摘 要: 随着遥感技术的快速发展，遥感图像全色锐化在矿产勘探、城市规划和地质灾害监测等领域中得到广泛应用。然而，现有全色锐化方法在融合全色图像与多光谱图像时普遍存在计算复杂度高、局部特征与全局信息提取不充分等问题。为此，构建了一种基于全局感知卷积（GAConv）与 Transformer 的融合网络（GCTNet）。该网络采用双分支多尺度架构分别提取全色图像与多光谱图像的空间和光谱特征，并结合 GAConv 模块与 Transformer 模块，有效捕捉局部细节与全局上下文信息，从而提升融合图像质量。实验结果表明，在多个遥感数据集的全色锐化任务中，GCTNet 相较于当前先进方法取得了更优性能，显著提升融合图像质量并降低了模型计算复杂度。

Abstract: Pansharpening plays a crucial role in remote sensing applications such as mineral exploration, urban planning, agricultural monitoring, and geological hazard analysis. High-resolution remote sensing imagery is often obtained by combining a high-resolution panchromatic (PAN) image with a low-resolution multispectral (MS) image, making fusion quality essential for subsequent interpretation tasks. However, existing pansharpening methods often suffer from high computational complexity, limited representation capability, and insufficient extraction of both local spatial details and global contextual dependencies. To address these limitations, this study proposes a novel Global-Aware Convolution and Transformer-based fusion network, termed GCTNet, designed to enhance feature extraction, improve spatial–spectral consistency, and maintain computational efficiency. GCTNet is constructed using a dual-branch multi-scale architecture that independently extracts spatial and spectral features from the PAN and MS inputs. A Global-Aware Convolution (GAConv) module is introduced to replace conventional convolution operations through a learnable linear combination of two reference kernels, significantly reducing redundant parameters while preserving detailed spatial modeling capability. A Global Feature Harmonization (GFH) mechanism further injects global contextual cues into the convolution process. To capture long-range dependencies that traditional convolutions struggle to represent, a lightweight Transformer module is embedded into the network. This module includes a Multi-DConv Head Transposed Attention (MDTA) mechanism for efficient global context modeling and a Gated Depthwise Convolutional Feed-Forward Network (GDFN) for adaptive local feature refinement. Subsequently, an attention-based feature fusion module is employed to adaptively integrate spatial and spectral information from the two branches, ensuring balanced and effective fusion. Multi-scale encoders and decoders process hierarchical representations at different resolutions, and a reconstruction module produces the final high-resolution MS images. Comprehensive experiments were conducted on three widely used remote sensing datasets—WorldView-3, WorldView-2, and QuickBird—under both reduced-resolution and full-resolution evaluation protocols. Experimental results show that GCTNet achieves leading performance among 13 recent pansharpening methods, including both traditional models and deep learning-based approaches. Quantitative evaluations across multiple metrics, including PSNR, SAM, ERGAS, and SCC, demonstrate significant improvements in spatial detail preservation and spectral fidelity. Visual assessments further confirm that GCTNet produces fused images with sharper structural edges, clearer textures, and reduced spectral distortion. Full-resolution experiments validate the robustness and practical applicability of the proposed approach, while cross-sensor generalization tests—trained on WV3 and tested on WV2—demonstrate strong transferability across different satellite sensors. This study presents GCTNet, a hybrid pansharpening framework that integrates GAConv and Transformer modules to effectively capture both local spatial details and global contextual dependencies. By reducing redundant parameters, enhancing global feature perception, and incorporating adaptive multi-scale fusion, GCTNet achieves state-of-the-art pansharpening performance with high robustness and strong cross-sensor generalization capability. The proposed method provides an effective and practical solution for high-quality remote sensing image fusion and holds great potential for real-world applications in Earth observation, environmental monitoring, and geospatial analysis.