Jon Atli Benediktsson

Land cover change detection (LCCD) with heterogeneous remote sensing images (Hete_RSIs) is important for disaster rescue applications. However, Hete_RSIs are acquired by different sensors, and different imaging modalities mean that bitemporal Hete_RSIs cannot be compared directly for LCCD. In this paper, an unsupervised global-local graph feature learning network (GLGFLN) is proposed to learn the shared global and local features from bitemporal Hete_RSIs for LCCD. First, global-local graphs are constructed by considering contextual information to describe the global and local structural features. Then, graph convolution coupled with cross-attention is designed to learn the deeply shared structural features and align the learned structural features from global and local perspectives. Finally, a fusion strategy based on frequencies and empirical weights is developed to fuse the global change magnitude feature and local change magnitude into the final change magnitude map. In the experiments, compared with seven state-of-the-art methods on four pairwise real Hete_RSIs, the results effectively demonstrate the superior performance of the proposed approach. For example, the improvements on the four datasets are 3.22%, 5.56%, 3.58%, and 3.79% in terms of overall accuracy. The source code of the proposed network is available at https://github.com/ImgSciGroup/2025-GLGFLN.

Although subsurface-reflectance-based Satellite-Derived Bathymetry (SDB) effectively mitigates sun-glitter effects, current methods struggle to disentangle subtle bathymetric features from complex background environmental noise, severely limiting their generalization across diverse waters. This limitation fundamentally stems from the inability of existing architectures to simultaneously capture high-frequency local textures and long-range semantic dependencies required for precise depth inference in degraded optical environments. To address this challenge, SDB-Mamba, a Cross-Scale Collaborative Visual State Space Network, is presented for robust bathymetric retrieval from multispectral imagery. First, a Dual-Encoder Mamba-ResNet architecture is developed to synergize the linear complexity of Visual State Space (VSS) models for modeling global contextual dependencies with residual connections for preserving fine-grained local details. Second, to enhance feature discriminability, a Spatial Attention Block (SAB) is integrated to suppress environmental noise in spatially heterogeneous regions, while a Channel Attention Block (CAB) is employed to adaptively recalibrates channel weights based on local optical water characteristics. SDB-Mamba was rigorously evaluated across six diverse study areas with distinct optical properties and bottom types, including unseen domains for external testing. Experimental results demonstrate that the proposed model achieves Mean Absolute Errors (MAE) of 0.36-0.64 m, reducing root mean square error (RMSE) by up to 86% compared to SwinUNet and improving Peak Signal-to-Noise Ratio (PSNR) by 66% relative to TransNet. Ablation studies confirm the critical role of the VSS, SAB, and CAB components, which improved Structural Similarity Index Measure (SSIM) by 25%, 68%, and 62%, respectively, establishing SDB-Mamba as a significant advancement in robust, high-precision nearshore bathymetry.

This paper presents a nonlinear discriminant analysis (NDA) method based on an orthogonal rational function (ORF) system for the classification of hyperspectral images (HSI). The proposed method, abbreviated ORF-NDA, is formulated from the standpoint of function theory. It can make good use of the spectral curve characteristics of HSIs and establish a methodology more suitable for HSI processing in the function space. This is a new paradigm that is different from mainstream hyperspectral data processing methods. First, a functional data fitting model based on an adaptive ORF system is designed for the reconstruction of the spectral curve for each pixel in HSIs. Second, the stochastic adaptive Fourier decomposition (SAFD) method is used to find the optimal parameters for building the adaptive ORF system. Third, based on the reconstructed spectral curve, an NDA method based on an ORF system is built for feature extraction of HSIs. Finally, a spectral-spatial classification framework is designed for HSIs in scenarios with limited sample sizes. Experimental results on five public HSIs demonstrate that the proposed ORF-NDA method outperforms several state-of-the-art deep learning methods in classification accuracy.

Hyperspectral images (HSIs) are often corrupted by various noise during acquisition, which leads to a reduction in HSI quality. Therefore, HSI denoising is a necessary preprocessing step for their further application. In this paper, we propose an HSI denoising framework based on the sparsity constraint and low-rank approximation (SCLA). SCLA simultaneously uses the global low-rank structure, sparse prior, and local spatial-spectral smoothness of HSIs to construct a denoising model for removal of heterogeneous Gaussian noise, impulse noise, and stripe mixed noise. The Split Bregman method is used for SCLA to solve the optimization problem of our model. An ablation study is utilized to show the effectiveness of the structures of SCLA. Compared with other denoising methods, based on simulated and real experiments, SCLA achieves significant advantages in quantitative and visual assessments for removing Gaussian noise, impulse noise, and stripe mixed noise.

