Heterogeneous change detection (HeCD) is a highly valuable yet challenging task in remote sensing. However, existing HeCD methods primarily focus on well-registered images, without considering unregistered heterogeneous image pairs, which are more common in real-world applications. In this paper, we first analyze why unregistered images significantly complicate the HeCD problem: they not only cause boundary errors in change detection but also undermine the image transformation process required for making heterogeneous images comparable. These effects are further amplified by the intrinsic complexity of HeCD. In light of this, we propose an unsupervised iterative global mapping-local searching method (IGmLs) for HeCD subject to registration errors. Specifically, IGmLs utilizes the global mapping to transform images into a common structural space to enable the comparison of heterogeneous images, and uses the local searching to reduce the direct influence of registration errors on change metrics, which is based on the analysis that misregistration would cause an unwanted increment on the change metric in unchanged region. Then, IGmLs builds a Markov random field (MRF) model to combine the global mapping and local searching processes, which enhances the robustness to misregistrations by considering spatial correlations. Finally, an iterative framework is employed to backpropagate the matching and changing results to refine the global mapping and local searching processes, which further eliminates the indirect influence of changes and misregistration on the image transformation and change metrics. Extensive experiments on five datasets have demonstrated the effectiveness of the proposed IGmLs. The codes will be released at https://github.com/yulisun/IGmLs .

Heterogeneous change detection (HeCD) enables the identification of land-cover changes using remote sensing imagery obtained from different sensors. Most existing methods overly emphasize modality transformation and shared feature extraction to bridge the gap between heterogeneous images. While these strategies facilitate comparable representations, they tend to neglect the intrinsic characteristics of the changes themselves, which limits their effectiveness in complex scenarios. To overcome this limitation, we propose a change prior-guided image transformation model (CPIT) for unsupervised HeCD. Specifically, starting from the definition of change detection, we analyze the connections among pairwise object relationships, change labels, and change semantics, and then derive change semantic consistency and inconsistency rules solely from the inherent nature of the change detection problem, without relying on data-specific assumptions. These rules are subsequently encoded as change semantic consistency and inconsistency constraints, which, from the perspective of graph signal processing, correspond to low-pass and high-pass spectral properties of the change signals. Finally, by integrating these semantic constraints with sparsity priors and image transformation constraints, we formulate a more precise transformation model for HeCD. Solving this model produces change detection results that conform to the change priors, thereby improving the detection performance. The derivation, formulation, and utilization of change priors in this work offer valuable insights for broader change detection research. Extensive experiments on five datasets validate the effectiveness of CPIT. The code will be released at https://github.com/yulisun/CPIT.

