Abstract: In order to study and analyse the influence the number of and quality of the training samples on the classification accuracy better, Helen city in Heilongjiang Province was chosen as the research required experimentation area, using Landsat 8 remote sensing images as the data source, the effects of the number and quantity of training samples on the classification accuracy were studied respectively by using the maximum likelihood, neural network and support vector machine three kinds of methods, and several experiments were made on these three kinds of classification methods. The final result shows that:(1) when the training sample quality is relatively constant, the degree of response of the same classification method to the same number of training samples as well as the degree of response of the different classification methods to the number of training samples are different, and the classification accuracy has different degree of volatility, with the increase of the number of training samples, the volatility will decrease, when the number of training samples reaches a certain degree, the mean of classification accuracy will tend to be relatively stable; (2) when the number of training samples is constant, the same classification methods as well as the different classification methods have different degree of response to the training samples of the same quality grade; the degree of response of the same classification method to the different training samples quality level is also different.