Reflectance spectral curve reveals the unique physical characteristic of different materials. Through spectral match and recognition, different materials could be distinguished. Because of the great amount of spectral data and the ambiguous absorption feature of original spectral curve, feature extraction of reflectance spectral curve is one of the essential techniques in hyperspectral image classification and recognition. Using wavelet decomposition technique, the present paper proposes a new spectral feature extraction algorithm to compress data amount while reserve spectral feature selectively. Through selecting the appropriate decomposition level by measuring the objective absorption feature frequency, the original signal would be projected into a new feature space with less data amount and more obvious objective feature than the original one. The experiments show that the method proposed can reduce the spectrum dimensions effectively and improve the recognition precision with the spectrum matching.