Significance Optical satellite remote sensing imagery offers advantages such as wide coverage,efficient information acquisition,and the ability to operate unrestricted by national borders.It plays an irreplaceable role in Earth observation applications and is widely used in both civilian and military fields,including environmental monitoring,disaster assessment,and intelligence reconnaissance.Remote sensing image target recognition refers to the technology of automatically locating and identifying specific targets,such as ships,aircraft,and vehicles,within remote sensing images.Driven by advances in deep learning technologies and practical application demands,target recognition in remote sensing imagery has evolved from traditional methods reliant on handcrafted features to data-driven deep learning models,achieving significant improvements in both performance and efficiency.Furthermore,in terms of recognition granularity,the focus has shifted from coarse-grained identification of broad target categories to fine-grained recognition capable of accurately distinguishing subcategories within the same general class,which better aligns with human interpretation needs.Overall,remote sensing image target recognition is steadily advancing towards greater intelligence and refinement. As a high-value target in remote sensing image interpretation,aircraft detection and identification are of great significance in both civil aviation management and military reconnaissance applications.In the field of civil aviation management,it helps to improve airport operational efficiency and safety management.For example,by accurately identifying and counting different types of aircraft on the tarmac,intelligent flight scheduling and rational allocation of jet bridge resources can be achieved.In the field of military reconnaissance,this technology can support the monitoring of the types,numbers,and deployment dynamics of aircraft in enemy airfields,thereby gaining insight into their combat readiness and operational intentions.During conflicts,it can also provide key target designations to fire units,enabling precision strikes against high-value targets,such as early warning aircraft and reconnaissance aircraft.In summary,fine-grained aircraft recognition is a crucial driving force for the automation and intelligence of civil aviation and national defense. Developing various technologies for fine-grained aircraft recognition in optical remote sensing images is crucial.Currently,there are two primary technical approaches:fine-grained image recognition and fine-grained object detection.The former treats aircraft detection and fine-grained recognition as two independent steps.Specifically,aircraft targets are first detected in the images through manual or automated learning-based algorithms to obtain target image patches.Then,fine-grained recognition algorithms are designed to classify the aircraft types within these patches.The latter integrates fine-grained aircraft recognition directly into the object detection process,achieving a unified framework for detection and recognition.Comparatively,the latter offers higher accuracy and efficiency,making it more practical for real-world applications. Over the past few decades,researchers have proposed numerous algorithms based on the two technical routes mentioned above to address challenges in the field of fine-grained aircraft recognition,ranging from traditional manually designed features to deep learning-based approaches.Although significant progress has been made,a series of challenges remain unresolved.Hence,it is particularly important and necessary to systematically review existing research to more reasonably guide the future development of this field. Progress This paper provides a systematic review of the research progress in the field of fine-grained aircraft recognition in optical remote sensing images.First,the definition and research value of fine-grained aircraft recognition,along with the associated challenges,are introduced.These challenges include difficulties in data acquisition and annotation,class imbalance,intra-class diversity and inter-class similarity,multi-scale and morphologies diversity,and the difficulty in capturing subtle differences.Then,the current state of research,both domestically and internationally,is organized and summarized according to two technical approaches:fine-grained image recognition and fine-grained object detection.Specifically,fine-grained image recognition algorithms are further divided into algorithms based on handcrafted features and methods based on deep learning.The latter encompasses multiple branches such as traditional and deep feature fusion,deep neural networks combined with template matching,data augmentation-based,pre-trained model fine-tuning,deep feature fusion-based,part feature learning-based,discriminative feature learning-based,few-shot learning-based,and other approaches.The advantages and limitations of various algorithms within each branch are compared and analyzed.For fine-grained object detection algorithms,two major technical approaches,two-stage and single-stage approaches,are summarized,and the advantages and disadvantages of different algorithms within each category are compared.In addition,publicly available fine-grained aircraft recognition datasets for optical remote sensing images in recent years are compiled,along with commonly used accuracy evaluation metrics.Finally,future development trends in this field are discussed,including building standardized and high-quality datasets,robust recognition in open-world environments,development of fine-grained recognition foundation models in the field of remote sensing,multimodal data fusion for aircraft recognition,and research on model lightweighting and edge deployment. Conclusions and Prospects Deep learning technologies have significantly improved the performance of fine-grained aircraft recognition in optical remote sensing images.However,to achieve reliable application in complex real-world scenarios,further exploration at both the theoretical and practical levels is still needed.This will facilitate breakthroughs in key technologies and promote the practical application and sustainable development of this field.

Cross-sensor Synthetic Aperture Radar (SAR) target detection suffers from severe performance degradation due to the inconsistency of sensors between training and testing data. Structural characteristics of SAR targets, as important intrinsic properties, play a crucial role in robust detection. However, variations in sensor types often cause inconsistencies in structural details, leading to distribution differences and domain shifts. To address these challenges, we propose a cross-sensor SAR target detection method based on semantic scattering graph structure alignment. First, a semantic scattering graph is constructed from sampled scattering points to characterize the target’s intrinsic structure. Meanwhile, the semantic node associations are enhanced with cross-domain statistical distributions to enrich node prior information and a graph convolutional network to strengthen contextual awareness. Then, a hierarchical structure alignment mechanism is introduced to calibrate structural consistency across domains. This includes the cross-domain perceptual interaction to narrow the semantic distribution differences. Furthermore, the hierarchical alignment is operated at both the local node level and the global structure level, achieving the alignment of target structural consistency. Finally, the experiments conducted on cross-sensor tasks demonstrate that our method significantly outperforms the state-of-the-art methods, achieving improvements of 5%-40% in mAP and F1-score, which highlights the effectiveness of the proposed approach.

