To address the catastrophic forgetting problem in lifelong learning, this paper proposes a lifelong learning method for hyperspectral image classification using multi-level knowledge distillation. Initially, a feature extractor based on multi-modal alignment is designed to fully leverage the spatial-spectral information and label text information of hyperspectral images. Additionally, a multi-level knowledge distillation strategy is devised to effectively preserve the multi-modal knowledge from previous phases. The proposed method was experimented on two public hyperspectral datasets. Compared to current state-of-the-art methods, the proposed method showed an average accuracy improvement of 15%-18% on the Pavia University dataset and 1%-8% on the Botswana dataset.