Determination of the optimum ripeness of New Rice for Africa (NERICA) using Alexnet

Abstract

The New Rice for Africa (NERICA) is a product of hybridizing African rice (Oryza glaberrima Stued) and Asian rice (Oryza sativa L.) in order to produce an improved variety of rice with very high yield, good resistance to draught, resistant to phosphorus deficiency and striga conditions. The research attempts to address the problem of food shortage as a result of increasing population and Draught in sub-Sahara Africa. The major two varieties of NERICA are Lowland and Upland varieties. One of the most acceptable varieties for uplands is NERICA-4, having delicious taste when compared to the other upland varieties. However, NERICA-4 suffers loss of grains during harvest, which leads to low productivity. In this paper, 75 images of different ripening stages were obtained from a cultivated farm land with NERICA-4 rice variety. The images were resized to a dimension of 277X277 with three channels to suit the input layer of Alexnet pre-trained network. The images are augmented for classification. The stochastic gradient descend with momentum training option is used with initial learning rate of 0.0001 for the settings. The training set achieved 100% classification accuracy from iteration 10 to 20 and zero loss at the 20^th epoch. The test set obtained 87.9% classification accuracy which gave 15.46% improvement over an earlier classification accuracy using Alexnet for classifying maize comb images. The loss of grains can be reduced and yield increased when made operational.