基于深度卷积神经网络的多源遥感图像时空融合方法研究

The previous spatio-temporal fusion models mainly focused on small datasets and the simple linear mapping model or the image degradation model, and lacked consideration for the big difference in spatial scale between different types of remote sensors, leading to the improvement space in both accuracy and practicality. This project aims to study the application of deep convolutional neural network in spatio-temporal image fusion by utilizing the properties of massive remote sensing data and proposes solutions for existing problems in previous models in aspect of model, algorithm and application. The main research contents include: utilizing the non-linear mapping of deep convolutional neural network to build a dynamic and adaptive mapping model; designing layer-adaptive cascaded convolutional network to deal with the big difference in spatial scale between different types of remote data, thereby increasing the accuracy of the designed algorithm; exploring the online update mechanism of convolutional neural network to deal with massive remote sensing data, thereby enhancing the practicality of the designed algorithm.

以往的时空融合模型主要基于小规模数据，通过简单的线性映射模型或者图像降质模型在两类传感器图像之间建立映射关系，未考虑异源遥感数据之间空间尺度差异大的问题，因此在融合精度和实用性方面都有待提高。考虑到海量遥感数据的特性，本项目研究深度卷积神经网络在遥感图像时空融合中的应用，针对以上融合模型存在的问题分别从模型、算法和应用角度提出相应的解决方案。主要研究内容包括：利用深度卷积神经网络隐层的非线性映射关系，在异源遥感图像之间建立动态自适应的映射模型；设计层数自适应级联卷积神经网络来处理异源遥感数据之间的空间尺度差异大问题，从而提高融合精度；探索卷积神经网络的在线更新机制来处理海量遥感数据，以增加算法的实用性。

针对如何设计遥感图像时空融合中非线性映射模型这一核心问题，项目组提出了卷积神经网络非线性映射模型、非线性映射模型和超分模型级联策略、残差网络等创新性技术，相比以往的时空融合算法，所提算法在融合精度、鲁棒性和适用范围方面均得到了提升。目前已在国内外知名期刊发表论文14篇，包括IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing 3篇，TIP、IEEE Transactions on Cybernetics、International Journal of Remote Sensing、Pattern Recognition、Computer Vision and Image Understanding、Neurocomputing各1篇，Electronics Letters 2篇, 中文核心期刊3篇。其中，SCI（E）源刊11篇。申请发明专利3项。发表高质量论文数和申请专利数均超出验收指标。已基于研究成果搭建了一套遥感数据时空融合原型系统，并成功应用于分别具有物候变化和地物类型变化的标准数据集。研究成果在农作物长势监测、土地覆盖变化检测和灾害监测等领域具有广泛的应用前景。