In order to achieve rapid detection of the contents of four physicochemical indicators related to freshness of mutton, a multi-task spectral feature extraction algorithm named ERC-RF is proposed in this study. By integrating the Ensemble Regression Chain (ERC) and Random Forest (RF), the correlation between freshness indicators was used to assigns varying weights to the characteristic spectra to ascertain the importance of spectral band features. During the experiments, data were first preprocessed by combining the SG (Savitzky-Golay) smoothing filter and Multiplicative Scatter Correction (MSC). Subsequently, the RF and ERC-RF models were used to select characteristic bands, respectively, and the spectral data were used for two freshness indicator contents prediction models: models for integrating Support Vector Regressio (SVR) with Multi-Output Regressor (MOR): SVR-MOR, and integrating SVR with Multi-Output Regression Chain (MORC): SVRMORC. The results showed that the Coefficient of Determination (R2), average Relative Root Mean Squared Error (aRRMSE), and Mean Absolute Error (MAE) values of the spectral feature data extracted by ERC-RF were 0.974 5, 1.302 7, 0.385 2, and 0.975 2, 1.495 6, 0.391 8, respectively in the two prediction models, which were significantly improved compared with models based on other band selection, with the number of bands being reduced from 125 to 31. These results confirm that ERC-RF is an effective multi-task feature wavelength screening method, which not only improves the accuracy of multioutput regression model, but also reduces the model complexity.