基于深度卷积神经网络和超像素CRF模型的高空间分辨率遥感图像多尺度征提取与分类识别研究

It is popular and important to apply spectral, spatial features and contextual information synthetically for high spatial resolution remotely sensed image classification. However, there are some drawbacks for the methods developed in this research area. For the feature extraction methods, they are separated from classification procedure and need professional prior knowledge to design. For the contextual information application, it is complicated in computing and difficult to capture long distance spatial context. According to these problems, in this study, we investigated to develop a framework based on convolutional neural networks and superpixel Conditional Random Fields (CRF) model to overcome those drawbacks. This framework includes three parts: hierarchical feature extraction based on deep convolutional neural networks. This model is based on the deep learning theory, fuses feature extraction and classification into one procedure without prior knowledge and can extract low level features, such as edge and corner, and high level features, like texture and shape, at the same time. The second part is superpixel based CRF model construction. This model can capture the long spatial context on superpixel scale. Except that, this model improves the computing efficiency of CRF significantly which makes it suitable for high spatial resolution remotely sensed image with huge data amount. The last part is superpixel based CRF model inference by applying Graph cuts algorithm. This project will provide a new way and framework for spectral-spatial based classification and improve the level of processing and the application for high spatial resolution remotely sensed images.

综合应用光谱、纹理、形状特征和空间上下文信息是目前高空间分辨率遥感图像分类识别研究的热点。针对目前研究中特征提取与分类识别过程相分离且依赖于领域先验知识来设计提取模型，以及难以捕获长距离空间上下文信息等问题，本研究提出了一种基于深度卷积神经网络模型和超像素条件随机场模型（CRF）的特征提取和分类识别框架，并进行实验验证。研究内容包括：基于深度卷积神经网络的多尺度特征提取研究。该模型依据新型的深度学习理论，融合特征提取和分类识别于统一过程，无需领域先验知识，并能够完成从低级特征到高级特征的一体化多尺度提取；基于超像素的CRF模型构建研究。该模型能够捕获长距离的空间上下文信息，并大大减少CRF模型的计算复杂度；基于Graph cuts算法的CRF模型最优化推断(inference)。本研究将为融合光谱、空间特征和空间上下文信息高空间分辨率遥感图像分析提供一种新的思路，并提高其处理和应用水平。

综合应用光谱、纹理、形状特征和空间上下文信息是目前高空间分辨率遥感图像分类识别研究的热点。针对目前研究中特征提取与分类识别过程相分离且依赖于领域先验知识来设计提取模型，以及难以捕获长距离空间上下文信息等问题，本研究提出了一种基于深度卷积神经网络模型和超像素条件随机场模型（CRF）的特征提取和分类识别框架，并进行实验验证。研究内容包括：基于深度卷积神经网络的多尺度特征提取研究。该模型依据新型的深度学习理论，融合特征提取和分类识别于统一过程，无需领域先验知识，并能够完成从低级特征到高级特征的一体化多尺度提取。融合多层卷积神经网络空间信息和光谱信息的遥感图像分类研究，光谱信息作为一个独立的特征输入显著的改善了分类的效果。开展了基于FPGA的遥感图像实时化预处理和空间特征提取的相关工作，基于FPGA的编程模式是遥感图像分析的全新方式，该研究为后续的遥感图像软硬件相结合分类识别研究打下坚实基础。本研究将为融合光谱、空间特征和空间上下文信息高空间分辨率遥感图像分析提供一种新的思路，并提高其实时化处理和应用水平。