Recently, the application of artificial neural networks (ANN) into classification has raised a great deal of interest. The standard back-propagation (BP) algorithm is suitable for training neural networks. Generally, the topological structure of the back-propagation neural network (BPNN) used in the image classification consists of neurons arranged into three layers, namely input layer, hidden layer and output layer. Obviously, the number of input layer nodes in a BPNN generally corresponds to the number of features (spectral bands), which influences the iterative (or training) time. Nowadays, sensors provide more and more spectral bands. Hence, the representation of multi-spectral remote sensing data in ANN has become a mayor problem. Selection of the effective image bands in order to reduce the size of the input data is therefore necessary using, for example, the Principal Component Analysis (PCA). In this paper, an improved BPNN model, using the PCA and BPNN, has been developed. This proposed method can reduce the number of input layer nodes aiming at attaining the effective "bands" for classification. The experimental results show that the proposed method can cut down computational costs during training stage because of the reduction of the number of input nodes and improve the overall classification accuracy.