Building footprint extraction using optical and synthetic aperture radar (SAR) images enables all-weather capability and significantly boosts performance. In practical scenarios, optical data may not be available, leading to the missing-modality challenge. To overcome this challenge, advanced methods employ mainstream knowledge distillation approaches with hallucination network schemes to improve performance. However, under complex SAR backgrounds, current hallucination-network-based methods suffer from cross-modal information transfer failure between optical and hallucination models. To solve this problem, this study introduces a cross-modal hallucination collaborative learning (CMH-CL) method, consisting of two components: modality-share information alignment learning (MSAL) and multimodal fusion information alignment learning (MFAL). The MSAL method facilitates cross-modal knowledge transfer between optical and hallucination encoders, thereby enabling the hallucination model to effectively mimic the missing optical modality. The MFAL method aligns semantic information between OPT-SAR and HAL-SAR fusion heads to strengthen their semantic consistency, thereby improving HAL-SAR fusion performance. By combining MSAL and MFAL, the CMH-CL method collaboratively alleviates cross-modal transfer failure problem between the optical and hallucination models, thereby improving performance in missing-modality building footprint extraction. Extensive experimental results obtained on a public dataset demonstrate the effectiveness of the proposed CMH-CL. The source code is available at https://github.com/TINYWAI/CMH-CL.

To address the positioning accuracy limitations of trajectory correction fuzes in complex battlefield environments, this study innovatively integrates the fourth-order Runge-Kutta (RK4) numerical integration method with the Extended Kalman Filter (EKF) to propose an improved ballistic trajectory estimation algorithm. The algorithm reconstructs the Jacobian matrix computation process within the EKF framework using the RK4 method and introduces a dynamic adjustment matrix to compensate for linearization errors. A full-state estimation model for four-degree-of-freedom ballistic motion is re-derived and established. Simulation results demonstrate that the new method achieves significant improvements over the traditional EKF algorithm. Compared to other optimization methods, it provides comparable positional estimation accuracy while achieving superior velocity estimation precision, also maintaining robust performance under extreme conditions. Furthermore, field firing tests validate the algorithm's effectiveness and stability in real-flight environments, laying a foundation for subsequent high-efficiency correction control.

To address false detections and resource constraints in offshore platform PPE monitoring, we propose YOLOv11-G, a lightweight detection model. It integrates Efficient Channel Attention (ECA) to recalibrate SPPF features and suppress background noise, alongside a Global Fusion Feature Module (GFFM) for decoupled cross-scale interaction. Furthermore, a Granularity-Enhanced Neck (GEN) is introduced to capture fine-grained information for small targets. Compared to the baseline, YOLOv11-G increases $\mathrm{mAP} \text{@} 0.5$ by $3.0 \%, \mathrm{mAP} \text{@} 0.5: 0.95$ by 3.6%, and Recall by 5.7%. This provides an efficient, feasible solution for real-time safety monitoring on edge devices.

On-board target detection demands high computational throughput under stringent energy and memory budgets, thereby necessitating architectures that tightly couple high parallelism to memory-efficient data orchestration. This article presents SHARP, a scalable field-programmable gate array (FPGA)-oriented convolutional neural network (CNN) accelerator tailored for real-time, resource-constrained remote-sensing deployments. SHARP increases sustained concurrency by adopting a packet-based multilevel parallel convolution strategy that reduces intermediate storage while enabling high-degree parallel processing of large multichannel feature maps. Building on this compute fabric, a scalable multilevel parallel CNN compute engine is provided so that the effective parallelism can be tuned to match available digital signal processor (DSP), lookup table, and block random access memory (BRAM) resources; this flexibility enables efficient mapping across devices with different resource budgets. To preserve throughput in practice, an integrated off-chip and on-chip data access strategy combines DDR interleaving with multi-BRAM dislocated caching and ping-pong buffering, thereby improving burst bandwidth utilization and removing read/write conflicts so I/O can be overlapped with computation. A theoretical resource-analysis framework ties these elements together by quantifying achievable parallelism under DSP and BRAM constraints and guiding hardware dimensioning. SHARP is implemented on a Xilinx XC7VX690T. The prototype attains a peak throughput of 1480.95 GOPS at 250 MHz with 20.525 W power and delivers 72.15 GOPS/W energy efficiency while maintaining 98.47% mAP. Compared with reported FPGA YOLO accelerators, these results demonstrate that SHARP achieves a superior tradeoff between throughput and energy efficiency for real-time, on-board remote-sensing applications.

