As in other remote-sensing applications, collecting ground-truth information from the earth's surface is expensive and time-consuming process for hyperspectral imaging. In this study, a deep learning-based semisupervised learning framework is proposed to solve this small labeled sample size problem. The main contribution of this study is the construction of a deep learning model for each hyperspectral sensor type that can be used for data obtained from these sensors. In the proposed framework, the "trained base model" is obtained with any dataset from a hyperspectral sensor, and fine-tuned and evaluated with another dataset. In this way, a general deep model is developed for extracting deep features which can be linearly classified or clustered. The system is evaluated with three different clustering techniques, the modified k-means, subtractive, and mean-shift clustering, for selecting initial representative labeled training samples comparatively. Another contribution of this study is to exploit the labeled and unlabeled sample information with linear transductive support vector machines. The proposed semisupervised learning framework is proven by the experimental results using different number of small sample sizes.