Comparative Study on Hyperspectral Unmixing Errors Considering Preprocessing Differences

Before conducting the unmixing of mixed pixels, the question of how to preprocess the original spectral data is a controversial issue. Mixed pixel samples with varying black‐to‐white ratios were fabricated, and their original spectral data were acquired using an ASD FieldSpec3 spectrometer in a laboratory setting. The eight commonly used preprocessing methods were divided into four groups: baseline correction (first derivative [FD] and second derivative [SD]), scattering correction (multivariate scattering correction [MSC] and standard normal variate [SNV]), smoothing (moving average [MA] and Savitzky–Golay [SG]), and scaling (maximum normalization [MN] and deviation standardization [DS]). The constrained linear spectral mixture model (CLSMM) was employed to unmix the original spectra and the preprocessed spectra, and the influence of preprocessing on unmixing was quantitatively analyzed by using the root‐mean‐square error (RMSE) and abundance error (AE). The results show the following: (1) the spectral signatures of mixed pixels are altered by the choice of preprocessing method. The SG and MA remain consistent with the original spectra. However, the spectral curve after DS processing approaches a straight line. (2) MSC has the lowest unmixing accuracy, with an RMSE of 0.531. Smoothing processing reduces the amount of spectral data while retaining the original spectral features. The RMSE of SG and MA is only 0.021. (3) By analyzing the abundance differences of different mixed pixel samples, it was found that as the black proportion within the pixel increases, the AE of most preprocessing methods shows a significant upward trend. However, MA and SG exhibit minor fluctuations, and the abundance difference is only 0.018. The results of this study establish a statistical foundation to guide the selection of appropriate preprocessing methods.