In order to realize the spoilage warning of bananas during storage, hyperspectral data of bananas with different storage time was collected. The original spectrum was smoothed by Savitzky-Golar (SG) so as to obtain the spectral data with less noise interference. And a method of extracting the feature wavelength from hyperspectral information based on Wilks Λ statistic combined with principal component information fusion is proposed. The method could be described as follows: firstly, principal component analysis (PCA) was applied to the hyperspectral data after SG treatment, then the corresponding Wilks Λ statistics value was calculated for the each obtained principal component, and then the principal component variable with the lowest Wilks Λ value (the third principal component, PC3) could be selected. And the variation curve of combination weight coefficient corresponding to the original variable (each wavelength) of the PC3 was plotted. The wavelengths variable corresponding to the wave peaks and valleys in the weight coefficient curve were picked to be as the feature wavelengths. According to this method, totaling 9 feature wavelengths were extracted. At the same time, the color difference data of the banana was analyzed. By analyzing the trend of the L^*, a^*, b^* and ΔE values of the test samples with the storage time, the position of the abnormal data point (inflection point) was found. Combined with the actual sensory situation, the spoilage benchmark was initially determined as the sixth storage day. In order to further illustrate the rationality of the given spoilage benchmark, the feature spectral data was used to make the average spectral reflectance curves of different storage time samples. It was found that the spectral reflectance values reached the minimum value at the 6th day of storage for the different feature wavelengths, which was consistent with analysis results based on the color difference index, it was further determined that the spoilage benchmark was indeed the sixth storage day. Thus, the benchmark information could be characterized by the feature spectral data of the sixth storage day, and then the feature spectral representation vector and corresponding covariance matrix of the spoilage benchmark could be generated. Finally, took the hyperspectral feature vector characterizing the spoilage benchmark as a reference point, a spoilage early warning model based on feature wavelength spectrum information during banana storage was established by Mahalanobis distance (MD) and was verified effectively. The results showed that the MD between the test samples and spoilage benchmark given by the early warning model was getting closer and closer with the extension of banana storage time, which was consistent with the actual process of banana spoilage. Therefore, the proposed hyperspectral feature wavelength extraction method, determination method of spoilage benchmark, the representation method and the hyperspectral warning model are suitable and effective for early warning of banana spoilage.