基于深度学习的高光谱图像空-谱联合特征提取

发布时间：2018-06-15 22:28

本文选题：图像处理 + 高光谱图像分类　；参考：《激光与光电子学进展》2017年10期

【摘要】：由于高光谱遥感数据具有波段多、特征非线性、空间相关等特点,提出一种基于深度学习的空-谱联合(SSDL)特征提取算法来有效提取数据中的空-谱特征。该算法利用多层深度学习模型——堆栈自动编码机对高光谱数据进行逐层学习,挖掘图像中的深层非线性特征,然后再根据每个特征像元的空间近邻信息,对样本深度特征和空间信息进行空-谱联合,增加同类数据聚集性和非同类数据分散度,提升后续分类性能。在帕维亚大学和萨利纳斯山谷高光谱数据集上进行地物分类实验:在1%样本比例下,地物总体分类精度达到了91.05%和94.16%;在5%样本比例下,地物总体分类精度达到了97.38%和97.50%。结果表明:由于SSDL特征提取算法融合了数据中深层非线性特征和空间信息,能够提取出更具鉴别特性的特征,较其他同类算法能够获取更高分类精度。
[Abstract]:Because hyperspectral remote sensing data have many characteristics, such as multi-band, nonlinear features and spatial correlation, a space-spectrum combined with SSDL feature extraction algorithm based on depth learning is proposed to extract space-spectrum features from the data effectively. The algorithm uses a multi-layer depth learning model-stack automatic coding machine to study hyperspectral data layer by layer, mining the deep nonlinear features in the image, and then according to the spatial nearest neighbor information of each feature pixel. The spatial spectral combination of the depth features and spatial information of the samples can increase the aggregation of similar data and the dispersion of non-similar data, and improve the performance of subsequent classification. Based on the hyperspectral data sets of Pavia University and Salinas Valley, the classification accuracy of ground objects is 91.05% and 94.16%, and 97.38% and 97.50% respectively. The results show that the SSDL feature extraction algorithm can extract more discriminant features because of the fusion of the deep nonlinear features and spatial information in the data, and can obtain higher classification accuracy than other similar algorithms.
【作者单位】：重庆大学光电技术及系统教育部重点实验室;
【基金】：国家自然科学基金(41371338) 重庆市研究生科研创新项目(CYS16040) 重庆市基础与前沿研究计划(cstc2013jcyjA40005)
【分类号】：TP751

【相似文献】