To address imaging degradation in visible and infrared modalities under complex environments, as well as the limitations of existing multimodal detection methods in cross-modal complementarity, conflict suppression, and fusion transfer, this paper proposes a hybrid-level dual-input mamba fusion (HDMF) framework for UAV remote-sensing vehicle detection. HDMF consists of visible, infrared, and fusion branches. At the feature level, dual-branch simAM–wavelet attention (DSWA) performs wavelet-based complementarity only in highly complementary regions through discrepancy-aware gating, thereby suppressing noise amplification and false-response transfer. A binary fusion mamba (BFMamba) module is further introduced to jointly model visible, infrared, and interaction features in the hidden-state space, enabling long-range cross-modal dependency modeling with linear complexity. At the decision level, a tri-branch aware decision fusion head (TDFH) explicitly incorporates illumination priors and cross-branch consistency constraints during inference to compensate for residual modality mismatch in complex scenes, improving overall stability and reliability. Experiments on Drone Vehicle, LLVIP, and FLIR achieve 56.62\%, 69.82\%, and 48.11\% mAP, respectively, demonstrating that dual-input hybrid-level fusion at the state-space level is an effective way to balance modeling capability, robustness, and efficiency.

Deep learning-based hyperspectral image (HSI) classification has significant applications in remote sensing scene understanding. Whole-image propagation classification methods can achieve high-precision classification at the full-pixel level with a significant efficiency advantage. However, such methods generally face challenges such as the smoothness of high-frequency details and difficulty in efficiently modeling deep features. To address these issues, this article proposes a fast multikernel graph convolutional network (FMGCN) for fast whole-image classification. The model designs a multikernel graph convolution (MKGC) algorithm to enhance the expression ability of deep spectral-spatial features. This algorithm generalizes the graph convolution process from a single kernel propagation process to a propagation process with an arbitrary number of kernels to more fully exploit spectral-spatial features in non-Euclidean spaces. In addition, a symmetric compressed refinement network (SCRN) is constructed to restore the weakened local structural information in the whole-image propagation, and an embedded edge compensation strategy (EECS) is combined to further strengthen the high-frequency features in the boundary regions. Based on this design, FMGCN is able to capture more comprehensive global dependency information in a single whole-image propagation, achieving a uniform balance of high accuracy and high efficiency. Experiments on four representative hyperspectral datasets show that FMGCN achieves the best classification performance, effectively alleviates over-smoothing in complex structural scenarios, and outperforms batch propagation models in terms of inference efficiency.

Small weak object detection (SWOD) is a significant but neglected task in remote sensing image interpretation. Due to limitations in imaging resolution and inherent characteristics of the objects, detection networks struggle to effectively extract semantic features, which are crucial for object identification and recognition. In recent years, graph convolutional networks (GCNs) have been developed to handle non-Euclidean data. Through GCNs, node data are enriched via aggregation and propagation across the graph. In this article, we explore the feasibility of GCNs in semantic clue extraction to address the lack of key semantics in small weak objects. First, we propose a multihead graph reasoning learning model (MGRL) that projects initial feature representations into graph space and utilizes a two-layer multihead graph network to extract essential semantic information. Second, we introduce a foreground-background binary masking technique that roughly segments the foreground region of the image. The mask is converted into a prior prompt, which is then incorporated into the adjacency matrix, emphasizing object reasoning in MGRL. Next, we present a cross learning-based feature alignment learning module to resolve feature misalignment issues caused by spatial projection. Finally, we adopt a cross-layer semantic interaction module to facilitate cross-layer communication and aggregation of features. Extensive experiments are conducted on five remote sensing datasets: DIOR, AI-TOD, NWPU VHR-10, DOTA-v1.0, and STAR. The experimental results demonstrate the superior performance and advantages of our method.

