基于深度学习的高分辨率遥感图像融合研究

Image processing and information extraction using high-resolution remote sensing imagery is a hotspot and the frontier in remote sensing research. The fusion of high-spatial-resolution (HSR) panchromatic (PAN) and multispectral (MS), which is also referred as pansharpening, is one of standard steps in image process. Quality of fused products generated by current pansharpening methods are affected by several facts, including misregistration and correlations between PAN and MS bands. How to reduce spectral distortions of fused images and sharpen boundaries between different objects is a challenging problem in pansharpening. With significant advantages in deep feature extraction and complex relationship expression, deep learning provides powerful tool to solve the challenges in pansharpening. Though the research on pansharpening using convolutional neural networks (CNNs), the major objects of this project include: 1) developing new pansharpening methods using deep CNNs, which are less sensitive to misregistration and low correlations between PAN and MS bands, and reduce spectral distortions and sharpen boundaries between different image objects; 2) exploring a universal deep CNN model, which can be used by other users without time-consuming training, for the fusion of HR PAN and MS bands recorded by WorldView-2/3. The research productions of this project will break through the key problems in pansharpening, obtain fused images with reduced spectral distortion and sharpened boundaries between different objects, and promote the use of HSR fused images in applications such as land cover classification, change detection, and objects recognition.

高分辨率遥感图像处理与信息提取是目前遥感研究领域的热点与前沿。全色-多光谱图像融合是高分辨率图像处理的标准流程之一。目前全色-多光谱图像融合受几何配准误差、波段相关性等影响，如何进一步降低光谱失真和锐化地物边界是面临的挑战之一。鉴于深度学习具有强大特征提取和表达能力优势，本项目针对上述问题，开展基于卷积神经网络的高分辨率全色-多光谱图像融合研究，达到以下研究目标：（1）发展基于深度卷积网络的新融合方法，降低融合图像光谱失真并锐化地物边界，降低几何配准误差、波段相关性等对融合图像质量的影响；（2）探索面向某一特定传感器遥感图像融合的通用深度网络模型，直接用于特定传感器图像的融合，为面向其他传感器遥感图像的通用模型研究奠定基础。通过本项目的研究，突破高分辨率全色-多光谱图像融合中的关键问题，降低融合图像光谱失真和锐化地物边界，提高高分辨率融合图像用于地物分类、目标识别等应用的精度。

高分辨率遥感图像处理与信息提取是目前遥感研究领域的热点与前沿。全色-多光谱图像融合是高分辨率图像处理的标准流程之一。目前全色-多光谱图像融合受几何配准误差、波段相关性、混合像元问题等影响，如何进一步降低融合图像光谱失真和锐化地物边界是面临的挑战之一。针对上述问题，本项目开展基于卷积神经网络的高分辨率全色-多光谱图像融合研究，取得了以下研究成果：（1）建立了考虑混合像元的图像融合新框架，通过改善混合像元融合效果降低了基于部分CS和MRA方法融合图像的光谱失真；（2）建立了基于残差神经网络的图像融合模型，达到了降低融合图像光谱失真和锐化地物边界的效果，显著地提高了融合图像质量；（3）发展了基于卷积神经网络的未完全配准图像融合问题，解决了全色-多光谱图像融合图像质量受配准误差影响存在的边界模糊问题；（4）建立了面向WorldView-2/3 图像融合的通用深度网络模型研究，通过少量微调训练即可用于生成高质量WordView-2/3融合图像，解决了深度学习用于图像融合过程中需要大量训练的问题。具有通用强、鲁棒性好、速度快的优势。本项目的研究成果，实现利用深度学习技术进一步降低融合图像光谱失真，对提高高分辨率融合图像用于地物分类、目标识别等应用的精度具有重要意义。