Mutton freshness is affected by many factors, and the detection is generally carried out based on sensory properties, physical and chemical products of decomposition, microbial reproduction and other aspects. However, the freshness detection of mutton based on a single indicator has great limitations and low applicability, and it is not easy to evaluate the mutton freshness comprehensively. Moreover, the traditional detection methods are complex and inefficient, which cannot meet the daily actual needs. As a fast, nondestructive and efficient intelligent detection technology, hyperspectral imaging technology can effectively collect the surface, internal composition and physical and chemical changes in the process of mutton putrefaction. This paper proposes an evaluation model for the freshness of chilled mutton based on the improved deep forest algorithm, which adds feature screening to mine the spectral information related to multiple evaluation indicators. It adds layer growth control to prevent the model from over fitting effectively. This paper collects 400~1 000 nm hyperspectral data of mutton samples stored at 4 ℃ for 0~14 days .Total volatile base nitrogen (TVB-N), pH, total aerobic plate count (TAC) and the approximate number of coliforms (ANC) indicator values are measured by laboratory methods. The representative spectra of mutton samples are extracted in the regions of interest. The original spectral data is preprocessed using the smoothing filtering and multivariate scattering correction methods. 18 spectral feature bands are extracted by using the continuous projection method, and the samples of the training set and the testing set are divided at a ratio of 3∶1. Establishment of the freshness classification model is used in the improved deep forest algorithm proposed in this paper. The results show that the overall accuracy of freshness classification is 0.985 7, and use Hamming loss, One-error, Ranking loss and Marco-AUC multi-label metrics to evaluate the performance of the model, which are 0.025 7, 0.014 3, 0.014 2 and 0.998 6 respectively. Theyare better than the traditional multi-label classification algorithm. The research shows that the multi-indicator freshness classification model can be used for rapid, nondestructive testing of mutton freshness. It improves the limitations of single-indicator model classification and provides a research method for multi-indicator nondestructive testing of subsequent hyperspectral imaging technology.