Local binary pattern (LBP) is a simple and effective method for texture classification. LBP-based methods have been widely used in texture classification research. However, Although the existing LBP-based texture classification methods have certain robustness to noisy images, they have some defects when dealing with texture images with high levels of noise. Therefore, to address the problem of sensitivity to noise at high levels, we propose a simple and robust texture descriptor for texture classification, named multiscale group variance binary pattern (MGVBP). Firstly, we propose the local median group pattern (LMGP) to extract the texture information around the adjacent pixels in the local area. Secondly, we propose local variance binary pattern (LVBP) to construct histograms of texture images. Finally, we construct multi-scale grouped variance binary patterns by concatenating multiple single-scale grouped variance binary patterns. Our proposed MGVBP descriptor is evaluated on popular texture datasets for classification tasks (KTH-TIPS, CUReT, UIUC, Brodatz and Kylberg) and under different high-noise conditions. Compared with state-of-the-art texture classification methods, the MGVBP descriptor achieves satisfactory classification performance.