The explosive growth of hyperspectral data poses severe challenges to the storage, transmission, and real-time processing capabilities of spaceborne platforms. Existing band selection methods often struggle to balance computational efficiency with feature representation, typically neglecting spatial context and relying on computationally expensive iterative search procedures. To address these limitations, this paper proposes a hardware-friendly multi-Feature enhanced curve-fitting subspace partition (MFE-CFSP) strategy and implements a highperformance field-programmable gate array (FPGA) accelerator. Algorithmically, we introduce a block-wise maximum entropy pre-screening mechanism to filter background redundancy at an early stage, and propose a multi-dimensional feature enhancement model that fuses spectral variance, Sobel operator edge intensity, and sum of absolute differences (SAD)-based texture details. Furthermore, a two-stage global optimization strategy employing recursive Cholesky updates is designed to reduce the overall search complexity from 4 ()OK to 3 ()OK, effectively bounding the per-iteration evaluation cost to 2 ()OK .Regarding hardware implementation, we propose a novel two-pass dataflow scheduling architecture on the Xilinx Zynq UltraScale+ platform that decouples global scanning from local feature extraction to overcome the memory wall bottleneck. Additionally, a triangular systolic array co-processor is implemented to accelerate the recursive optimization process. Finally, to address the heterogeneous precision sensitivity of different operators, we introduce an evolutionary mixedprecision quantization strategy. Experimental results on four benchmark datasets demonstrate that MFE-CFSP consistently outperforms five state-of-the-art band selection methods across multiple classifiers, exhibiting superior performance particularly with a low number of target bands. Beyond classification, the selected subsets also achieve the highest detection accuracy on target detection (CEM) and anomaly detection (RX) tasks, and the best spectral reconstruction fidelity (SSIM), confirming the generalizability of the proposed strategy. On the Xilinx ZCU102 platform, the proposed accelerator achieves a speedup of up to 219.8 and an energy efficiency improvement of 308.5 compared to an embedded ARM Cortex-A53 CPU. Furthermore, on the resource-constrained KV260 edge device, the mixedprecision strategy reduces look-up tables and block random access memory utilization by 8.8% and 12.5%, respectively, with negligible impact on inference time and accuracy.

Hyperspectral image (HSI) classification models commonly analyze fixed-size patches or rely on predefined scales in multi-scale approaches. This rigidity prevents them from adaptively selecting the optimal field of view, which varies for different ground objects. To overcome this limitation, this paper proposes a novel HSI classification framework named dynamic scale and graph optimal transport (DSGO). DSGO features two synergistic branches: an adaptive scale vision transformer (AS-ViT) branch and a dynamic graph optimal transport (DGOT) branch. In AS-ViT, which integrates multi-scale information, a learnable scale selector (LSS) dynamically updates the spatial sampling range for each input. The DGOT branch simultaneously learns the optimal graph structure by updating both nodes and their adjacency matrix. To effectively fuse features from these branches and mitigate potential conflicts, DSGO incorporates a dual-channel graph optimal transport module (DGOM). Inspired by domain alignment and graph comparison theory, DGOM facilitates gradual and compatible feature fusion by optimizing cross-domain optimal transport distances between the graph Laplacian matrices of the two branches. Extensive experiments on four benchmark HSI datasets (Pavia University, Houston, Salinas Valley, and Indian Pines) demonstrate that DSGO achieves state-of-the-art performance, with overall accuracies of 99.43%, 98.50%, 99.42 %, and 98.82 %, respectively, representing improvements of about 1-2 % over existing methods. Future work will explore advanced lightweight strategies and structural optimization techniques to enhance the efficiency and deployability of DSGO.

