<p>Despite the success of deep neural networks (DNNs), insufficient labeled samples remain a key challenge for DNN based hyperspectral image (HSI) classification. To perform better with limited labeled samples, some HSI classification methods try to acquire spatial information through superpixel segmentation (SPS). However, the existing methods seldom use multi-band and multi-scale SPS simultaneously to extract features, and lack effective means to mine the superpixel-wise information fully. This paper proposes a semi-supervised learning method combined with regression pre-training of Siamese network based on superpixel segmentation (SLRSS). The method innovatively introduces two key components: (1) an average edge-weighted graph using SPS results that provides the regression labels for the regression pre-training, and (2) a model-selected labeling (MSL) mechanism that enhances the subsequent semi-supervised learning process. First, we design an average edge-weighted graph to generate the similarity labels of each sample pair by using the multi-band multi-scale SPS results obtained from the unlabeled HSI. Next, we randomly sample a proportion of sample pairs with their corresponding similarity labels and add them into the regression training (RT) set. To fully utilize the limited labeled samples in the training set for classification, these limited labeled samples are also used to construct the sample pairs with the similarity labels according to their true labels and added into the RT set. Then, the parameters of the well pre-trained Siamese network after the regression training are used to initialize the parameters of the feature extraction module in the classification training. Finally, we design a novel semi-supervised learning strategy named as MSL to augment the training set in the classification training. Experimental results on two HSI datasets show that the proposed SLRSS outperforms several state-of-the-art methods significantly, with only five labeled samples per class.</p>