One of great challenges in neural network-based analysis of remotely sensed imagery is to find an adequate pool of training samples without prior knowledge for the network so that that these unsupervised training samples can describe the data. A judicious selection of training data can be tremendously difficult due to the presence of subpixel targets and mixed pixels, particularly, when no prior knowledge is available. Surprisingly, the above issues have been largely overlooked in the past, where most of the efforts have been focused on exploring network architecture parameters such as the arrangement and number of neurons in the different layers. Very little has been done in regard to the selection of a set of good training samples for networks in mixed pixel classification. This paper revisits neural network-based mixed pixel classification from an aspect of training sample generation and further demonstrates that the selection of training samples can be more important than the choice of a specific network architecture. Since the training samples must be obtained directly from the data to be processed in an unsupervised fashion, four types of pixels: pure pixel, mixed pixel, anomalous pixel and homogeneous pixel are used to demonstrate this concept. A pure pixel is a pixel whose spectral signature is completely represented by a single marterial substance as opposed to a mixed pixel whossee spectral signature is made up of more than one material substance. A homogeneous pixel is defined as a pixed whose spectral signature remains nearly constant subject to small variations within its surroundings. Therefore, a homogeneous pixel can be considered as an opposite of an anomalous pixel whose signature is spectrally distinct from the signatures of its neighboring pixels. In this paper, various scenarios are designed for experiments to substantiate the impact of using these four types of pixels as training samples for mixed pixel classification.