Structured pruning has become a mainstream approach for network lightweighting due to its advantages in terms of compression efficiency and deployment friendliness. Existing methods generally struggle to generate a complete Pareto front and cannot easily incorporate practical inference performance metrics directly into the optimization objective. This significantly constrains their capacity to recommend pruning subnet schemes and prevents them from adequately satisfying the actual requirements of specific tasks. To address the aforementioned challenges, we propose DRL-Pruner, a pruning framework guided by multi-objective reinforcement learning that caters to diverse budget optimization requirements. Leveraging the powerful generative capability of diffusion model, this method gradually iteratively refines randomly initialized pruning strategy into compression schemes that can fully cover the entire objective space. We statistically measure data such as the global computational variation and structural sensitivity of the network as state representation, providing the pruning framework with an observation set of structural states and performance trends. For remote sensing vision tasks, we conduct pruning experiments on ResNet-50, ViT-Tiny, and Faster R-CNN under the scene classification benchmark NaSC-TG2 and the object detection benchmark NWPU VHR-10 (v2). The results demonstrate that DRL-Pruner achieves an effective tradeoff between network accuracy and various inference budgets, providing pruning decisions that cover the Pareto front.

Despite advances in remote sensing object detection, accurately identifying small, weak objects remains challenging. Their limited pixel representation often fails to capture distinctive features, making them susceptible to environmental interference. Current detectors frequently miss these subtle feature variations. To address these challenges, we propose FCDet, a feature contrast-based detector for small, weak objects. Our approach introduces: (1) a spatial-guided feature upsampler (SGFU) that aligns features by adaptive sampling based on spatial distribution, thus achieving fine-grained alignment during feature aggregation; (2) a feature contrast head (FCH) that projects GT and RoI features into an embedding space for discriminative learning; and (3) an instance-controlled label assignment (ICLA) strategy that optimizes sample selection for feature contrastive learning. We conduct comprehensive experiments on challenging datasets, with the proposed method achieving 73.89% mAP on DIOR, 95.04% mAP on NWPU VHR-10, and 26.4% AP on AI-TOD, demonstrating its effectiveness and superior performance.

In current super-resolution (SR) research, blind SR models capable of handling multiple degradations have attracted significant attention. Inspired by variational autoencoders (VAEs) that model data distributions through latent representations, this paper proposes a VAE framework for unsupervised remote sensing image (RSI) SR. VAEs excel at learning rich latent representations, modeling probabilistic distributions of input data and unsupervised learning, making them inherently well-suited to real-world blind SR scenarios. The proposed framework consists of an encoder that maps low-resolution (LR) images into a latent space and a decoder that reconstructs super-resolved images from the latent representations. To enhance latent modeling, an alternating optimization strategy is implemented for training the encoder and decoder. Furthermore, a comprehensive loss function and a latent coding regularization strategy are designed to constrain latent representations while maintaining image domain consistency. Experimental results demonstrate that on synthetic data, our method achieves favorable performance in both visual quality and quantitative metrics. It also demonstrates competitively performance compared to supervised methods, particularly in 4× and 8× SR tasks. Additionally, evaluations on Jilin-1 satellite RSIs further validate the effectiveness of our approach.

Existing mainstream hyperspectral image (HSI) classification frameworks, such as multilayer perceptron (MLP) networks, combine learnable linear transformations with nonlearnable nonlinear activation functions. While these feature extraction algorithms achieve high classification accuracy, they require significant computational and memory resources, and many models lack adequate informative representations. Recently, fast Kolmogorov-Arnold networks (FKANs) have emerged as a lightweight alternative to MLPs, offering competitive performance with fewer parameters and lower computational costs. This article presents a hierarchical progressive fusion network (HPFN) based on FKAN to address these challenges. The proposed network explores HSI information more comprehensively through branches at three scale levels (pixel, patch, and global) while remaining lightweight. The orthogonality of information at different levels is preserved to some extent for diverse practical applications. We develop two FKAN-based lightweight convolutional networks to extract pixel- and patch-level features. In the global-level branch, a boundary-enhanced U-Net with an edge-constrained module based on Sobel operators mitigates oversmoothing. The cross-scale guided fusion mechanism allows for dynamic awareness and accurate pairing of pixel-level features with local and global features, reducing information redundancy. Experimental results show that the proposed network achieves competitive classification accuracy on four public datasets at low computational cost, validating its effectiveness.

