometric correlation. The hybrid constellation with BDS augmentation significantly outperforms standalone lunar constellations, achieving 20-30 m 3D accuracy with just 2 d arcs. The key contribution is the synergistic integration of heterogeneous orbits, leveraging NRHO's lunar proximity, DRO's wide coverage, and BDS's stable Earth reference to enhance autonomy, accuracy, and robustness. This framework reduces ground dependency and provides a scalable solution for future cislunar exploration.

The rapid expansion of large-scale Low Earth Orbit (LEO) satellite constellations offers new opportunities for space-based monitoring of Global Navigation Satellite Systems (GNSS). However, leveraging these constellations introduces challenges such as data redundancy and onboard resource constraints, necessitating efficient methods for selecting symmetric satellite subsets. This paper proposes a selection and reconfiguration framework based on 2D Necklace Flower Constellation (2DNFC) theory. By formulating the satellite selection problem as a number-theoretic “Necklace Problem” and solving linear congruence equations, the framework systematically identifies optimal subsets from large-scale parent constellations while preserving spatial geometric symmetry. Furthermore, leveraging the concept of “equivalent necklaces,” we propose a rapid reconfiguration mechanism that maintains monitoring performance without degradation in the event of satellite failures. Using the CENTISPACE LEO constellation as a candidate pool and full-arc monitoring of BeiDou-3 (BDS-3) as the objective, we compare the proposed method with conventional selection strategies. Results show that the 24-satellite subsets selected via the 2DNFC and double necklace methods achieve 100% four-fold coverage of the BDS-3 full arc, with significantly enhanced five-fold and six-fold coverage percentages. Moreover, in failure scenarios, subsets based on equivalent necklaces enable rapid reconfiguration without orbital maneuver and with no loss in monitoring performance. This study provides a theoretical foundation and empirical validation for constructing low-cost, efficient, and survivable GNSS space-based monitoring networks.

Low Earth orbit (LEO) constellations are increasingly deployed to provide integrated services such as communication, navigation augmentation, and remote sensing. However, traditional constellation designs-such as Walker, streets-of-coverage (SoC), and Flower constellations-rely on distinct parameter sets, making unified optimization of hybrid configurations challenging. Existing unified approaches like the continuous coefficient (C2) method suffer from high-dimensional parameter spaces, leading to suboptimal solutions and convergence issues. To address this, we propose an improved continuous coefficient (i - C2) method that reduces the number of real-valued parameters while maintaining the ability to represent both symmetric and asymmetric constellation types. Using evolutionary algorithms, we apply the i - C2 method to design several LEO constellations, including single- and dual-coverage SoC configurations, a navigation-augmentation hybrid constellation, and an integrated positioning, navigation, timing, remote sensing, and communication (PNTRC) system. Results demonstrate that the i - C2 method successfully reproduces the Iridium NEXT constellation and identifies "inclination family" structures. Compared to combined orthogonal circular and Walker constellations, the i - C2 method improves the uniformity of the average number of visible satellites by 20% and 77.92% for 100- and 150-satellite configurations, respectively. This study highlights the i - C2 method's capability to enable efficient, flexible, and high-performance LEO constellation design within a unified optimization framework, offering significant potential for more integrated and cost-effective satellite systems.

Current semisupervised synthetic aperture radar (SAR) ship detection and classification methods mainly focus on closed-set scenarios, where the labeled training set includes all target classes present in the unlabeled training data and test data. In practice, this assumption rarely holds, as unknown targets absent from the labeled training set are often encountered. To address the limitation of closed-set methods in detecting and identifying unknown targets in open-set scenarios, this article proposes an open-set semisupervised SAR ship detection and classification method via joint known and unknown targets exploitation (JoTE). First, considering both known and unknown ships share more similar general properties with each other than with clutter, we design a suspected target detection module that combines class-agnostic target property learning with a curriculum of pseudolabels. This module enables the detector to better distinguish between unlabeled targets that have not been encountered and background clutter, thereby enhancing the detection performance in complex scenes. Then, considering the prior knowledge of SAR ships, we propose a physical- and deep-feature-guided open-set classification module, which can accurately classify known targets and identify unknown targets. It continuously updates both physical and deep features using training data and assigns reliable pseudolabels to unlabeled targets based on feature similarity measurements in the training process. In addition, we introduce extreme value theory to model the decision boundaries of the known targets stored in the feature banks, thereby adaptively setting the similarity thresholds for identifying unknown targets. Experiments on the SRSDD-v1.0 dataset demonstrate that JoTE achieves satisfactory performance over existing methods.

