Abstract: In this paper, the Byes method, Backpropagation Neural Network (BP), Probabilistic Neural Network (PNN) method and Clustering method are used to classify the data rooted from the data of the second earth investigation result in Guangdong province. Moreover, based-on supervised-unsupervised cluster algorithm is presented in order to take full advantage of high accuracy of supervised learning classification and no necessity demarcated study samples. Then, the estimation result of the above five methods is compared and analyzed. According to the experiment, the gaining of higher classification accuracy just needs a few of demarcated study samples if based-on supervised-unsupervised learning is used. And especially, better earth estimation result than the Byes method, BP Network, PNN method and so on, can be gotten when a few of samples are used. In land evaluation, supervised and unsupervised learning method can be combined to implement a tradeoff between classification accuracy and expense.