Most of the existing deep learning methods for hyperspectral image (HSI) classification use pixel-wise or patch-wise classification. In this paper, we propose an image-wise classification method, where the network input is the original hyperspectral cube rather than the spectral curve of each pixel (i.e., pixel-wise) or neighbor region of each pixel (i.e., patch-wise). Specifically, we propose a minimalistic fully convolution network (MFCN) and a semi-supervised loss function, which can perform pixel-level classification for HSI with few labeled samples. The comparison experiments demonstrated the progress of our methods, using three new benchmark HSI datasets (WHU-Hi-LongKou, WHU-Hi-HanChuan and WHU-Hi-HongHu) with wavelength range from 400 to 1000nm. In the comparison experiments, we randomly selected 25 labeled pixels from each class for training, equivalent to only 0.11%, 0.16%, and 0.14% of all labeled pixels for the three datasets, respectively. In addition, through ablation studies and theoretical analysis, we verified and analyzed the effectiveness and superiority of our design choices.