Spiking neural networks (SNN) have emerged as a promising energy-efficient alternative to artificial neural networks (ANN), leveraging event-driven operation for efficient neuromorphic implementation. However, achieving low-latency inference without compromising accuracy remains challenging, as SNNs typically require numerous time steps for effective processing.To address this, we propose a discrete knowledge distillation method applied to both output and multiple intermediate features, enabling low-latency SNN training. To our knowledge, this is the first work introducing multi-layer knowledge distillation to SNN training. Specifically, we first train a high-performance ANN, then employ ANN-to-SNN conversion to obtain an SNN matching the ANN's performance at long time steps. We then employ a cyclical teacher-student distillation process: the initial long-time-step SNN serves as teacher to guide training of a short-time-step student SNN. Upon convergence, this student becomes the new teacher, supervising an even shorter-time-step student. This cycle progressively compresses the temporal dimension through multi-layer distillation until obtaining a compact SNN matching ANN performance with minimal time steps. In multi-layer distillation, we use both output logits and intermediate features for comprehensive knowledge transfer, but SNN's discrete nature makes direct feature distillation infeasible. To overcome this, we propose a probability distribution-based spike feature distillation method that converts spike trains into continuous probability distributions and minimizes distributional divergence between teacher and student. Experiments on public datasets demonstrate our work significantly reduces inference time steps while maintaining competitive accuracy.

Short-baseline Connected Element Interferometry (CEI) is widely used for monitoring Geostationary Earth Orbit (GEO) satellites owing to its advantages in operational concealment and cost-effectiveness. However, due to the geostationary characteristics of GEO satellites, CEI system errors are difficult to separate from satellite orbit errors. In this study, a high-precision calibration method based on co-located GNSS and CEI observations is proposed. To avoid pointing errors introduced by CEI antenna tracking, GEO satellites are selected as calibration sources. By accounting for the inter-channel time delay differences of the GNSS receiver and establishing a CEI calibration model, the CEI system errors are decoupled, and explicit estimates of the system errors are obtained. Co-located observation experiments were conducted using the QZSS J07 and BDS C03 satellites to calibrate the system errors of a CEI system located in Zhengzhou, China. The results show that the CEI system error calibration values obtained from observations of the two satellites at four frequencies are all stable at approximately −1128.0 m. Further short-term extrapolation validation indicates that the CEI system error calibration values have good internal precision and stability, providing reliable support for the subsequent correction of CEI observation data.

The automatic astrogeodetic survey method based on total stations has emerged as the second near-Earth astrogeodetic technology after digital zenith cameras, offering high automation, precision, portability, and cost-effectiveness. The method employs an automatic scheduling algorithm for optimal observation sequencing and integrates Global Navigation Satellite System (GNSS) receiver synchronization with a precision oscillator to ensure time accuracy. This enables high-precision surveys with flexible field deployment, effectively replacing manual observations. Using the improved Image Total Station Astrogeodetic Positioning and Orientation System (ITSAPOS) as the measurement platform, we examined its operating principle, focusing on accuracy and reliability. Between 2022 and 2024, 76 experiments were conducted at 24 observation sites across plains, plateaus, hills, and mountains given varying seasonal conditions. ITSAPOS, the entire system weighing no more than 15 kg, consistently outperformed the Astrogeodetic Surveying System Based on Electronic Theodolites (ASSET), achieving standard deviations approximately half those of ASSET, thereby doubling precision. Longitude standard deviations ranged from 0.002 to 0.009 s and latitude standard deviations from 0.003 '' to 0.110 '' based on multinight observations at the same points. ITSAPOS also achieved stable closure errors for the meridian triangle within shorter times, reaching 0.01 s accuracy. Longitudinal external point accuracy was high, with a maximum difference of 0.012 s, average absolute difference of 0.007 s, and root mean square error (RMSE) of 0.009 s. For latitude, the maximum difference was 0.481 '', with average absolute difference of 0.253 '' and RMSE of 0.301 ''. In comparison, ASSET showed 0.053 s and 0.532 '' for maximum differences, 0.018 s and 0.300 '' for average differences, and 0.025 s and 0.334 '' for RMSE. Overall, the automatic astrogeodetic survey method based on total station images provides autonomous measurement, portability, cost-effectiveness, and stable high precision. The method allows rapid adjustment of observation points and cycles, offering transformative advantages for remote deployments and continuous monitoring.

