In recent years, more and more multi spectral satellite data have become available, and accordingly, the method to transform multi-dimensional data into less dimensional ones (called as "dimensional compression") is desired for the purpose of effective and comprehensive use of all the data. Principal Component Analysis (PCA) is one of the most traditional method for the dimensional compression in remote sensing field, but its limitation is proved that PCA cannot function so flexibly because of linear transformation. Neural Network (NN) models can behave like a non-linear transform function as a whole, because non-linear functions are embedded in a model and the complexity of those functions connection can implement non-linear performance as a whole. In this paper, authors examined the feasibility of the NN model for dimensional compression, which had been theoretically reported to be superior to PCA and had been examined only by applying for simulated data. Firstly, the architecture of the NN model for dimensional compression is described. Then, the NN model is applied for the compression of 6 bands of satellite data, or Lansdat TM data into 1, 2 or 3 dimensional ones. The compression efficiency of the NN model is compared with PCA's one by two indices, correlation coefficient and standard deviation of errors between the original data and the restored one. The result demonstrates that in the case of compression into 1 and 2 dimensions, the NN model is more appropriate for the dimensional compression than PCA, and that in the case of compression into 3 dimensions, there is no meaningful difference between both methods.