This paper describes a neural network-based technique to compress multispectral SPOT satellite images losslessly. The technique harnesses the pattern recognition property of one-hidden-layer back propagation neural networks to exploit both the spatial and the spectral redundancy of the three-band SPOT images. The networks are initially trained on samples of the SPOT images with a unique network for each of the bands. The resultant trained nonlinear predictors are then used to predict the target SPOT images. Predicted errors are entropy-coded using multi-symbol arithmetic coding. This technique achieves compression ratios of 2.1 times and 3.2 times for urban and rural SPOT images respectively which are above 10% better than using lossless JPEG compression techniques. In comparison with JPEG2000 lossless compression, the proposed technique is 5% better.