The joint semantic segmentation of multi-modal remote sensing images can make up for the problem of insufficient features of single-modal images and effectively improve the classification accuracy. Some deep learning methods have achieved good performance, but they face problems such as complex network structure, large number of parameters, and difficulty in deployment. In this paper, more attention is paid to front-end and branch-level feature transformation to obtain multi-scale semantic information. The multi-scale dilated extraction module (MDEM) is constructed to mine the specific features of different modalities. The multi-modal complementary attention module (MCAM) is designed for further acquiring prominent complementary content. The concatenated features are transmitted and reused by the dense convolution to complete the encoding. Ultimately, a general and concise end-to-end model is proposed. Comparative experiments are carried out on three heterogeneous datasets, and the model put forward performs well in qualitative analysis, quantitative comparison and visual effect. Meanwhile, the dexterity and practicability of the model are more prominent, which can provide support for lightweight design and hardware deployment.

Benefiting from the development of deep learning, the super-resolution technology for remote sensing hyperspectral images (HSIs) has achieved impressive progress. However, due to the high coupling of complex components in remote sensing HSIs, it is challenging to achieve a complete characterization of the internal information, which in turn limits the precise reconstruction of detailed texture and spectral features. Therefore, we propose the multi-attitude hybrid network (MAHN) for extracting and characterizing information from multiple feature spaces. On the one hand, we construct the spectral hypergraph cross-attention module (SHCAM) and the spatial hypergraph self-attention module (SHSAM) based on the high and low-frequency features in the spectral and the spatial domains, respectively, which are used to capture the main structure and detail changes within the image. On the other hand, high-level semantic information in mixed pixels is parsed by spectral mixture analysis, and semantic hypergraph 3D module (SH3M) are constructed based on the abundance of each category to enhance the propagation and reconstruction of semantic information. Furthermore, to mitigate the domain discrepancies among features, we introduce a sensitive bands attention mechanism (SBAM) to enhance the cross-guidance and fusion of multi-domain features. Extensive experiments demonstrate that our method achieves optimal reconstruction results compared to other state-of-the-art algorithms while effectively reducing the computational complexity.

Object detection is of great importance for remote sensing image interpretation work and has received significant attention. However, small weak object detection has always been a challenge. The main reason is that the critical information of these objects, such as textures and boundaries, is suppressed by the background and cannot effectively express their own characteristics. To address this issue, we introduce a novel texture and boundary-aware network (TBNet) in this paper. Firstly, we propose a texture-aware enhancement module (TAEM) to explore the texture details within the images. TAEM captures pixel correlations to perceive the distribution of texture in the objects. Secondly, a boundary-aware fusion module (BAFM) is introduced to emphasize spatial positions. BAFM can extract the edge information to guide the prediction of small weak objects. Finally, a task-decoupled RCNN (TD-RCNN) is designed to separate classification and regression tasks. TD-RCNN achieves fine-grained detection, avoiding compromises between subtasks. Comprehensive experiments on four public datasets, DIOR NWPU VHR-10, RSOD, and AI-TOD, demonstrate that TBNet achieves state-of-the-art performance compared to competitors. The model is also evaluated on UAVOD-10, which collects numerous small weak objects. TBNet achieves state-of-the-art results while significantly outperforming competitors, proving its ability to detect small weak objects.

