Attention and Transform hyperspectral image classification method and system

• Main Authors: YANG SHUMING, ZHANG GUOFENG, GE LINGYU, LIN KAILIANG, DENG HUIWEN, MENG YAYUN
• Format: Patent
• Language: Chinese; English
• Published: 09.05.2023
• Subjects: CALCULATING; COMPUTER SYSTEMS BASED ON SPECIFIC COMPUTATIONAL MODELS; COMPUTING; COUNTING; PHYSICS

The invention discloses an attention and Transform hyperspectral image classification method and system. The method comprises the following steps: acquiring a hyperspectral image data cube and a ground feature label; transposing the dimension of the data cube; carrying out pixel-by-pixel division on the ground feature label matrix according to a set proportion; dividing the image data into blocks according to ground object label matrix division; fusing the images of the divided blocks by using an attention 3D convolution layer; and inputting the data cube experiencing the attention 3D convolution layer into an improved Transform coding layer, and outputting a category through a decoding layer to obtain an optimal model. According to the method, local and global spatial spectrum information of a hyperspectrum can be obtained without complex pretreatment such as dimension reduction, and feature extraction and fusion are carried out from a global direction by using an improved Transform, so that the whole networ