Real-time intelligent processing of remote sensing data on resource-constrained edge platforms requires lightweight models with high performance. Conventional lightweight artificial neural networks (ANNs) incur substantial computational costs due to dense multiply-accumulate operations. In contrast, spiking neural networks (SNNs) reduce computation by simplifying calculations to accumulation-based operations driven by sparse spikes. However, existing SNN design methods typically optimize network architecture, quantization, and time steps separately. This fragmented approach often relies on expert heuristics, thereby limiting joint optimization of model efficiency. To address this, we propose a unified differentiable architecture search (UDAS) framework for SNNs targeting resource-constrained remote sensing applications. It constructs a unified search space integrating critical design parameters for lightweight models, including network topology (operators and channel counts), quantization bit-widths, and time steps. Our framework innovatively treats the time step as a layerwise optimizable variable through the proposed temporal integration strategy, enabling joint spatiotemporal optimization. To accelerate convergence, we introduce the target bit-width-aware co-search method, which refines the search space. Experiments on public datasets (UC-Merced, WHU-RS19, aerial image dataset (AID), and NWPU VHR-10) demonstrate that UDAS-designed SNNs achieve performance competitive with state-of-the-art (SOTA) lightweight ANNs under strict resource constraints. For instance, on WHU-RS19, our SNN achieves 96.94% accuracy (versus 96.63% for the ANN baseline) while achieving an 8 & times; reduction in storage (1.69 versus 13.89 MB) and up to 33.23 & times; compression over its full-precision counterpart. Our UDAS framework offers a practical way to fully exploit the efficiency advantages of SNNs in real-time remote sensing applications on resource-constrained devices. Code is released at https://github.com/GitHub-ljh/UDAS.

With the rapid expansion of the low-altitude economy, unmanned aerial vehicles (UAVs) equipped with vision systems have become essential platforms for various aerial perception tasks; nevertheless, optimizing the accuracy-efficiency trade-off for small object detection in aerial imagery remains a major challenge. To address this challenge, we propose a new real-time detector for aerial imagery based on YOLO11, named RSS-YOLO. First, a Reciprocal Interaction Fusion (RIF) module is designed to replace the C3k2 module in YOLO11. This module enables complementary information exchange across multiple feature levels via bidirectional feature interaction. By preserving spatial localization details while refining semantic cues, it alleviates insufficient integration of spatial and semantic information within the feature extraction layers. The neck architecture is further optimized by introducing a P2 small object detection layer integrating an improved SCSA attention mechanism, while removing the redundant P5 detection layer. Additionally, replacing standard convolutions with the L-Down module in deep networks strips away redundant parameters and computation without sacrificing performance. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method improves mAP @0.5 and mAP @0.5:0.95 by 3.8

Cross-domain few-shot semantic segmentation (CD-FSS) aims to segment novel classes under limited supervision and domain shifts. In this work, we focus on a more challenging yet practically important subtask, namely natural-to-remote-sensing CD-FSS, where knowledge learned from natural-scene data is transferred to diverse remote sensing target domains. Compared with generic unseen-domain transfer, this setting is particularly challenging due to the complex background structures, large scale variations, and substantial imaging discrepancies in remote sensing imagery, which make accurate segmentation in such scenarios difficult. To address these issues, we propose DFANet, a unified framework tailored for natural-to-remote-sensing CD-FSS. We first design a multi-dimensional prototype reconstruction (MPR) module, which refines class prototypes through query-guided attention to enhance prototype discriminability under complex scene structures and ambiguous foreground-background boundaries. We then introduce a dynamic feature alignment (DFA) module to align support and query features from both local spatial structures and global channel statistics, thereby alleviating cross-scene feature mismatches caused by large scale variations and appearance differences. Furthermore, we develop a test-time anchored support finetuning (TASF) strategy, which performs lightweight target-domain adaptation using a few annotated support samples as anchors to further reduce the residual domain discrepancy. On three representative remote sensing datasets, including Deepglobe, Vaihingen, and Potsdam, DFANet achieves average MIoU scores of 40.24%, 47.07%, and 48.95% under 1-shot, 5-shot and 10-shot settings, respectively, outperforming existing CD-FSS methods for natural-to-remote-sensing evaluation. Code is publicly available at https://github.com/zhangxiaogang-111/DFANet.