Accurate estimation of bloodstain aging duration is a critical challenge in forensic biology, as it underpins event timeline reconstruction and behavioral interpretation at crime scenes. Existing spectroscopic methods face limitations including biochemically unguided spectral selection, insufficient model interpretability, and limited environmental robustness. To address these issues, this study proposes a one-dimensional convolutional network, termed BloodAttention-CNN, which integrates channel and spatial attention mechanisms for aging classification of porcine bloodstains using 900–1800 cm−1 near-infrared spectra. After applying a standardized preprocessing pipeline, we compared the proposed network with conventional machine learning models. BloodAttention-CNN achieved 92.35% accuracy on the test set, substantially outperforming all competing methods. Attention visualization revealed that the model consistently prioritized regions near 1120–1130 cm−1, 1650 cm−1, and 1720 cm−1—peaks closely corresponding to hemoglobin oxidation and protein unfolding during stain maturation. Systematic perturbation experiments further confirmed that perturbing these high-attention regions caused up to a 35.29% accuracy drop, whereas perturbation of the negative control region caused only a 5.88% drop, providing perturbation-based evidence that these regions influence model decisions and are associated with biochemically relevant spectral features. Cross-species and cross-temperature validation further demonstrated robust generalization (82.35–90.12% accuracy). This framework offers a rapid, non-destructive, and forensically accountable strategy for bloodstain deposition time estimation, establishing a traceable link between spectral inputs, deep learning decisions, and underlying molecular mechanisms.

Global navigation satellite system (GNSS) integrity monitoring relies on continuous, worldwide observation of satellite signals. While ground-based monitoring networks are constrained by geographical coverage and political boundaries, low Earth orbit (LEO) constellations offer a promising space-based alternative due to their global distribution and dynamic characteristics. Current performance evaluations of satellite constellations often rely on real or simulated ephemerides. Although the geometry and probability model (GAPM) can estimate Earth coverage without such data, it cannot model the 'inverse coverage' of LEO satellites monitoring high-altitude GNSS satellites. To address this gap, we propose the LEO GAPM, which explicitly models this 'inverse coverage' by introducing the concept of relative inclination and reformulating the observation probability function. This model estimates key performance metrics, such as monitoring coverage fold and satellite geometry dilution of precision (SGDOP), using only a small set of orbital parameters. Validation experiments show that the average difference in estimated coverage fold between LEO-GAPM and orbital propagation methods is less than 0.14, and the underestimation rate of SGDOP ranges between 5% and 13%. The model was further employed to analyze two representative LEO constellations (Iridium NEXT and CENTISPACE) for BDS-3 monitoring, assessing how coverage capability varies with the number of LEO satellites. This analysis demonstrates that constellation configuration, orbital inclination, and satellite count are key determinants of monitoring performance. Additionally, the study quantifies the number of satellites required to achieve an average coverage fold of 3-10. Overall, our study provides a computationally efficient tool for preliminary design and resource allocation in space-based GNSS monitoring systems.

