Hyperspectral image (HSI) technology is widely applied in agriculture, geology, and other fields. However, challenges such as complex data acquisition and low accuracy have long constrained the development of agricultural remote sensing applications in China. Accurately selecting bands under information-scarce conditions will significantly advance agricultural remote sensing technology and contribute to resolving agricultural issues in China.

The research team, led by Professor Huimin Zhao and Professor Jinchang Ren from our institute, has proposed a deep learning framework called CAE-UBS, which is based on the Autoencoder method. This framework enables effective band classification even with scarce labeled information, thereby realizing an unsupervised band selection learning system for HSI.

The CAE-UBS framework is an end-to-end deep learning architecture that, for the first time, applies Gumbel-Softmax technology in the field of unsupervised band selection. It achieves the transformation from continuous real-numbered weights to binary weights, which contain only 0 and 1. Experiments conducted on four publicly available datasets demonstrate that CAE-UBS outperforms existing methods with superior performance and stability. This research achievement has been published in the internationally renowned IEEE Transactions on Geoscience and Remote Sensing (IF = 8.125).