A Classification Method for Polsar Images using SLIC Superpixel Segmentation and Deep Convolution Neural Network
Feng Gu, Hong Zhang, Chao Wang · 2018
Deep convolution neural networks (DCNN) have been successfully introduced in the field of Polarimetric SAR image classification. However, the commonly used DCNN will classify each pixel in the image and neglect the fact that neighboring pixels may have similar intensity. Besides, the fixed size input in DCNN cannot be well adopted in remote sensing image which includes a great deal of different-scale information. Thus, superpixel segmentation (SS) and the input pyramid are introduced in this paper to improve the performance of DCNN. The former will guide the DCNN to classify superpixel instead of single pixel and the latter will include different-scale information around the pixel. Experiments carried out on two scenes of ALOS-2 PALSAR-2 POLSAR images demonstrate that the introduced technic can help DCNN achieve good accuracy and smooth boundary adherence with highly efficiency.