Identification and validation of key wavelengths for on-line beef tenderness forecasting

The objective of this study was to identify and validate key wavelengths in the visible/near-infrared (VNIR) spectral region for beef tenderness assessment using hyperspectral images of beef carcasses acquired on-line in a commercial beef packing plant. A prototype on-line spectrograph hyperspectral imaging (HSI) system was developed and used to acquire images of exposed ribeye muscle on hanging beef carcasses (n = 274) at 2 days post-mortem. After image acquisition, a strip steak was cut from each carcass and vacuum packaged. After aging for 14 days, the steaks were cooked and Warner-Bratzler shear force values were collected as a measure of tenderness. Hyperspectral and multispectral image analysis approaches were conducted and compared to each other in terms of accuracy and speed of implementation. In the hyperspectral approach, principal component analysis (PCA) was conducted to reduce the spectral dimension of the hyperspectral images and create principal component (PC) bands. Image textural features were extracted from the PC bands and modeled to classify samples into two tenderness categories (tender or tough). In the multispectral approach, the loading vectors obtained from the PCA procedure were used to identify five key wavelengths (584, 619, 739, 832, and 950 nm). Image textural features were then extracted from these key wavelength images and modeled. In both approaches, the image features were extracted from the images acquired at 2 days post-mortem and used to forecast 14-day beef tenderness. Using a true validation procedure involving 101 beef samples, the hyperspectral approach had a tender certification accuracy, tender identification accuracy, tough identification accuracy, and accuracy index value of 86.7%, 65.8%, 63.6, and 66.8%, respectively. The corresponding metrics for the multispectral approach were 87.5%, 62%, 68.2%, and 67.8%, respectively. The tenderness prediction results for these two analysis methods were not considerably different; however, multispectral image analysis took far less time (about 8 s compared to 105 s on a quad-core computer with 8 GB of memory) to classify a beef sample into one of the two tenderness categories (tender or tough). Therefore, the multispectral approach shows great potential for real-time beef tenderness assessment in commercial beef packing plants.