Condition assessment serves as the foundation for studying the deterioration mechanisms of stone heritage and formulating conservation measures. However, traditional photogrammetry-based identification methods for heritage deterioration often suffer from excessive subjectivity, over-reliance on engineer experience, and low efficiency. To address these issues, this study proposes an intelligent identification method based on hyperspectral imaging technology, using the rock carvings at Fengshan Temple in Dazu as a case study. A hyperspectral imaging system covering the 400-1000 nm wavelength range is employed. Signal-to-noise ratio is enhanced through preprocessing techniques including Savitzky-Golay smoothing filter, normalization, and continuum removal. To reduce spectral redundancy, key wavelength features are extracted by using Competitive Adaptive Reweighted Sampling (CARS), Random Frog Leap Algorithm (RFLA), and Successive Projections Algorithm (SPA). The performance of Random Forest (RF) classification models incorporating these features is compared. Results show that spectral feature separability has significantly improved after preprocessing, with a noise reduction rate of 89%. The CARS-RF model achieves the best overall performance. It can accurately distinguish deterioration types such as flaking and powdering exhibiting lower boundary recognition errors in micro-regions compared to other models, and its identification results are highly consistent with field survey findings. Quantitative analysis reveals that mild weathering is the predominant deterioration type, accounting for 31.71%, followed by powdering (28.65%) and flaking (23.02%). Painted residue weathering is least prevalent, at only 16.62%. These proportions closely match the field investigation records. This method provides robust technical support for the digital conservation of stone heritage, and its hyperspectral feature extraction strategy can be extended to the inspection of similar cultural heritage.