Using Mathematical Models to Identify the Chlorophyll Status of Wheat from Ordinary Digital Camera Images in the Field

Abstract

Listen

Translate article

ABSTRACT The chlorophyll content of field crops is directly indicative of their health and fertilizer requirements. However, existing approaches for assessing chlorophyll content are either laboratory-bound or very costly. The future prospects for developing effective variable-rate fertilizer machines therefore depend upon the development of better ways for assessing the chlorophyll status of field crops. This paper looks at the scope for doing this by using digital camera images of field scenes. Images of a wheat crop canopy were taken in the field by a digital camera at different shooting angles. These were then subjected to correlation analysis in relation to physical measurements of the chlorophyll content to identify a set of the most effective features for nutritional diagnosis from multiple color spaces. The shooting angle had an effect on these features, so an optimal angle needed to be determined. Correlation-based feature selection (CFS) in association with three different algorithms – Ordinary Least Square Regression (OLS) Random Forest (RF) and a Multi-Layer Perceptron neural network (MLP) – was used to develop multivariable models. In fact, tests suggest that the prediction accuracy of best feature-selection models is better than that of all-feature models. Thus, it is necessary to establish which mathematical models are best-suited to identifying the most effective features to work with. Out of all the tested models, an MLP model using just 8 input features had the highest prediction accuracy (R2 = 0.822, RMSE = 3.072).