RGB-infrared (RGB-IR) vehicle detection in UAV imagery is essential for applications such as traffic monitoring and object tracking. However, existing methods often suffer from heterogeneous feature responses across modalities, degraded RGB feature representations under adverse illumination, and insufficient capture of fine-grained structural and edge details, which collectively impede accurate cross-modal modeling and localization. To alleviate these issues, we propose an Illumination-aware Semantic-Guided Mamba (ISGM) network. First, we design a Feature Representation Refinement (FRR) module that stabilizes RGB and IR features through scale-specific channel remapping and normalized nonlinear refinement, yielding more reliable representations for subsequent cross-modal interaction. Furthermore, to enhance robustness to illumination variations and better preserve detailed structural cues and boundary information, we develop a cross-modal feature interaction mechanism comprising the Illumination-Aware Fusion Modulation (IAFM) module and the Detail-enhanced Semantic-Guided Mamba (DSGM) module. Specifically, the IAFM module estimates illumination-aware modality reliability weight maps, thereby improving robustness under challenging illumination conditions. These weight maps guide the DSGM module to integrate high-level semantic information and low-level detail cues into multi-scale guidance features. These features are subsequently used to modulate the scanning parameters for adaptive RGB-IR feature interaction. This design improves the modeling of target-region features while minimizing interference from complex backgrounds. Extensive experiments on the DroneVehicle and VEDAI datasets demonstrate that ISGM outperforms state-of-the-art methods in detection performance while achieving a favorable accuracy-efficiency balance among comparable methods.

Heterogeneous change detection (HeCD) is a significant yet challenging task in remote sensing. Most existing methods rely on well-registered images, making them vulnerable in real-world scenarios. To address this issue, we propose a unified model that jointly integrates graph-based image transformation and optical flow-based image registration, named GiToF, for HeCD task in the presence of co-registration errors. Unlike conventional approaches that separately perform registration and change detection, GiToF formulates them into a collaborative framework and employs three tailored constraints to guide the unified model: the structural consistency for image transformation of unchanged image, the Lucas-Kanade optical flow constraint for image registration of matched image, and the prior sparsity for changed image. This unified model allows GiToF to mitigate the impact of co-registration errors on image transformation and change estimation via optical flow-based matching, while image transformation and change estimation ensure that the brightness constancy and smooth motion assumptions of optical flow are satisfied, which are typically violated due to image heterogeneity and drastic changes in the HeCD problem. Therefore, GiToF provides a more robust HeCD solution by seamlessly coupling image transformation, registration, and change estimation. Extensive experiments on five benchmark datasets demonstrate that GiToF achieves superior performance and robustness compared with state-of-the-art methods under co-registration errors.

Adversarial purification with diffusion models has emerged as a promising defense strategy. While existing frequency-based approaches better preserve global semantics, they overlook the distinct temporal dynamics of low-and high-frequency components during the reverse process and lack adaptive correction mechanisms. Meanwhile, Diffusion Transformers (DiTs) offer stronger multi-condition alignment and superior scalability, but remain underexplored for adversarial purification, primarily due to prohibitive inference cost. Motivated by these challenges, we propose FEAP-DiT, a frequency-decoupled efficient adversarial purification framework built upon DiTs. FEAP-DiT designs complementary processing schemes for low-and high-frequency features and unifies them through frequency-aware decoupling and fusion. Specifically, low-frequency features are sparsely updated and largely reused with mild selective correction to preserve global semantics. High-frequency features, where adversarial perturbations concentrate, are updated more frequently to suppress perturbations and approximated at intermediate steps via a Hermite-based sequential predictor due to their short-term temporal continuity. To further mitigate prediction deviations in high-frequency components, FEAP-DiT introduces an efficient control-aware refinement mechanism that fuses structural constraints into the DiT backbone. Control blocks are inserted only at selective layers, and control computation is restricted to high-energy tokens over informative regions. Extensive experiments on CIFAR-10 and ImageNet-1K demonstrate the effectiveness of FEAP-DiT, achieving the largest average gains of 2.43% on CIFAR-10 and 3.40% on ImageNet-1K under white-box attacks. Moreover, FEAP-DiT reduces inference latency compared with existing multi-step diffusion-based adversarial purification baselines, while achieving a 15.61% reduction in computational complexity relative to the representative DiT architecture PixArt-alpha.

