Aiming at overcoming the limitations in extracting impervious surface by traditional methods, two non-linear spectral mixture models, Mixture Tuned Matched Filtering (MTMF) and Multi-Layer Perceptron(MLP) neural network, are used to decompose all pixels to the four fraction images representing the abundance of four endmembers. In these models, MTMF performs a “partial” unmixing by only finding the abundance of a single, user-defined endmember, by maximizing the response of the endmember of interest and minimizing the response of the composite unknown background. The MLP is a hierarchical structure of several perceptrons, and capable of learning a rich variety of nonlinear decision surfaces. The Maximum Noise Fraction(MNF) is used to transform the six bands of TM image into a new feature space and the first three components accounting for the majority (more than 90%) of total information content are used to endmember extraction. After that, the Pure Pixel Index(PPI) is used to select pure pixels. The N-dimensional visualizer is used for assisting selection of four endmembers: vegetation, high-albedo objects, low-albedo objects and soil. The fraction images are derived to represent the abundance of the above four endmember. Impervious surface is estimated by analyzing high-albedo and low-albedo fraction images. QuickBird multi-spectral image is used to evaluate the accuracy of impervious surface extraction by different methods. Experimental results indicate that the accuracy of artificial neural network is higher than others, which means non-linear spectral mixture models is also effective to impervious area extraction, even better than linear models.