Building extraction from remote sensing imagery is essential for urban planning, population monitoring, and emergency response. However, traditional methods often struggle with accurate building delineation due to complex architectural features and varying imaging conditions. In this paper, we propose ABIBE, an Adaptive Building Information-Based Extraction framework that leverages Vision-Language Models (VLMs) for high-precision building extraction from remote sensing imagery. Our approach introduces two key innovations: (1) a hierarchical feature transfer mechanism that selectively extracts and adapts visual representations from the JanusPro vision-language model, and (2) a dynamic multi-scale feature fusion technique that adaptively integrates these features with spatial information through cross-modal attention mechanisms. Extensive experiments on the WHU Building Dataset demonstrate that our method significantly outperforms state-of-the-art approaches, achieving 91.39% IoU and 95.50% F1 score. We also provide comprehensive performance analysis including computational complexity, runtime efficiency, and resource requirements to evaluate the practical applicability of the proposed framework. The ABIBE framework provides a promising solution for high-precision building extraction from remote sensing imagery, with particular advantages in areas with complex architectural structures and high building density.

Few-shot building segmentation in remote sensing imagery remains challenging due to limited annotated data and complex geometric structures inherent in building footprints. Traditional prototype-based methods ignore rich geometric priors of buildings, leading to suboptimal feature representations and poor generalization across different geographic regions. We propose GAPL-SegNet, a novel architecture that integrates Geometry-Aware Prototype Learning with adaptive feature aggregation for superior few-shot performance. Our approach introduces several key innovations: (1) a geometric feature extractor that captures building-specific structural patterns including edges, corners, and rectangularity through dedicated detection branches; (2) a geometry-aware prototype learning framework that leverages spatial importance weighting for more discriminative prototype construction based on geometric significance; (3) a progressive adaptive training strategy with dynamic geometric loss weighting that ensures effective integration of geometric priors; and (4) systematic cross-dataset analysis quantifying domain adaptation challenges. Extensive experiments demonstrate strong performance across multiple datasets and scenarios. On the WHU Building dataset with 100 training samples, GAPL-SegNet achieves 85.75±0.43% IoU (95% CI: [85.22, 86.29]) based on five independent runs, with notable stability (CV < 0.5%), outperforming the best baseline by 3.67% IoU. Cross-dataset evaluations reveal significant challenges from resolution differences: WHU to Inria achieves 51.68% IoU (41.4% with resolution alignment) and Inria to WHU reaches 53.71% IoU (33.6% with resolution alignment). Multi-resolution analysis demonstrates that spatial resolution differences (0.3m to 1.2m) cause up to 46% IoU degradation, identifying resolution adaptation as the primary domain shift challenge that exceeds the scope of geometric priors alone. Computational efficiency analysis reveals that our method achieves a favorable performance-efficiency balance with 22.5M parameters and 5.3ms inference time. In building segmentation’s specific few-shot setting, geometric priors bring better boundary consistency and deployment efficiency compared to general-purpose foundation models.

Remote sensing spatiotemporal image fusion is a promising approach to acquire remote sensing data with high spatial and temporal resolution. While most deep neural network-based models have demonstrated high accuracy, they heavily depend on temporal information from the dataset, necessitating a pair of coarse and fine resolution image sets ($C_{1}-F_{1}$) taken near the prediction time. This reliance on temporal data makes these models challenging in practical applications. To tackle this challenge, this study introduces a time-insensitive spatiotemporal fusion model. An adaptive spatial distribution transformation (ASDT) module is proposed to enhance the spatiotemporal consistency of fine-resolution images by incorporating the spatial structure of coarse-resolution images. This module aims to improve the model’s performance in tasks less sensitive to time. Furthermore, an encoding-decoding structure is devised to address significant resolution variations in remote sensing images. The encoder uses a multilevel feature extraction (MLFE) module to capture features at multiple levels, reducing information loss and enhancing feature utilization. The decoder includes a cross-scale attention feature fusion and reconstruction (CSAFFIR) module to integrate features of different scales and semantic levels, thereby improving the overall performance of the fused image. Experimental findings from datasets collected by two satellite types demonstrate that the proposed HPLTS-GAN model surpasses existing two-input models in subjective and objective evaluations. In addition, our approach shows competitiveness with current three-input models.

Variational optimization (VO) pansharpening method proposes a model for optimizing the energy function by delineating the acquisition process of remote sensing images. However, traditional VO methods face challenges in solving the operators. Currently, some model-driven methods alleviate this problem by employing deep unfolding techniques. However, most model-driven approaches do not take into account the unknown and variable degradation process inherent in remote sensing images when solving the optimization model. They typically utilize a fixed Gaussian downsampling operator or a two-layer convolutional neural network module to directly simulate the degradation process during model unfolding, which leads to models that lack sufficient generalization. Therefore, this article proposes a pansharpening method based on deep nonlocal unfolding with degradation kernel estimation. Specifically, we expand the iterative process of solving the energy function into three modules: degradation kernel estimator, image generator, and three-branch prior network. First, the degradation kernel estimator is employed to fit the real degradation process of the image and optimize the generation of an adaptive degradation kernel. Simultaneously, we incorporate local priors, nonlocal spatial priors, and nonlocal spectral priors into the three-branch prior network to adaptively capture local and nonlocal prior features. These prior features are then fused with the generated image through residual connections to approximate the real remote sensing image. Experimental results on datasets from two different types of satellites demonstrate the superiority of our approach.