Out-of-distribution (OOD) detection is essential for reliable automatic target recognition (ATR) in open-world scenarios. Although widely studied in optical imagery, research on OOD detection in synthetic aperture radar (SAR) images remains limited. SAR imagery poses unique challenges due to high intra-class variability and low inter-class separability, which existing distance-based methods with simplistic class representations fail to handle. To address these issues, we propose two dynamic multi-prototype representation methods: Dynamic Multi-Prototype Clustering (DMPC) and Uncertainty-Aware Dynamic Multi-Prototype Learning (DMPL). DMPC is a post-hoc method that applies hierarchical clustering to assign adaptive prototypes per class, enabling accurate distance-based OOD scoring without retraining. In contrast, DMPL is an end-to-end framework that enhances discrimination in SAR imagery by improving intra-class compactness and inter-class separability. It further incorporates an uncertainty-aware filter that updates prototypes only with low intra-class uncertainty samples, ensuring stable and diverse representations. Extensive experiments on the evaluated SAR-OOD benchmark show that the proposed methods achieve strong average OOD detection performance compared with representative post-hoc and training-based baselines, demonstrating the effectiveness of dynamic multi-prototype representation for SAR OOD detection. Our code is publicly available at https://github.com/Xiaoyan-Zhou/DMPC-DMPL.

Multiclass Object Change Detection (MCOCD) has emerged as a pivotal task in remote sensing image analysis, detecting the locations, categories, and change types (disappeared, unchanged, appeared) of multiclass, time-sensitive objects across bitemporal imagery. Existing end-to-end architectures are primarily designed for conventional change detection tasks that ignore unchanged objects, which limits their applicability to MCOCD. Typical MCOCD approaches are designed in a non-end-to-end manner, which leads to error accumulation and susceptibility to pseudo-changes. To address these limitations, this study presents the Feature Interaction and Fusion Network (FIF-Net), an end-to-end framework that jointly optimizes object detection and change analysis with two novel modules: the Cross-Agent Attention Module (CAAM) to effectively suppress pseudo-changes and the Sum-Difference Feature Fusion Module (SDFFM) for joint optimization of features for both changed and unchanged objects. Experimental results in the Aircraft Change Detection dataset (ACD)-v1.5 dataset demonstrate that FIF-Net achieves state-of-the-art performance of 67.6% mAP, significantly outperforming existing approaches such as Dual Correlation Attention-guided Detector (DCA-Det) and Temporal Mutual Attention and Contextual Network (TMACNet), and ablation studies further validate the efficacy of each core module.

