Classification of rice plant nitrogen nutrient status using k-nearest neighbors (k-NN) with light intensity data

Abstract

Crop management including the efficient use of nitrogen (N) fertilizer is important to ensure crop productivity. Human error in judging the leaf greenness when using the leaf color chart (LCC) to estimate the rice plant N nutrient status has encouraged numerous researchers to implement a machine-learning algorithm but experienced some issues in calibration and lighting. The datasets are created at 6.00-7.00AM (consistent lighting) and including light intensity, so each dataset contains RGB value and light intensity as inputs, and LCC value as a target. A system consists of a smartphone with an application that prevents user from taking an image if the light intensity is not in 2000-3500 lux, and a computer for preprocessing and classification purposes were developed. The preprocessing included cropping, splitting the rice leaf images, and calculating the average RGB values. A k-NN classifier is implemented and by using a cross-validation method is found k=5 gives the best accuracy of 97,22%. The in-site test of the system also works with an accuracy of 96.40%.