Recent theoretical advances reveal that the Hadamard product induces nonlinear representations and implicit high-dimensional mappings for the field of deep learning, yet their practical deployment in efficient vision models remains underdeveloped. To address this gap, we introduce the Adaptive Cross-Hadamard (ACH) module, a novel operator that embeds learnability through differentiable discrete sampling and dynamic softsign normalization. This enables parameter-free feature reuse while stabilizing gradient propagation. Integrated into Hadaptive-Net (Hadamard Adaptive Network) via neural architecture search, our approach achieves unprecedented efficiency. Comprehensive experiments demonstrate state-of-the-art accuracy/speed trade-offs on image classification task, establishing Hadamard operations as fundamental building blocks for efficient vision models.

Driven by the rapid expansion of the low-altitude economy, unmanned aerial vehicles (UAVs) have emerged as pivotal platforms for aerial perception. However, balancing detection accuracy and computational efficiency for micro-targets remains a formidable challenge. To address this, we propose LSO-YOLO, a novel real-time detector optimized for UAV-based small-object detection. Our primary contributions are threefold: first, a multi-scale dilated context aggregation (MDCA) module is designed to replace the final down-sampling convolution in the backbone. By employing sequential dilated depth-wise convolutions and a lightweight channel attention mechanism, the MDCA expands the receptive field and suppresses background noise, effectively mitigating the loss of fine-grained details caused by aggressive down-sampling; second, the C2f module in the neck is redesigned by substituting the second convolution in the bottleneck with parallel grouped convolutions, which enhances feature diversity and multi-scale representation while reducing computational overhead; finally, we optimize the overall architecture and introduce a lightweight down-sampling strategy to mitigate inter-channel interference. Experimental results on the VisDrone2019 dataset demonstrate that the proposed method improves mAP@0.5 and mAP@0.5:0.95 by 2.9 https://github.com/diapfer/LSOYOLO .

In unsupervised domain adaptation for semantic segmentation of remote sensing imagery, identical land-cover classes across different domains often exhibit substantial variations in appearance, scale, and class distribution, which seriously hinder cross-domain generalization. Moreover, even within a single domain, land-cover classes present highly complex and diverse intra-class distributions that cannot be effectively captured by a single class representation, further increasing the challenge of generalization. To this end, we propose a collaborative framework integrating local pixel-level contrast and global Gaussian multi-prototype bidirectional alignment. At the local level, we introduce probability-masked contrastive learning, which adaptively increases the sampling probability of minority classes to mitigate the class imbalance issue. Meanwhile, pixel contrastive learning is incorporated to enhance the robustness to cross-domain variations in appearance and texture. At the global level, we employ a Gaussian mixture model to represent each source domain class with multiple Gaussian prototypes rather than a single one, thereby yielding richer and more fine-grained class representations. Building on this, a bidirectional alignment strategy is proposed. Concretely, the forward alignment serves multi-prototypes as semantic anchors that progressively guide target-domain features to align with the source-domain class distributions, reducing the intra-class variance. Meanwhile, the reverse alignment dynamically refines the prototypes to increase anchor accuracy, further enhancing the stability and discriminability of cross-domain alignment. Experimental results on the widely used ISPRS and LoveDA datasets demonstrate the superiority of our proposed method over state-of-the-art approaches.

Unsupervised Domain Adaptation (UDA) for semantic segmentation transfers knowledge from an annotated source domain to an unlabeled target domain, reducing the cost of pixel-level annotation. Most existing methods rely on dense discriminative classifiers that estimate pixel-wise class posterior probabilities using a single weight vector per class. However, in remote sensing imagery, intra-class variability often gives rise to multi-peaked feature distributions with multiple sub-clusters, which may not be adequately captured by a single weight vector, leading to intra-class confusion. Moreover, such discriminative classifiers primarily focus on learning class decision boundaries, while the consistency of same-class sub-clusters across domains is only weakly constrained. Under domain shift, these sub-clusters may drift inconsistently, thereby aggravating cross-domain structural mismatch. To this end, we propose GMMDA, a Gaussian Mixture Model (GMM)-based generative domain adaptation method that predicts pixel labels through Bayesian inference based on class-conditional feature likelihoods. Specifically, we fit a class-wise GMM using an online momentum Sinkhorn EM procedure to model class-conditional feature densities, where each Gaussian component corresponds to an intra-class sub-cluster, enabling a richer characterization of multi-peaked distributions. Furthermore, component-aware contrastive learning treats Gaussian components as fine-grained semantic references, pulling target-domain features toward their best-matched components using a variance-aware distance that accounts for component scales. In addition, we introduce component-wise distribution alignment to align each source component with the corresponding source-guided target local statistics by jointly constraining their means and principal directions, aiming to reduce cross-domain structural mismatch. Experimental results show that GMMDA achieves competitive mIoU scores on the ISPRS and LoveDA benchmarks compared with existing methods.

