Hyperspectral images contain rich spectral information and have better performance in ground target recognition than panchromatic and multispectral images. An extremely randomized clustering forests (ERC-Forests) algorithm is introduced after analysis of the decision tree algorithm. Hyperion hyperspectral images and IRS-p6 image data of Qilian County, Qinghai Province, are used in the experiment. After dimension reduction with subspace methods and based on the spectral range, support vector machine (SVM), neural network (NN) and maximum likelihood (MLC) are used for classification. The results are compared with that of random decision trees algorithm, showing that the extremely randomized clustering forests algorithm is better, without dimension reduction. The method is widely applicable to hyperspectral remote sensing.