少量样本条件下的高光谱图像集成深度神经网络分类方法

Lack of training samples will not only lead to over-fitting of hyperspectral image classification method based on deep neural network, but also affect the design of deep neural network structure. This project focus on hyperspectral image classification method based on the depth of neural network under the condition of a small amount of samples. For neural Recurrent Neural Network (RNN) and Convolutional Neural Network (CNN), We will fully explore the existing data resources and optimize the network structure design, and then, propose a spectral library data assisted RNN for classification and a multi-scale feature fused CNN for classification. On this basis, the deep neural network method integrated by RNN and CNN is studied to give full play to their advantages in extracting spectral features and spatial features, so as to improve the accuracy and stability of the deep neural network algorithm for hyperspectral image classification. Finally, the model accuracy and adaptability of the proposed method are verified by using hyperspectral images at different spatial resolutions. The study of this project can effectively solve the problem of insufficient training samples and make up the defect that convolutional neural network is unable to extract large-scale spatial features under the condition of a small number of samples. It is of great significance to improve the robustness and generalization ability of the deep learning method for large-scale hyperspectral image classification.

训练样本不足不仅会导致基于深度神经网络的高光谱图像分类方法容易陷入过拟合，也会影响深度神经网络结构的设计进而影响模型精度。本项目聚焦少量样本条件下基于深度神经网络的高光谱图像分类方法，针对循环神经网络RNN和卷积神经网络CNN两种优势算法，充分挖掘现有数据资源并优化设计网络结构，分别提出光谱库数据辅助的循环神经网络分类方法，和多尺度空间特征融合的卷积神经网络分类方法；在此基础上，研究RNN与CNN集成的深度神经网络方法，充分发挥二者在提取光谱特征与空间特征方面的优势，提高面向高光谱图像分类的深度神经网络算法的精度和稳定性；最后，利用不同空间分辨率的模拟和真实高光谱图像，对本项目所提出方法的模型精度与适应性进行验证。本项目研究可以有效解决训练样本不足问题，弥补少量样本条件下卷积神经网络无法提取大尺度空间特征的缺陷，对提高面向大区域高光谱图像分类的深度学习方法的鲁棒性和泛化能力具有重要意义。

利用深度学习技术对光学遥感图像或者高光谱图像进行地物分类和目标检测具有重要意义。但是一般情况下深度学习技术需要大量的样本进行模型训练，如何在少量样本条件下提高基于深度学习的遥感图像分类与检测精度，已成为近年来研究的热点。本项目首先通过物理模型生成大量模拟样本数据，用于源域中模型的预训练，开展了基于深度学习网络（LSTM）的高光谱图像分类模型研究；其次，研究了基于多源多时相遥感影像的深度学习作物类型分类算法；最后，利用深度卷积神经网络，开展了面向精细化多尺度特征的遥感图像目标检测方法研究。通过本项研究，提出了基于物理模型生成模拟样本、多源多时相数据融合、精细化多尺度特征融合等三种在少量样本条件下基于深度学习的遥感图像分类与目标检测方法。