Tiny object detection in remote sensing images has long been challenging due to the low resolution of tiny objects and complex backgrounds. However, existing adaptive receptive field methods mainly focus on object and contextual information, making them susceptible to interference from complex backgrounds. This interference exacerbates the imbalance in sample assignment and hinders accurate adaptation of receptive fields to object scales, resulting in insufficient representation of fine-grained object details. In addition, most loss functions are prone to gradient instability in scenarios involving small objects. To address the aforementioned problems, this paper proposes a DCP-Net, which learns detail–context perception under spatial-frequency guidance. To enhance the representation of details of tiny objects, the network is the first to leverage spatial frequency to distinguish the structural features of objects and background. By introducing a spatial-frequency guided dual-dilated convolution, it effectively enhances local object details and global semantic associations while suppressing interference from complex backgrounds. Based on this convolutional unit, we further design a cross-layer fine-grained fusion module to recover lost object details at a lower computational cost. Furthermore, we introduce a Gaussian classification-regression loss to address the issues of imbalanced sample allocation and unstable gradient, while applying a decoupled probabilistic metric strategy and a multi-scale enhanced decoupled head to improve localization accuracy. Extensive experimental results demonstrate the effectiveness of the proposed method. Specifically, DCP-Net achieves an average precision (AP) of 27.5% on the AI-TODv2 dataset and an AP50 of 78.0% on the DOTA-v1.0 dataset. The code will be found at: https://github.com/yangyvni/DCP-Net.

Single-source Domain Generalization (SDG) in object detection aims to develop a detector using only source-domain data while maintaining strong generalization capability on unseen target domains. Existing methods are mainly built on CNN-based detectors and improve generalization through carefully designed data augmentation strategies. However, augmentation-based methods are inherently effective only when the distribution of augmented samples can approximate or cover unseen scenarios, which limits their ability to improve generalization consistently across diverse target domains. Moreover, although recent detection transformers have achieved strong performance in general object detection by means of effective global feature extraction, their potential for SDG object detection remains largely unexplored. To this end, we propose a strong DETR-based detector, termed the Style-Adaptive DEtection TRansformer (SA-DETR), which improves generalization by representing unknown target styles with a compact set of reusable style bases rather than attempting to explicitly cover the full diversity of target-domain styles. SA-DETR introduces an Online Domain Style Adapter (ODS-Adapter), which builds a self-organizing multi-basis style memory from channel-wise backbone statistics and updates it through discrepancy-aware merge and replacement operations. During inference, a discrepancy-weighted AdaIN-style projection maps target statistics onto the learned source-style manifold, and an optional test-time assimilation mechanism is further used to alleviate under-adaptation under severe shifts. In addition, we introduce an Object-aware Contrastive Learning (OCL) module to encourage the extraction of domain-invariant object features through contrastive learning. By using carefully designed object-aware gating mechanisms to restrict the range of feature aggregation in both spatial position and semantic category, this module enables effective cross-domain contrast at the instance level and further improves the detector’s generalization capability. Code is released at https://github.com/h751410234/SA-DETR.

Onboard real-time synthetic aperture radar (SAR) imaging is essential for time-sensitive applications but faces strict power constraints on satellites. Heterogeneous coarse-grained reconfigurable arrays (CGRAs) with circuit-switched networks offer an efficient and flexible platform for SAR computing, yet existing mapping methods fail to jointly optimize power and compilation time efficiency. This article introduces SAR-aware Power-and-Compilation-Efficient Mapping (SPCMap), a SAR-aware mapping framework for heterogeneous CGRAs that tackles this joint challenge. SPCMap features four innovations. 1) a power-aware modulo routing resource graph modeling method for accurate power estimation. 2) An anchor diffusion method leveraging the multisource convergence pattern of SAR data flows to prune the search space. 3) a nongreedy progressive search method to escape local optima. 4) A parameterized active backtracking method for robustness. Evaluated on 13 real SAR data flow graphs (DFGs) using a 12-nm CGRA platform, SPCMap achieves 100% mapping success. It reduces router power by up to 62.7%, with an average deviation of only 2.4% from the theoretical optimum, while keeping compilation time within 1.4 & times; of the fastest greedy baseline. Ablation studies confirm the contribution of each component. In cross-platform comparisons, the proposed CGRA with SPCMap attains a leading normalized energy efficiency of 3.44 & times;10(7) pixel(2)/s/W, outperforming existing processors. SPCMap thus balances power, compilation time, and reliability, providing a practical compilation solution for energy-efficient onboard SAR imaging.