Autoencoder-based methods have gained prominence in Blind hyperspectral Unmixing (BHU), with most approaches based on predefined mixing assumptions, such as the linear mixing model. However, real-world scenes often involve complex nonlinear light-material interactions that are difficult to model explicitly, leading to systematic biases in endmember and abundance estimation when these assumptions deviate from the actual mixing mechanism. In this article, we extend conventional autoencoder-based BHU by jointly and adaptively learning the mixing mechanism directly from hyperspectral images while estimating endmembers and abundances, and propose a novel BHU framework termed UNmixing TRansformer (UNTR). Specifically, an adaptive mixing mechanism modeling module is designed to parameterize the mixing mechanism, in which a specially designed transformer fuses heterogeneous endmember and abundance information. Furthermore, we develop an end-member query mechanism integrated into the adaptive mixing mechanism modeling module under a distribution alignment constraint, enabling robust endmember estimation under random initialization. As the first attempt to adaptively model the mixing mechanism as data-driven network parameters, UNTR achieves remarkable accuracy in both abundance and endmember estimation. Without relying on predefined mixing assumptions and well-initialized endmembers, UNTR exhibits strong adaptability to complex real-world scenarios. Our code will be available at https://github.com/Preston-Dong/UNTR.

A innovative frequency-domain alignment method is proposed to address a feature shift issue that is encountered by the Segment Anything Model (SAM) when cross-domain segmentation on remote sensing images. Due to the inherent differences is performed in imaging mechanisms between natural and remote sensing images, as well as significant scale variation challenges, existing SAM-based pre-trained models often struggle to effectively perceive the semantic discrepancies between natural domain and remote sensing domain objects. To address this, a Cosine Similarity Gated Smoothing (CSGS) method is proposed, built upon features extracted by pre-trained SAM, incorporating a consistency constraint strategy. The proposed approach quantifies the different semantic expressions of features from different domains under the same target using cosine similarity, while refining the propagation of semantic information in 8 directions through a graph propagation approach. Furthermore, a Deformable Cross-Scale Attention Bridge (DCAB) is introduced, which adaptively samples low-level local features and performs weighted fusion with a temperature gating mechanism, significantly enhancing the model’s segmentation capability in fine-detail regions. Finally, to improve SAM’s ability in cross-domain feature learning for remote sensing images, we propose the Multi-Scale Frequency Domain Adapter (MSFA), which decouples and fuses low-frequency feature components. In experiments, the proposed approach achieved mean IoU (Intersection over Union) scores of 0.8972%, 0.9085%, and 0.5033% on three public datasets: Vaihingen, Potsdam, and LoveDA, respectively. Codes are available at https://github.com/CHINASCBZ/FreqAlign.

Currently available transformer models have achieved satisfactory results in hyperspectral image (HSI) classification. However, the existing transformer models face two key challenges: First, the collaborative modeling of spatial–spectral features remains inadequate, hindering the models’ ability to effectively extract and fuse complementary information between spatial patterns and spectral responses. Second, in practical scenarios with extremely limited labeled samples, complex deep models are highly susceptible to overfitting due to the severe mismatch between their massive capacity and limited training data. Significantly constraining its generalization capability and robustness. Therefore, developing classification algorithms that can achieve an excellent performance even under conditions of small sample sizes holds significant research value and is of practical importance. To remedy these drawbacks, a dual-branch token selection fusion network (DBTSFN) for HSI classification, is proposed in this article. Specifically, DBTSFN employs a multimodule collaborative mechanism through a spatial–spectral cascade strategy to perform deep mining of spatial and spectral features, respectively, within a dual-branch architecture. A gate-enhanced token selection block fully models spatial–spectral complementary information. Finally, through the classification decision fusion module, the overfitting issue of the model under small-sample conditions was effectively mitigated, significantly enhancing its generalization performance in scenarios with limited labeled data. Since each decision prediction may exhibit different overfitting tendencies, combining multiple predictions can mitigate overfitting to enhance classification accuracy and stability. The results show that the proposed method significantly outperforms state-of-the-art approaches on all considered datasets and demonstrates the effectiveness and superiority of the proposed model for HSI classification when few labeled samples are used.

The rapid expansion of multi-source satellite imagery is driving innovation in Earth observation, opening unprecedented opportunities for Remote Sensing Foundation Models to harness diverse data. However, many existing models remain constrained by fixed spatial resolutions and patch sizes, limiting their ability to fully exploit the heterogeneous spatial characteristics inherent in satellite imagery. To address these challenges, we propose FlexiMo, a flexible remote sensing foundation model that endows the pre-trained model with the flexibility to adapt to arbitrary spatial resolutions. Central to FlexiMo is a spatial resolution-aware module that employs a parameter-free alignment embedding mechanism to dynamically recalibrate patch embeddings based on the resolution and dimensions of the input images. This design not only preserves the geometric fidelity of tokenization under varying image sizes, resolutions, and patch granularities, but also enables efficient feature extraction without requiring modifications to the underlying network architecture. In addition, FlexiMo incorporates a lightweight channel adaptation module that leverages prior spectral information from sensors. This mechanism allows the model to process images with varying numbers of channels while maintaining the data’s intrinsic physical properties. Extensive experiments on diverse multimodal, multi-resolution, and multi-scale datasets demonstrate that FlexiMo significantly enhances model generalization and robustness. In particular, the proposed method achieves outstanding performance across a range of downstream tasks, including scene classification, land cover classification, urban building segmentation, and cloud detection. We also explicitly validate physical consistency through wavelength-channel permutation and wavelength-perturbation tests, showing that FlexiMo is sensitive to physically incorrect spectral metadata while remaining robust to small wavelength deviations. By enabling parameter-efficient and physically consistent adaptation, FlexiMo improves the practical deployability of RSFMs under real-world sensor heterogeneity and scale variations.