Object tracking in multi-drone videos shows great potential in disaster emergency response, scene reconstruction and low-altitude economy due to its advantages of wide-area coverage and comprehensive perception. Most existing trackers adopt a decoupled framework, where each drone independently performs multi-stage feature extraction before multi-drone interaction. However, this paradigm suffers from inherent limitations, including task conflicts, mutual feature contamination and inconsistent target representations. To overcome above issues, this paper proposes JointTrack, a temporal-spatial prompt learning-based framework densely joints feature extraction and interaction in an end-to-end manner. Unlike prior methods, JointTrack incorporates multi-drone interaction at every feature extraction stage to achieve collaborative optimization. Specifically, a temporal prompt encoder (TPE) is proposed to capture temporal prompts from historical frames, guiding the network to disentangle target-related dynamics from background noise and thereby suppress feature contamination caused by occlusions or cluttered scenes. Meanwhile, a spatial prompt encoder (SPE) constructs scale-aware and component-aware prompts by integrating multi-drone templates, enabling feature complementarity that alleviate inconsistent feature representations under varying viewpoints. Additionally, a dynamic gated decoder (DGD) adaptively fuse these temporal-spatial prompts into original features through a gating mechanism with contextual dependency modeling. Extensive experiments on the MDOT benchmark demonstrate that JointTrack outperforms previous state-of-the-art trackers by 4.1% and 3.0% in success and precision. Moreover, JointTrack maintains competitive tracking speed and exhibits strong generalizability on the MDMT and UAV123 datasets, as well as in real-world scenarios. Codes will be released at https: //github.com/HanJin0326/JointTrack.

Cross-sensor Synthetic Aperture Radar (SAR) target detection faces significant domain shift challenges. In practical applications, source data is often inaccessible due to privacy, storage, and transmission constraints, making Source-Free Domain Adaptation (SFDA) an ideal solution. However, directly applying conventional SFDA methods to SAR images leads to severe performance degradation. Specifically, the strong clutter and discrete scattering characteristics of SAR targets induce semantic space drift, while severe geometric distortions destroy continuous visual contours, thereby leading to bounding box drift and degraded localization accuracy. To overcome these challenges, we propose a novel SFDA framework for cross-sensor SAR target detection by deeply mining the intrinsic physical invariants of SAR targets in the deep feature space. First, we design a Semantic Prototype Anchoring Calibration (SPAC) module to calibrate stable semantic space center. By leveraging inherent source model prototypes as semantic anchors, it effectively guides target domain features to cluster correctly under strong clutter interference, thereby calibrating cross-domain semantic drift. Second, we introduce the Foreground-Aware Manifold Constraint (FAMC) module to regulate regression scattering distributions. By leveraging spatial centerness priors and mapping intrinsic foreground structures into a manifold space, FAMC jointly regularizes spatial centroids and inter-channel geometries. This precisely aligns feature distributions, effectively mitigating localization degradation caused by geometric deformations. Experiments on cross-sensor tasks demonstrate that our framework reaches improvements of about 10%–25% in mAP and F1 score without accessing any source data, significantly enhancing the model's generalization capability in unknown scenarios.

Multimodal change detection (MCD) is a topic of increasing interest in remote sensing. Due to different imaging mechanisms, the multimodal images cannot be directly compared to detect the changes. In this article, we explore the topological structure of multimodal images and construct the links between class relationships (same/different) and change labels (changed/unchanged) of pairwise superpixels, which are imaging modality-invariant. With these links, we formulate the MCD problem within a mathematical framework termed the locality-preserving energy model (LPEM), which is used to maintain the local consistency constraints embedded in the links: the structure consistency based on feature similarity and the label consistency based on spatial continuity. Because the foundation of LPEM, i.e., the links, is intuitively explainable and universal, the proposed method is very robust across different MCD situations. Noteworthy, LPEM is built directly on the label of each superpixel, so it is a paradigm that outputs the change map (CM) directly without the need to generate intermediate difference image (DI) as most previous algorithms have done. Experiments on different real datasets demonstrate the effectiveness of the proposed method. Source code of the proposed method is made available at https://github.com/yulisun/LPEM.