Deep learning-based detectors for remote sensing object detection have achieved remarkable success, but their performance heavily depends on large-scale annotated datasets, which are costly and time-consuming. Few-shot remote sensing object detection has therefore emerged as a promising paradigm to recognize novel categories using only a few labeled samples, where transfer-learning-based methods can achieve excellent performance via simple fine-tuning. However, high inter-class similarity still poses a significant challenge, leading to misidentification among confusing categories. To address this problem, a text semantic prior few-shot object detection network (TSP Net) is proposed for few-shot remote sensing object detection. A remote sensing object characteristic prompt is designed, and a CLIP text encoder is used to obtain class-level semantic relationships, which serve as the text semantic prior. This prior is then coupled with contrastive learning in the detection network, where inter-class feature repulsion is adaptively weighted according to semantic similarity, thereby enlarging distance among confusing categories. Integrating text semantic prior with contrastive learning in a unified detection network, TSP-Net effectively enhances feature discrimination, improving detection performance. Extensive experiments on the DIOR and NWPU VHR-10 benchmarks demonstrate that the proposed method consistently improves novel-class detection performance, achieving notable gains under challenging 3-shot and 5-shot settings. The source code will be made publicly available at https://github.com/panghumt/TSP-Net.

Tiny object detection in remote sensing typically faces the challenges of being submerged in backgrounds, limited feature representation, and high sensitivity to prediction errors due to the small size and diverse shapes. To address these challenges, a Geometric Guided Noise Reduction Super-Resolution Network (GNSRNet) is proposed. First, a dual-branch denoising and super-resolution feature pyramid network is proposed, which integrates an adaptive dynamic noise reduction module and an inference decoupled auxiliary super-resolution branch, while a progressive loss-annealing strategy is further introduced to reduce reliance on the super-resolution branch during inference, meeting the requirements of lightweight and high-performance remote sensing tiny object detection. Second, a geometric characteristic regression metric is proposed, which comprehensively considers the location accuracy and the shape similarity between the prediction and ground truth boxes, thereby improving bounding box quality and detection precision. Extensive experiments have been conducted on the remote sensing tiny object datasets AI-TOD v1, AI-TOD v2, USOD, and VisDrone. Specifically, it reaches an AP of 31.6 on AI-TOD v1, 30.5 on AI-TOD v2, 37.4 on USOD, and 30.5 on the VisDrone, demonstrating its capability for tiny object detection in remote sensing. The code is available at https://github.com/worthlessjl/GNSRNet.git.

Drone-based RGBT person detection, which has garnered significant attention, seamlessly integrates the spatial flexibility of drones with the around-the-clock time flexibility of RGBT data for continuous information acquisition. However, the drone-captured images cover a wide range of areas, resulting in illumination imbalance and thermal background clutter issues in RGB and thermal images. The varying quality of the two modalities in different regions of an RGBT image set increases the difficulty of complementary information fusion, thereby leading to false negatives in the advanced drone-based RGBT tiny person detection methods. In this context, a novel Cross-modal Complementary Region-aware framework for drone-based RGBT tiny person Detection (CCRDet) is proposed for effective object detection under poorly illuminated RGB regions and thermally cluttered backgrounds. CCRDet employs the proposed cross-modal region-aware guidance to be aware of these regions and guide the counterpart modality to enhance valid target features accordingly. After that, it leverages the proposed modality-difference feature gated fusion to deliver these valid target features to the fused features with effective preservation, thereby enhancing their response intensity after fusion and providing high-quality inputs for the detection head. Extensive experiments on two drone-based RGBT tiny person detection datasets, RGBT-DronePerson and VTUAV-det, demonstrate the effectiveness of the proposed method. The code is available at https://github.com/G-pz/CCRDet.