Arbitrary-oriented ship detection is a challenging task in optical remote sensing images (ORSIs). Existing methods mainly focus on improving model structures but seldom consider ship characteristics, such as shape and multiscale distribution. This results in prior ship-related information not being effectively utilized. To address these issues, a ship detector based on ellipse feature encoding (EFE) and balanced decoupling label assignment strategy (EBDet) is proposed, aiming to effectively integrate ship features during the crucial detection process. Specifically, the EFE strategy incorporates ship shape to represent the oriented bounding boxes (OBB) with seven parameters, and addresses the boundary discontinuity problem by resolving the contradiction between angular discreteness and network continuity. The ellipse balanced sampling (EBS) strategy fits the size and shape of ships to achieve a balance in the number of positive samples between ships of different scales. The dynamic decoupling label assignment (DDLA) strategy adaptively evaluates the sample space sensitivity, and mitigates the feature misalignment problem by decoupling positive samples in classification and regression. Extensive experiments demonstrate the effectiveness and advancement of our EBDet. Specifically, on the HRSC2016 dataset, EBDet achieves an mAP of 90.53% and 90.64% with ResNet-50 and ResNet-101 backbones, respectively. On the DIOR-Ship dataset, EBDet achieves mAP of 86.69% .

With the thriving development of deep learning, the hyperspectral image (HSI) super-resolution (SR) technology has made significant progress. Existing methods often ignore the task orientation of network learning and the global distribution characteristics of the structural and semantic features in images, leading to spatial and spectral distortions. To address these issues and achieve a more comprehensive information representation, we propose a feature fusion method based on a hypergraph neural network and convolutional neural network (HGCNet). Specifically, we constructed a three-layer hypergraph based on texture features (T), spectral clustering (S), and positional relationships (P), called TSP, and designed a hypergraph attention module (HGAM). The constructed hypergraph imposes a unique learning tendency on the network, enabling it to exploit the feature spaces specific to the HSI SR task better. In addition, we constructed a multi-order spatial interaction module (MSIM) to realize cross-guidance and fusion of different types of features in the network. To the best of our knowledge, this is the first application of hypergraph neural networks for solving an HSI SR task. An extensive comparison of the experimental results on three public datasets demonstrated that HG-CNet performed better than other cuttingedge models.

The joint semantic segmentation of multimodal remote sensing (RS) images can make up for the problem of insufficient features of single-modal images and effectively improve classification accuracy. Some deep learning methods have achieved good performance, but they face problems such as complex network structure, large number of parameters, and deployment difficulty. In this letter, more attention is paid to front-end and branch-level feature transformation to obtain multiscale semantic information. The multiscale dilated extraction module (MDEM) is constructed to mine the specific features of different modalities. The multimodal complementary attention module (MCAM) is designed for further acquiring prominent complementary content. The concatenated features are transmitted and reused by the dense convolution to complete the encoding. Ultimately, a general and concise end-to-end model is proposed. Comparative experiments are carried out on three heterogeneous datasets, and the model put forward performs well in qualitative analysis, quantitative comparison, and visual effect. Meanwhile, the dexterity and practicability of the model are more prominent, which can provide support for lightweight design and hardware deployment.

The complementary characteristics of SAR and optical images are beneficial in improving the accuracy of land cover classification. Deep learning-based models have achieved some notable results. However, how to effectively extract and fuse the unique features of multi-modal images for pixel-level classification remains challenging. In this article, a two-branch supervised semantic segmentation framework without any pretrained backbone is proposed. Specifically, a novel symmetric attention module is designed with improved strip pooling. The multiple long receptive fields can better perceive irregular objects and obtain more anisotropic contextual information. Meanwhile, to solve the semantic absence and inconsistency of different modalities, we construct a multi-scale fusion module, which is composed of atrous spatial pyramid pooling, varisized convolutions and skip connections. A joint loss function is introduced to constrain the backpropagation and reduce the impact of class imbalance. Validation experiments were implemented on the DFC2020 and WHU-OPT-SAR datasets. The proposed model achieved the best quantitative values on the metrics of OA, Kappa and mIoU, and its class accuracy was also excellent. It is worth mentioning that the number of parameters and the computational complexity of the method are relatively low. The adaptability of the model was verified on RGB–thermal segmentation task.

