The application of the hyperspectral image (HSI) classification has become increasingly important in industry, agriculture and military. In recent years, the accuracy of HIS classification has been greatly improved through deep learning based methods. However, most of the deep learning models tend to classify all the samples into categories that exist in the training data. In real-world classification tasks, it is difficult to obtain samples from all categories that exist in the whole hyperspectral image. In this paper, we design a framework based on convolutional neural networks and probability thresholds(CNPT) in order to deal with the open-category classification(OCC) problem. Instead of classifying samples of categories that do not exist in the training process to be any known class, the proposed method mark them as unseen category. We first get samples of unseen class from labeled data. With a lightweight convolutional network that fully uses the spectral-spatial information of HSI, we obtain the probabilities for each seen class for every sample. By adding a threshold to the maximum probabilities, we classify some samples to unseen category. A balanced score called Fue which considers both the recall rate of unseen class and the overall accuracy of seen classes is proposed in this paper, and we use it to select the threshold and evaluate the performance of CNPT. The experimental results show that our proposed algorithm performs well on hyperspectral data, and has generalizability on different datasets. © 2019 Association for Computing Machinery.