Abstract With the recent advances achieved by deep neural networks in image processing applications, researchers have begun exploring deep learning in pansharpening and obtained remarkable results. However, the existing methods are generally limited by their weak feature representation ability, often leading to spectral distortion or spatial blur. To generate high‐quality pansharpened images, this article proposes a novel neural network for pansharpening that includes both feature extraction and excitation mechanisms to consider important features. The neural network is modified with domain knowledge in pansharpening to fully extract spectral and spatial structures, and the proposed method outperforms traditional methods.

Although deep neural networks (DNNs) have achieved great success in pansharpening, most of them lack transparency and interpretability. Currently, some DNNs methods utilize deep unfolding techniques to alleviate this problem. However, they do not consider the regularization term separately when solving the energy function that represents the image degradation process, making it difficult to extract complex prior information in the unfolding module. Therefore, this paper proposes a pansharpening method based on deep non-local unfolding. Specifically, we expand the iterative process of solving the energy function into the corresponding neural network modules, making each module have a certain physical meaning. Then, we decouple the prior operator containing the prior knowledge of the remote sensing image and approximate the solution using the network module. Meanwhile, we incorporate local and non-local self-similarity priors into the prior operator and design a two-branch prior module for learning the prior features and contribution weights adaptively. Finally, the fused image is corrected with the learned prior features to approximate the real image. Experimental results on datasets from two different types of satellites demonstrate the superiority of our approach.

Remote sensing images often contain many similar components, such as buildings, roads, and water surfaces, which have similar spectra and spatial structures. Although convolutional neural networks (CNNs) based on residual learning can provide excellent performance in pansharpening, the existing methods do not make full use of intrinsically similar information in images. Moreover, since the convolution operation is focused on the local region, even in a deep network, position-independent global information is difficult to obtain. In this article, an efficient nonlocal attention residual network (NLRNet) is proposed to capture the similar contextual dependencies of all pixels. Specifically, to reduce the difficulty of network training caused by the original nonlocal attention, we propose an efficient nonlocal attention (ENLA) mechanism and employ residual with zero initialization (ReZero) technology to make the signal easy to spread through the network. Furthermore, a spectral aggregation module (SpecAM) is proposed to generate fused images and adjust the corresponding spectral information. The experimental results for the QuickBird and WorldView3 data sets show that the proposed method is competitive with other advanced methods based on quality assessment and visual perception.

In the remote sensing image fusion field, fusion methods based on deep learning are the latest techniques in panchromatic sharpening (pansharpening). However, existing pansharpening methods based on neural networks cannot adequately inject the spatial feature information of panchromatic (PAN) images into fusion images, and they do not exploit the feature relationships between spatial locations, such as rows and columns of feature maps. To solve these problems, a multidimensional channel attention residual neural network (MCANet) is proposed in this paper. To preserve the structural information in PAN images, a two-stream detail injection (TSDI) module is proposed, and the local skip connection operation is adopted to mine more spectral and structural information. A multidimensional channel attention (MCA) module is also designed to enable the network to learn the nonlinear mapping relationships between image spatial locations. In addition, a multiscale feature fusion (MSFF) module is designed to improve feature representation in the image fusion process, which is conducive to improving the pansharpening effect. The experimental results on the WorldView-2, GaoFen-2 and QuickBird datasets demonstrate that the proposed method outperforms state-of-the-art methods both visually and quantitatively.

With the continuous development of deep neural networks in the visual field, their application to panchromatic sharpening has received increasing attention from researchers; however, the existing panchromatic sharpening methods generally lack the ability to combine knowledge of the panchromatic sharpening field for feature extraction with neural networks, which have certain limitations in feature extraction and the discovery of new features. This article proposes a simple, modular, multibranched feature extraction and reuses network architecture designed not only to support feature reuse but to learn well-expressed new features for use in panchromatic sharpening approaches. In addition, we fused field knowledge of panchromatic sharpening to extract spatial structure information of panchromatic maps through gradient calculators and design structural and spectral compensation to fully extract and preserve the spatial structural and spectral information of images. We conducted experiments on the QuickBird and WorldView-3 satellite data sets, and the experimental results reveal that our proposed method has advantages over the best methods currently available, achieving excellent results not only on objective evaluation metrics, such as full-reference and no-reference metrics, but also on subjective visual evaluation.

In recent years, the powerful nonlinear modeling capability of convolutional neural networks has led to an increasing number of researchers focusing on deep learning-based pansharpening methods. However, due to the diversity of remote sensing image features and the limitations of the convolution operation, the existing methods are still inadequate in restoring the spatial details of complex remote sensing scenes. Therefore, in this paper, we propose a simple and effective network for pansharpening methods. Specifically, in our hierarchical feature integration architecture, a multi-scale grouping dilated block is designed to adequately capture fine-grained representations of multi-level scale features. At the same time, we propose a spatially self-attention block to adaptively improve the feature extraction process by establishing associations between features. The above blocks are connected in a hierarchical design, with selective reuse of features between layers, and a good ability to explore new levels of features while reusing low-level features. Our experiments with the GaoFen-2 satellite dataset, WorldView-2 satellite dataset, and WorldView-3 satellite dataset show that our proposed method is highly competitive with existing excellent methods in both objective indicator evaluation and subjective visual evaluation.