Domain Generalization (DG) has led to remarkable achievements in cross-scene Hyperspectral Image (HSI) classification. Inspired by contrastive language-image pre-training (CLIP), the language-aware domain generalization method has been explored for cross-scene HSI classification with language prior knowledge. However, existing methods face some challenges: 1) The weak capacity to extract long-range contextual information and inter-class correlation. 2) Due to the inadequacies of the special pre-training on HSI data, the spatial-spectral features of HSI and linguistic features can not be straightforwardly alignment. To tackle those dilemmas, a novel network has been proposed with a CLIP framework, which consists of an image encoder, based on an encoder-only transformer to obtain the global contextual information and inter-class correlation, a frozen text encoder, and a cross-attention mechanism, named Linguistic-Interact-with-Visual Engager (LIVE), enhances the interaction between two modalities. Extensive experiments demonstrating superior performance over state-of-the-art methods with a CLIP framework, with 83.39% and 83.94% in OA, on the UH dataset and Pavia dataset, respectively.

Graph convolutional networks (GCNs) can extract features of samples in non-Euclidean space, which can be used for hyperspectral image (HSI) classification in collaboration with convolutional neural networks (CNNs). The features of GCNs and CNNs are incompatible to a certain extent, and traditional graph convolution methods use a single-channel 2-D matrix to extract features. As a result, it is difficult to explore the relationships fully and flexibly between samples. To further exploit the potential of these two networks for collaborative extraction of HSI features, we propose a fusion framework based on multichannel GCNs and class-guided attention mechanism (MG2A). Specifically, a multichannel GCN (MGCN) module is designed for batchwise network training, where the adjacent matrix of each channel contains different information between samples. In addition, we develop a class-guided attention mechanism to adaptively fuse the features of multiple MGCN modules and learn the transformation process of the features. Finally, the features of CNNs and MGCN modules are fused at multiple layers through a fusion framework. Experimental results on four benchmark HSI datasets show that MG2A achieves better classification performance compared to other state-of-the-art methods.

Deep learning has made significant achievements in remote sensing object detection tasks. However, small weak objects located in complex scenes are still not effectively addressed. The lack of feature information and negligible contributions during the optimization stage are the main reasons. To solve the indicated issues, a novel remote-sensing object detection method is proposed in this letter. Firstly, the context feature integration module (CFIM) is designed to extract implicit clues co-occurring with the object to compensate for the lack of features in small weak objects. The receptive field expansive deformable convolution (RFConv) constructed in CFIM can adaptively adjust the information extraction range based on the object’s characteristics, thereby capturing suitable context features. Secondly, to make small weak objects competitive during the optimization process, we propose tailored optimization functions: the balanced sampling strategy (BSS) and the modulated loss function (MLF). BSS makes up for the sample deficiency of small weak objects by balancing sampling. More specifically, BSS dynamically mines potential positive samples from the ignored set in order to increase the chances of matching. MLF progressively adjusts the loss proportion of the object and pushes the detector to be more sensitive to small weak objects during the training stage. Our proposed method achieves 94.79% and 71.23% mAP on the NWPU VHR-10 and DIOR datasets, which also proves the effectiveness.

Hyperspectral image (HSI) super-resolution (SR) technology has made remarkable progress due to the development of deep learning (DL). However, the technique still faces two challenges, i.e., the imbalance between spectral and spatial information extraction, and the parameter deviation and high computational effort associated with 3-D convolution. In this article, we propose an SR method for remote sensing HSIs based on multidomain spatial information and multiscale spectral information fusion (MSSR). Specifically, inspired by the high degree of self-similarity of remote sensing HSIs, a spatial-spectral attention module based on dilated convolution (DSSA) for capturing global spatial information is proposed. The extraction of local spatial information is then accomplished by residual blocks using small-size convolution kernels. Meanwhile, we propose the 3-D inception module to efficiently mine multiscale spectral information. The module only retains the scale of the 3-D convolution kernel in spectral dimension, which greatly reduces the high computational cost caused by 3-D convolution. Comparative experimental results on four benchmark datasets demonstrate that compared with the current cutting-edge models, our method achieves state-of-the-art (SOTA) results and the model computation is greatly reduced.