Intelligent Precision Nitrogen Fertilizer Application Based on Speaking Plant Approach for Environmental Sustainability

Abstract

In vegetable farming, one of important issues is determining the right dose application of nitrogen fertilizer. The advantages and disadvantages of nitrogen fertilizer in spinach (Amaranthus sp.) will have an adverse effect on its productivity. The total nitrogen measurement by chemical analysis is too expensive, inefficient, and cannot be applied directly. The use of digital image analysis and intelligent modelling such as artificial neural network (ANN) can provide a real-time and accurate solution to predict the content of nitrogen in spinach leaf. This study aims to model the relationship between texture parameter based on colour co-occurrence matrix (CCM) and nitrogen content in spinach leaves. The texture analysis consists of 40 CCM textural features derived from RGB and grey colour. From the 40 CCM textural features, some of the best CCM textural features are selected to be used as ANN inputs in predicting nitrogen content. The best parameter selection method applying two approaches are the filter method including: (1) correlation-based feature selection; (2) correlation attribute evaluation; (3) linear regression; and (4) relief attribute evaluation and the wrapper method which is neural-genetic algorithm (N-GA). The best parameter selection result in the filter method based on the validation result is correlation-based feature selection (using 10 CCM textural features, training MSE = 0.0028; validation MSE = 0.00016; testing accuracy R2 = 0.96). However, when compared to all filter methods, the wrapper method using N-GA still demonstrates better results (using 8 CCM textural features, training MSE = 0.00039; validation MSE = 0.000038; testing accuracy R2 = 0.993). From CCM textural features which have been selected, it can be used as input ANN to predict the nitrogen content of spinach leaves accurately. The best ANN structure has been built by using 1 input layer (8 inputs), 1 hidden layer (20 nodes), and 1 output layer (1 node).