Global Navigation Satellite System Reflectometry (GNSS-R) has been successfully applied in fields including ground target detection and imaging. However, the low Power Flux Density (PFD) of signals reflected from the Earth's surface has long been a critical bottleneck restricting the performance improvement of target detection. The rapid development and large-scale deployment of Low Earth Orbit (LEO) navigation satellites have furnished external illuminators with superior operational performance. The optimization of the quantity and spatial distribution of these transmitters can effectively enhance the accuracy of target state estimation. This paper first analyzes the Signal-to-Noise Ratio (SNR) improvement effect of the LEO-R system from three key dimensions: the orbital altitude of LEO navigation satellites, detection range, coherent time and incoherent time. Furthermore, a moving target state estimation algorithm specifically designed for LEO-R scenarios is proposed. This algorithm extracts the bistatic range and target radial velocity of a single satellite via the Radon-Fourier Transform (RFT), establishes a path difference-based observation model, and adopts the Levenberg-Marquardt (L-M) algorithm to solve the nonlinear least squares problem for precise target position estimation. Subsequently, target velocity is resolved by integrating the position estimation results with a direct physical model. Both simulation and field experimental results validate that the comprehensive performance of the proposed algorithm is significantly superior to that of the conventional Time Difference of Arrival (TDOA) algorithm.

Imaging based on signals reflected from global navigation satellite systems (GNSS) represents a novel approach for surface observation. Owing to the global coverage of navigation satellite constellations and the short revisit periods of satellite sub-points, GNSS reflectometry imaging enables high-temporal-resolution change detection in sensitive areas. However, stationary dominant reflectors (SDRs) in the imaging area suppress the signal amplitudes of nearby weak reflection targets, thereby increasing the difficulty of their detection. Conventional coherent change detection and differential interferometry methods demand high phase accuracy and require multiple imaging operations within the imaging domain, thereby reducing detection efficiency. To address these issues, this study proposes a method for directly suppressing SDRs directly in the range-compressed domain. This method leverages the extensive cancellation algorithm batches (ECA-B) framework and utilizes a single satellite revisit period to construct a scattering characteristic model dominated by SDRs, which serves as the reference signal for ECA-B cancellation. To ensure model stability, two-dimensional range-compressed results from multiple prior revisit periods were accumulated to construct a weighted scattering characteristic model, incorporating noise mean and variance. Simulations and real-world experiments demonstrated that the proposed method significantly attenuates signals from strong reflectors in the range-compressed domain, thereby enhancing the detectability of weak change targets.

The Earth–Moon libration point orbit (LPO) offers unique advantages in terms of location and dynamical characteristics, providing new opportunities for designing communication and navigation constellations. However, LPOs cannot be described by Keplerian elements and require significant computational resources for initial value searches, limiting their optimization potential. This paper proposes a two-step optimization algorithm based on a two-layer initial value library. The first layer represents orbit families, while the second layer contains orbits with varying amplitudes within each family. The first step of the algorithm quickly filters orbit families, and the second step selects the orbits that form the constellation, significantly reducing both the optimization scope and the number of parameters. We introduce a novel grid-based division of key cislunar regions and expand the constellation service area using a three-phase construction strategy. To evaluate the optimized constellation’s navigation capabilities, we test its performance with typical orbits, including Earth–Moon transfer orbit (EMTO), elliptical lunar orbits (ELO), and geosynchronous orbit (GEO). Experimental results show that a single near rectilinear halo orbit (NRHO) satellite provides single coverage of the Earth–Moon transfer critical region at sampling times, with an orbit determination (OD) accuracy of 475.7 m for the EMTO. Seven satellites, positioned on the L1, L2, L4, L5, and NRHO orbits, achieve quadruple coverage for both the Earth–Moon transfer and near-Moon regions, with OD accuracies of 25.7 and 17.3 m for the EMTO and ELO, respectively. In the third phase, adding two L3 LPO satellites forms a nine-satellite constellation, extending coverage across the entire cislunar space, achieving OD accuracies of 15.1, 13.4, and 2.2 m for the EMTO, ELO, and GEO, respectively. This study provides valuable insights for the future design and deployment of cislunar communication and navigation constellations.

