Computational intelligence image processing for precision farming on-site nitrogen analysis in plants

Abstract

Nitrogen is one of the macronutrients which is essentially required by plants. To support the precision farming, it is important to analyse nitrogen status in plants in order to prevent excessive fertilisation as well as to reduce production costs. Image-based analysis has been widely utilised to estimate nitrogen content in plants. Such research, however, is commonly conducted in a controlled environment with artificial lighting systems. This thesis proposes three novel computational intelligence systems to evaluate nitrogen status in wheat plants by analysing plant images captured on field and are subject to variation in lighting conditions. In the first proposed method, a fusion of regularised neural networks (NN) has been employed to normalise plant images based on the RGB colour of the 24-patch Macbeth colour checker. The colour normalisation results are then optimised using genetic algorithm (GA). The regularised neural network has also been effectively utilised to distinguish wheat leaves from other unwanted parts. This method gives improved results compared to the Otsu algorithm. Furthermore, several neural networks with different number of hidden layer nodes are combined using committee machines and optimised by GA to estimate nitrogen content. In the second proposed method, the utilisation of regularised NN has been replaced by deep sparse extreme learning machine (DSELM). In general the utilisation of DSELM in the three research steps is as effective as that of the developed regularised NN as proposed in the first method. However, the learning speed of DSELM is extremely faster than the regularised NN and the standard backpropagation multilayer perceptron (MLP). In the third proposed method, a novel approach has been developed to fine tune the colour normalisation based on the nutrient estimation errors and analyse the effect of genetic algorithm based global optimisation on the nitrogen estimation results. In this method, an ensemble of deep learning MLP (DL-MLP) has been employed in the three research steps, i.e. colour normalisation, image segmentation and nitrogen estimation. The performance of the three proposed methods has been compared with the intrusive SPAD meter and the results show that all the proposed methods are superior to the SPAD based estimation. The nutrient estimation errors of the proposed methods are less than 3%, while the error using the renowned SPAD meter method is 8.48%. As a comparison, nitrogen prediction using other methods, i.e. Kawashima greenness index (𝐺𝐼𝑘𝑎𝑤) and PCA-based greenness index (𝐺𝐼𝑃𝐶𝐴) are also calculated. The prediction errors by means of 𝐼𝑘𝑎𝑤 and 𝐼𝑃𝐶𝐴 methods are 9.84% and 9.20%, respectively.