Sensor-specific Transfer Learning for Hyperspectral Image Processing

Transfer learning (TL) has shown its great advantage to solve small-training-sample problems using knowledge learned from existing large data with deep learning techniques, which can be used for hyperspectral image intelligent processing in which labeled data is very difficult and even impossible to be obtained. However, the mismatch of hyperspectral sensors results in lots of difficulty for transfer learning to be used in hyperspectral image (HSI) processing. In this paper, sensor-specific based transfer learning is proposed for hyperspectral images acquired from same sensors, in which knowledge learn from hyperspectral images, e.g., the network structure and parameters of a deep neural network, are limited to transfer to images of the same sensor only. Specifically, the validity of sensor-specific transfer learning is evaluated using three deep learning based tasks, including feature learning, super-resolution, and image denoising. Experimental results from two benchmark datasets from the well-known ROSIS sensor, i.e., Pavia Centre and Pavia University, have demonstrated that sensor-specific based transfer learning can achieve satisfying performance even without fine-tune by small-training-samples on the target scene.