The growing demand for lunar exploration and infrastructure development requires reliable, real-time navigation for spacecraft in cislunar space, particularly those in distant retrograde orbits and near-rectilinear halo orbits. However, global navigation satellite systems (GNSS) signals are weak at lunar distances, with few visible satellites, high position dilution of precision (PDOP), and low carrier-to-noise ratios (<15 dB-Hz), resulting in poor geometric stability (high PDOP) and moderate ranging errors (meter-level URE), resulting in poor real-time positioning accuracy often exceeding 100 m. Current augmentation strategies, such as deploying active lunar satellites, face challenges in deployment complexity and sustainability. This study proposes a passive position, navigation, and timing (PPNT) framework (relying solely on signal reception without active transmission) that combines lunar and terrestrial beacons to establish a robust spatial reference network. The system leverages dual-anchor signals from both Earth and the Moon to exploit spatial diversity and geometric redundancy, overcoming GNSS limitations. We developed mathematical models for pseudorange measurements, DOP metrics, and positional estimation errors. Over a 30 d simulation, we evaluated three PPNT models (PPNT1: Earth-based only; PPNT2: Earth with a nearside lunar station; PPNT3: Earth with additional lunar polar and farside stations) in terms of geometric visibility, PDOP, C/N-0, receiver noise error and user range error. URE analysis confirms that the PPNT architecture reduces ranging errors from the meter-level (weak GNSS) to the decimeter-level. Results show that PPNT3 significantly outperforms the other two methods due to its geometric configuration, reducing PDOP by 45.9%-54.3%, enhancing stability, and providing the highest carrier-to-noise ratio, achieving centimeter-level receiver noise error and decimeter-level user range error. This research bridges the gap between terrestrial PNT technologies and cislunar mission requirements, offering a novel pathway to improve real-time positioning accuracy in weak GNSS environments. It advances the PNT ecosystem, supporting future lunar infrastructure and deep-space exploration through scalable, passive navigation solutions.

Open set recognition (OSR), which can recognize known classes while reject unknown classes, has become a hot area of research for real world applications. The main challenge of OSR is to effectively discriminate unknowns not seen during training, especially the hard unknown class (HUC) easily confused with knowns. To this end, we propose an open-set prototypical network based on HUC generation (OSPN-HUCG), consists of three main steps: basic closed-set prototypical network (PN) learning for known class separability, data-level HUC generation, and open-set PN learning for knowns and unknowns discrimination. Concretely, on the basis of closed-set PN, we generate HUC data whose features near to known class feature boundaries. To achieve this goal, an HUC feature prior distribution-guided generative adversarial network (GAN) is designed, which aligns the feature distribution of generated data with the prior distribution constructed via perturbed known class boundary features. In open-set PN learning, we devise to reject all unknowns by learning the data of known classes and generated HUC. Motivated by the prior knowledge that unknowns follow lower feature responses than knowns , we progressively constrain the generated HUC towards lower response regions while pushing all knowns to higher response regions, thus both HUC and easy unknown class (EUC) with lower responses can be easily discriminated. Experiments on benchmark datasets and measured dataset validate the state-of-the-art performance of our method.

The Earth Orientation Parameters(EOP) provide a time-varying transition relationship between the International Terrestrial Reference Frame and the International Celestial Reference Frame. To support deep space exploration and the Beidou Navigation Satellite System, the Chinese New-generation Very Long Baseline Interferometry Network(CNVN) is under construction for independent monitoring of the EOP. This paper evaluates the performance of existing 4-antenna CNVN through a batch generated observation schedules followed by extensive Monte Carlo simulations. The optimal positions of the fifth and sixth antennas of CNVN are found from 24hypothetical antenna positions uniformly distributed in China. In this process, the weighted parameters are optimized, which not only reduce the possibility of large error of EOP estimation accuracy due to unreasonable combination, but also greatly reduce the calculation cost.