Fine-grained matching between Unmanned Aerial Vehicle (UAV) imagery and satellite imagery is a critical step for achieving UAV self-localization. Current fine-grained matching methods typically compare each UAV-captured image in turn against cropped patches from a large-scale satellite reference map to find the most similar region, then estimate the UAV's position via point correspondences. However, when the UAV lacks precise initial location information and the satellite map covers a large area, this approach consumes significant computational resources and time due to the need for large-scale matching to retrieve similar patches. Moreover, existing methods usually select only a single best-matching patch for subsequent finegrained matching, which limits both accuracy and robustness. To address these challenges, we have adapted the UAV-VisLoc dataset to suit our task requirements. Using the dense matching algorithm LoFTR as our baseline, we have developed a one-tomany fine-grained matching pipeline specifically for UAV selflocalization. Experimental results demonstrate the effectiveness of our proposed one-to-many fine-grained matching framework.

Accurate geolocation of high-altitude airborne synthetic aperture radar (HAA-SAR) imagery is pivotal for applications in Earth observation and remote sensing. However, errors in tropospheric delay modeling can severely affect positioning accuracy. Despite numerous existing compensation approaches, these methods inadequately address the complex atmospheric spatiotemporal variations encountered in HAA-SAR systems. This study presents the first comprehensive investigation of spatiotemporal variations of tropospheric delay in HAA-SAR systems, revealing their impact on positioning accuracy. To address these challenges, a spatiotemporal adaptive tropospheric delay compensation (SATDC) approach is introduced. The proposed method integrates high-resolution ERA5 meteorological data with ray-tracing techniques, while employing a cross-iterative algorithm to simultaneously optimize target height and delay estimates. The method was validated through flight experiments using an X-band HAA-SAR system across diverse atmospheric conditions and terrains. Experimental results show that the proposed method improves positioning accuracy compared to existing approaches. Specifically, the SATDC method achieves an average range position mean absolute error (MAE) of 2.24 m, representing a reduction of 33.12% outperforming our previous tropospheric delay compensation work tropospheric delay compensated range-Doppler (TDC-RD). Furthermore, the implementation of spatiotemporally adaptive compensation reduces the slant range measurement MAE from 12.94 m (without compensation) and 2.97 m (with conventional compensation) to 1.71 m. These findings highlight the method's effectiveness in compensating for tropospheric delay, enhancing the precision and reliability of HAA-SAR geolocation in challenging environments.

Optical-SAR image matching is a fundamental task for remote sensing applications. While existing methods perform well on some popular datasets such as SEN1-2 and WHU-SEN-City, their generalizability across diverse data sources such as satellites, spatial resolutions, and scenes remains insufficiently investigated, hindering the practical implementation of optical-SAR matching in various downstream tasks. Thus, 3MOS, the first multi-source, multi-resolution, and multi-scene optical-SAR dataset, was proposed in our study to address this gap. This dataset consists of 113k optical-SAR image pairs, with the SAR data collected from five satellites and resolutions ranging from 3.5 m to 12.5 m, further categorized into eight scenes, such as urban, rural, and plains through a simple but practical classification strategy. Based on this dataset, the performance of optical-SAR matching methods was evaluated through the data with diverse characteristics. Additionally, extensive experiments were conducted, and the following two findings were obtained. 1) None of the state-of-the-art methods achieved consistently superior performance across different sources, resolutions, and scenes, specifying significant generalization challenges for diverse downstream task data. 2) Training data distribution significantly impacted the matching performance of deep-learning models, highlighting the domain adaptation challenge in optical-SAR image matching. Furthermore, the practical utility of the dataset was comprehensively validated through multimodal change detection experiments, demonstrating its substantial value for a wide range of downstream applications.

Spiking neural networks (SNNs) are the third generation of neural networks that offer the advantages of low computational requirements, fast inference speed, and strong biological interpretability. This makes SNNs suitable for synthetic aperture radar (SAR) target recognition tasks, which are often constrained by limited computational power. This letter proposes SAR-TinySNN, a lightweight SNN architecture designed for SAR target recognition. Unlike existing SAR-related studies that predominantly rely on rate coding, SAR-TinySNN uses direct coding to encode SAR images, allowing for a more efficient coding method adapted to SAR images and achieving high target recognition accuracy, especially in scenarios with limited training samples. By integrating direct coding into a trainable SNN framework, SAR-TinySNN achieves competitive performance compared with traditional deep neural networks (DNNs) and deep SNNs on vehicle, aircraft, and ship SAR target recognition datasets, with faster inference times. The experimental results demonstrate the effectiveness of SAR-TinySNN for SAR target recognition.

