Prediction of Cooking Time for Soaked and Unsoaked Dry Beans (Phaseolus vulgaris L.) Using Hyperspectral Imaging Technology

The cooking time of dry bean varies widely by genotype and is also influenced by the growing environment, storage conditions, and cooking method. Thus high-throughput phenotyping methods to assess cooking time would be useful to breeders interested in developing cultivars with a desired cooking time. The objective of this study was to evaluate the performance of hyperspectral imaging technology for predicting dry bean cooking time. Fourteen dry bean genotypes with a wide range of cooking times were grown in five environments over 2 yr. Hyperspectral images were taken from whole dry seeds, and partial least squares regression models based on the extracted hyperspectral image features were developed to predict water uptake and cooking time of soaked and unsoaked beans. Relatively good predictions of water uptake were obtained, as measured by the correlation coefficient for prediction (Rₚᵣₑd = 0.789) and standard error of prediction (SEP = 4.4%). Good predictions of cooking time for soaked beans (ranging between 19.9–95.5 min) were achieved giving Rₚᵣₑd = 0.886 and SEP = 7.9 min. The prediction models for the cooking time of unsoaked beans (ranging between 80–147 min) were less robust and accurate (Rₚᵣₑd = 0.708, SEP = 10.6 min). This study demonstrated that hyperspectral imaging technology has potential for providing a nondestructive, simple, fast, and economical means for estimating the water uptake and cooking time of dry bean. Moreover, a totally independent set of 110 similar dry bean samples confirmed the suitability of the technique for predicting cooking time of soaked beans after updating the partial least squares model with 20 of the new samples, giving Rₚᵣₑd = 0.872 and SEP = 3.7 min. However, due to the genotypic and phenotypic variability of water absorption and cooking time in dry bean, periodical updates of these prediction models with more samples and new bean accessions, as well as testing other multivariate prediction methods, are needed for further improving model robustness and generalization.