ADAPTIVE MULTI-STAGE PANSHARPENING CNN FOR HYPERSPECTRAL IMAGES

Pansharpening for hyperspectral (HS) images is hindered by heterogeneous characteristics of observed scenes. Most of existing methods do not take into consideration such inherent characteristics, and handle all of observed scenes by means of the identical pansharpening strategy, resulting in potential over-pansharpenings on homogeneous scenes or under-pansharpenings on non-homogeneous scenes. To solve this problem, in this paper we propose a adaptive multi-stage pansharpening convolutional neural network (CNN), called AdapMSNet, which can adaptively select different upsampled paths for various observed scenes according to their heterogeneous characteristics. Experimental results have proven the effectiveness and the superior pansharpening performance of our method in terms of the spatial reconstruction and spectral restoration.