Many, historically or archeologically valuable paintings such as wall paintings in ancient tombs or Buddhist paintings in Japan or the other Asian countries have been drawn with natural mineral pigments (NMP). The digital archive of those paintings, the identification of pigments used in the paintings and the retrieval of color fading are strongly desired. Multispectral image acquisition of those paintings is very useful for both archive and analysis purposes. In this paper, we focus on the segmentation by pigment from multispectral images in the visible range. Here the kernel based nonlinear subspace method (KNS) is applied to the pigment-based segmentation of multispectral images of the paintings. At first, 55 NMP patches were made and the spectral reflectances were measured. Next, multispectral image acquisition of the color patch array and a Buddha painting drawn with those pigments were performed. Using the training sets of color patches, the segmentation of those images was performed. For comparison, image segmentation from three-band image and a conventional linear subspace method called CLAFIC were tested. It was found that the KNS method with multispectral images worked best than the other methods. Quantitative evaluation with color patches was carried out and the visual evaluation for the segmentation result of the Buddha painting was also performed.