Few-Shot Shrub Identification via Hyperspectral Deep Learning: A Case Study of Caragana microphylla Lam in Shrub-Encroached Grasslands

Accurately extracting the spatial distribution of shrubs is an important basis for scientific diagnosis, rational prevention, and control of shrub-encroached grasslands (SGs). In SGs, the landscape exhibits a high degree of spatial intermingling among shrubs, herbaceous vegetation, and bare soil, resulting in a serious issue of mixed pixels. The combination of UAV hyperspectral imaging and deep learning provides the most promising technical approach for shrub–grass separation at present. However, issues such as data redundancy and algorithmic adaptability urgently need to be addressed. In this study, Caragana microphylla Lam, a typical shrub species found in the SGs of Inner Mongolia, is the subject for extraction. This work aims to extract sensitive spectral bands from UAV hyperspectral data, construct a deep learning-based framework for shrub identification, and map the distribution of Caragana microphylla Lam in the study area. First, a dataset of hyperspectral images in Xilinhot City, Inner Mongolia, China, was constructed by manual visual interpretation for shrub identification. Second, a novel unsupervised band selection algorithm, US-BS-Net, was proposed to optimize the proxy task of BS-Conv-Net by introducing contrastive learning. It can be used to obtain sensitive spectral bands with discriminative feature representations. Finally, the US-BS-P-Net framework was proposed by combining the US-BS-Net with the prototypical network to construct high-precision shrub identification models in small-sample scenarios. Taking 80 bands as input, the highest overall accuracy reaches 93.15%, which is better than the full spectrum and other deep learning models. The deep synergy between hyperspectral technology and deep learning effectively overcomes traditional challenges such as spectral similarity between shrubs and grasses and their mixed spatial distribution. In particular, the proposed US-BS-P-Net, which combines the advantages of US-BS-Net in sensitive band extraction and the prototypical network in few-shot model construction, provides excellent fundamental data for precise monitoring of SGs.