Matching drone images to satellite reference images is a critical step for achieving UAV self-localization. Existing drone visual localization datasets mainly focus on target localization, where each drone image is paired with a corresponding satellite image slice, typically with identical coverage. However, this one-to-one approach does not reflect real-world UAV self-localization needs as it cannot guarantee exact matches between drone images and satellite tiles nor reliably identify the correct satellite slice. To bridge this gap, we propose a one-to-many matching method between drone images and satellite reference tiles. First, we enhance the UAV-VisLoc dataset, making it the first in the field tailored for one-to-many imperfect matching in UAV self-localization. Second, we introduce a novel loss function, Incomp-NPair Loss, which better reflects real-world imperfect matching scenarios than traditional methods. Finally, to address challenges such as limited dataset size, training instability, and large-scale differences between drone images and satellite tiles, we adopt a Vision Transformer (ViT) baseline and integrate CNN-extracted features into its patch embedding layer.

Heterogeneous change detection (HeCD) is a highly valuable yet challenging task in remote sensing. To enable the comparison of heterogeneous images with different imaging mechanisms, some structural consistency-based image transformation methods have been proposed, which utilize graph models to represent image structures and constrain the transformed images and original images to have the same structural characteristics on the graph model. Consequently, these graph-based methods face two challenges: adequately characterizing the image structure and effectively utilizing the change information. To address these challenges, this article proposes a signed graph-based image transformation (SGIT) method for unsupervised HeCD. First, we analyze the limitations of previous unsigned graph-based methods in capturing the image structure, which leads to the failure to detect changes in some scenes. In light of this, we construct signed graph models that utilize positive/negative weights to represent the similarity/dissimilarity relationships within the image, respectively, and employ adaptive weighting, negative sampling, and neighborhood expansion strategies to bolster the structure representation capability of signed graphs. Second, we analyze how the change would induce a bimodal distribution of vertex feature distances in original and transformed images. Subsequently, a distribution-induced reweighted graph Laplacian regularization (RGLR) is proposed to exploit this prior change information. Finally, a more accuracy image transformation model is obtained by incorporating three types of constraints: signed graph-based structural consistency term, bimodal distribution-induced RGLR, and change sparsity-based penalty term. Extensive comparative experiments on five real datasets have demonstrated the effectiveness of the proposed SGIT.

Heterogeneous change detection (HCD) is crucial for monitoring surface changes using various remote sensing data, especially in disaster emergency response and environmental monitoring. To facilitate the comparability of heterogeneous images, previous methods are devoted to designing various complex transformation functions to transfer heterogeneous images into a common domain for comparison. As a result, the performance of HCD is constrained by the accuracy and robustness of these transformation functions. Unlike existing comparison-based HCD methods that rely on complex transformations and feature alignments between heterogeneous images, this paper proposes an unsupervised rules-induced energy model (RIEM) that detects changes by independently analyzing intra-image relationships, without explicitly comparing the heterogeneous images. This frees HCD from the complicated and challenging transformations and interactions between heterogeneous images. Specifically, we first establish the connections between the class relationships (same/different) and change labels (changed/unchanged) of pairwise superpixels, and then derive six rules for determining the change label of each superpixel, which enables detecting changes by considering only the intra-image relationships within each image, without inter-image comparisons. Then, we build an energy-based model to release the ability of rules to identify changes, which implements four types of energy loss functions. Remarkably, since the rules used in the energy model are derived based on the nature of change detection problem, the proposed RIEM is highly robust to imaging conditions. Extensive experiments on seven datasets demonstrate the efficacy of RIEM in detecting changes from heterogeneous images. The code is released at https://github.com/yulisun/RIEM.