Hyperspectral spatial-spectral joint feature extraction method based on transfer learning

• Main Authors: LI XUEQIONG, YANG WENJING, ZHOU DONG, LAN LONG, REN JING, PENG YUANXI, XU LIYANG, ZHAO LIYUAN, LIU YU, HUANG DA, YANG SHAOWU, XU WEIXIA
• Format: Patent
• Language: Chinese; English
• Published: 30.03.2021
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; HANDLING RECORD CARRIERS; PHYSICS; PRESENTATION OF DATA; RECOGNITION OF DATA; RECORD CARRIERS

The invention discloses a hyperspectral spatial-spectral joint feature extraction method based on transfer learning, and belongs to the field of deep learning remote sensing. The method for extractingthe spatial-spectral combined characteristics of the hyperspectral data comprises the following steps: firstly, designing a 1D CNN and a 2D CNN to respectively extract spectral and spatial characteristics of the hyperspectral data, and then fusing the two parts of characteristics. In order to overcome the contradiction that a deep neural network needs a large amount of training data and hyperspectral data lacks marked samples, a method of migrating a model ResNet-18 pre-trained on an RGB image data set ImageNet to a hyperspectral image target domain is adopted, network parameter sharing is realized, and the calculation cost of a training model is reduced. A SoftMax layer is trained based on the extracted combined features to realize a hyperspectral target classification task. Finally, through a fine-tuning transf