Remote sensing data have been used for effective recognition and classification of land use and land cover features on Earth surface. This paper presents a new approach of handling mixed pixel problem present in remote sensing data and nonlinearity between class boundaries through incorporating the kernel functions with noise classifier (NC). Kernel functions have been combined with conventional noise clustering without entropy, classification method (KNC) to classify data obtained from Landsat-8 and Formosat-2 satellites. Simulated image technique has been introduced and used to handle mixed pixel problem and to assess the results of adopted classification method (KNC). The procedure includes optimization of parameters for nine different kernels, finding the best performing kernel, and image to image accuracy assessment using fuzzy error matrix. A comparative analysis of kernel-based noise classifier over conventional NC classifier has also been included in this paper.