Arti"cial neural networks are gaining widespread acceptance in cereal grain classi"cation and identi"cation tasks. The choice of a neural network architecture and input features can pose a problem for a novice user. This research is aimed at evaluating the most commonly used neural network architectures for cereal grain classi"cation using the frequently used morphological features as inputs. An evaluation of the classi"cation accuracy of nine di!erent neural network architectures was done to classify "ve di!erent kinds of cereal grains namely, Hard Red Spring (HRS) wheat, Canada Western Amber Durum (CWAD) wheat, barley, oats and rye. To evaluate the classi"cation accuracy of the di!erent neural network architectures, colour images of 7500 kernels (1500 kernels of each grain type) were taken. For each kernel, eight morphological features namely, area, perimeter, length of major axis, length of minor axis, elongation, roundness, Feret diameter and compactness were extracted and used as input to the neural networks. The networks were trained using 70% kernels for training and 20% kernels for validation of each grain type. Testing of the trained network was done on the remaining 10% kernels as well as the whole data set. The relative importance of the input features was also compared and the features that contributed the least to the classi"cation, were eliminated to decrease the complexity of the networks. The best results were obtained using a four-layer back-propagation network with each layer connected to the immediately previous layer. The classi"cation accuracies were in excess of 97% for HRS wheat, CWAD wheat and oats. The classi"cation accuracies for barley and rye were about 88%. The network required only four input features namely, Feret diameter area, minor axis length and compactness for classi"cation. A general regression neural network architecture was found to be the least suitable for grain classi"cation.