Open-set recognition (OSR) for synthetic aperture radar (SAR) targets has emerged as a significant research focus. However, existing methods overlook the importance of rejection threshold setting, leading to numerous misclassifications of unknown targets and missing detections of known targets. To tackle this challenge, this paper proposes a novel Rejection Threshold Adaptive OSR method for SAR targets (RTA-OSR). RTA-OSR employs an encoder-decoder network, where the encoder extracts deep features for known class recognition, and the decoder is used to discriminate between known and unknown targets based on the reconstruction errors. Due to the absence of training data for real unknown targets, RTA-OSR designs a pseudo-open environment training method based on the prior knowledge of distribution differences between known and unknown targets to simulate the distribution of unknown targets. Next, a rejection threshold calculation method (RTC) is designed. RTC performs probabilistic fitting between the maximum distribution of reconstruction errors for known targets and the minimum distribution of reconstruction errors for simulated unknown targets, ultimately adaptively calculating the optimal rejection threshold. Experiments based on the MSTAR dataset demonstrate that RTA-OSR can determine reasonable rejection thresholds and achieve satisfactory open-set performance.

Multimodal learning has set off a research hotspot in the remote sensing field due to its powerful performance. Although the paired multimodal training data can be collected offline, some of the data may be missing due to the limitation of the environment or the equipment in practical applications. Therefore, it is of great significance to effectively utilize the complete modal information in the training phase to assist the inference under missing modalities in the testing phase. To address this challenge, we propose a feature fusion-guided cross-modal distillation network. In our method, we design the modal interaction attention fusion module to mine the complementary information from multimodal data, as a multimodal learning model in the complete data scenario. When missing modalities, we use a combined knowledge distillation method to perform cross-modal learning. This is achieved by transferring sample feature similarity relationships and aligning logits output distributions, thereby reducing the redundancy of cross-modal learning feature representations. Finally, we use decision fusion to combine modality-specific information to complete the model inference. The proposed method is experimentally validated on the UNICORN dataset, which can effectively improve the model recognition rate under missing modalities.

To address the limitations of traditional navigation systems in lunar exploration missions regarding positioning accuracy and coverage, this study proposes a multi-objective optimization framework for the Lunar Navigation Satellite System (LNSS) to support the lunar re-exploration mission. A comprehensive simulation system encompassing the entire mission process-powered descent, landing, and surface operations-was developed, integrating a hybrid-orbital architecture (DRO, ELFO, and NRHO) and a multidimensional navigation performance evaluation model. An improved genetic algorithm (IGA) was employed to optimize the constellation configuration under constraints including satellite count (N = 9), geometric dilution of precision, signal attenuation, and cost (75% reduction compared to full-scale architectures), yielding the optimal D1E4N4 configuration. Experimental results demonstrate that the optimized LNSS achieves continuous four-satellite visibility during critical mission phases (landing preparation, powered descent, and surface exploration). The system effectively mitigates signal occlusion caused by complex lunar terrain. These findings provide critical technical support for China's crewed lunar missions and the International Lunar Research Station (ILRS), advancing the transition of Earth-Moon navigation systems from theoretical validation to practical implementation and laying a foundation for sustainable deep-space exploration.

Class-incremental learning enables models with the ability to continuously detect and classify newly appearing targets. However, the absence of old targets in large-scale incremental data makes it challenging to retain previously acquired knowledge. Therefore, this paper proposes a scene characteristic-guided class-incremental synthetic aperture radar (SAR) ship detection and classification, which better alleviates the forgetting of old knowledge in incremental models. Specifically, we propose a context-robust exemplar replay (ER) method to retain precise scene context between replayed old targets and clutter in the incremental data. Fully considering the scene characteristics of the ship targets, the context-robust ER method combines a local context-aware strategy with sea-land segmentation to pinpoint regions with the highest likelihood of being sea areas. Then, context-robust ER method replays exemplars within these regions, effectively reducing contextbias issues over existing ER methods. In addition, we also propose a multi-granularity knowledge distillation (KD) method, which transfers the knowledge learned by the old model to the incremental model progressively. Multi-granularity KD method combines the target localization information with spatialchannel attention mechanisms to constrain the incremental model to focus on the most important information instead of redundant information, thereby preventing the negative distillation issues over existing KD methods. Experiments on SRSDD-v1.0 dataset indicate that our method achieves satisfactory incremental performance over existing methods. Our codes is published in https://github.com/LiYiming98.