Visible/near infrared (Vis—NIR) spectroscopy was investigated to determine the dry weight of the cocoons layer of mulberry silkworm fresh cocoons. Optimal partial least squares (PLS) models were developed with different preprocessing, and the data preprocessed by Savitzky—Golay (SG) smoothing was chosen for the effective wavelengths selection. The selection was operated based on regression coefficients in PLS models, and reduced the original 601 varieties into 7. Then multiple linear regression (MLR) was used for calibration and prediction based on the seven effective wavelengths, compared with the PLS model built on full-spectrum data. The results showed that MLR model was the optimum model for the dry weight of the cocoons layer detection in the process of production and marketing, because of its simple arithmetic and accurate detection. The correlation coefficient and residual predictive deviation were 0.7587 and 2.0464.