Classification of hard red wheat by feedforward backpropagation neural networks

Abstract

Cereal Chem. 72(3):317-319 Because of breeding practices, visual determination of kernel morphology is becoming less dependable for classification of hard red wheat as winter or spring. Because of the price differential between these classes, it is important to develop rapid, accurate, and automatable alternative methods. This study was conducted to determine whether feedforward backpropagation neural networks applied to near-infrared (NIR) diffuse reflectance spectra of ground kernels could perform the classification. The NIR diffuse reflectance spectra (1,100-2,500 nm) of 2,403 ground hard red wheat samples representing the United States crop for 1987-1990 were used with commercially available neural network software. Mathematical pretreatments included logO( I/ reflectance) and second differences of the log data. Networks with and without hidden layers were used with various subsets of the full spectral region as inputs. When developed on samples from the 1987-1989 crop years, the best neural network models yielded 97.0 and 96.8% accuracies for calibration and validation sets, respectively, utilizing the full wavelength range. This performance declined slightly to calibration and validation accuracies of 96.3 and 95.9%, respectively, when the wavelength range of 2,142-2,472 nm was used. When applied to the 1990 crop year, the prediction accuracies of the full and abbreviated wavelength range models were 95.1 and 95.6%, respectively. These models performed better than a previously reported principal component analysis with Mahalanobis distance classifier. Neural networks, combined with second difference pretreatment, should be a very useful component of a NIR-based classification system. In the United States, classification of wheat has become increasingly difficult because of an increasing number of cultivars per class, more overlapping of growing regions of different classes, and more crossbreeding between cultivars belonging to two or more classes. Discrimination between hard red winter (HRW) and hard red spring (HRS) wheat is of particular importance, because of the volume of trade and the price differentials between these classes. The knowledge and experience required to accurately perform this classification is becoming too complicated for grain inspectors conducting visual inspection based on kernel morphology. Instrumentation that could classify wheat rapidly and with little training would be very useful to federal grain inspectors, traders, and millers. We have been conducting research to develop techniques to differentiate these two classes by near-infrared (NIR) diffuse reflectance spectroscopy. The advantages of the NIR method are that it is rapid, does not require much sample preparation, and can be used in field measurements. Delwiche and Norris (1993) used NIR spectra of ground wheat to compare various discriminant analysis models. They calibrated the models with 1987-1989 crop samples and found that a five-factor principal component analysis with Mahalanobis Distance (PCA/ MD) classifier was most accurate. The classification rate was 95% when the model was validated on the 1987-89 samples not included in the training set, and 92% when predicting a set of the 1990 samples. Artificial neural networks are widely applied to pattern recognition problems. Examples are optical character recognition, image classification, target recognition, and speech recognition. Recently, they were found to be effective using spectral data to classify poultry for quality control on processing lines (Chen 1992, Park and Chen 1993) and to classify undamaged and damaged peanut kernels (Dowell 1994). Neural networks have some potential advantage over previously reported mathematical classification methods because they are able to discover and use nonlinear rela'Research leader, agricultural engineer, and mathematician, respectively, Instrumentation and Sensing Laboratory, Beltsville Agricultural Research Center, ARS, USDA, Beltsville, MD. Mention of a tradename or propietary product does not constitute a guarante or warranty by the USDA and does not imply its approval to the exclusion of other products that may be suitable.