بهینه سازی شبکه عصبی نیتروژن برگ ذرت حس شده از راه دور با برنامه ریزی خطی و الگوریتم ژنتیک با استفاده از پنج پارامتر عملکرد حس شده

An algorithm was developed to select an optimum model among several neural network (NN) models using the Manhattan and Euclidean metric measures. The algorithm was implemented to find an optimum NN prediction model based on simultaneous comparison of five performance parameters. Weighted coefficients were given to each performance parameter based on their significance for specific condition. The associated weighted coefficients were optimised using two optimisation techniques: (i) genetic algorithm; and (ii) linear programming. The algorithm performed satisfactorily in determining acceptable models and selecting an optimum NN model. The radial basis function NN model based on green vegetation index texture yielded an average prediction accuracy of 92·1% for predicting leaf nitrogen content under field conditions.

Leaf nitrogen is a critical component for the growth and development of crops. Assessment of leaf nitrogen may be useful to accurately determine the variation of crop nitrogen status. Leaf nitrogen affects crop growth and crop yield. Previous studies have examined the relationships between leaf nitrogen, crop growth, and yield (Beatty et al., 2000; Cihacek & Kerby, 1991; Flowers et al., 2000; Hong et al., 1997; Reeves et al., 1993). The laboratory-based conventional technique for determining leaf nitrogen is time consuming, expensive, and labour intensive. Field techniques, such as the use of hand-held chlorophyll meters are tedious (Reeves et al., 1993; Scharf & Lory, 2000). Thus, there is a need for developing an alternative technique to determine leaf nitrogen in the field. There are many approaches to the assessment of leaf nitrogen under field conditions in the literature. Leaf nitrogen was determined by quantifying nitrate concentration in the laboratory (Filella et al., 1995; Greenwood et al., 1991). Leaf nitrogen was also related to crop growing stage (Huggins & Pan, 1993; Zhuang & Bemard, 1990) and was reported to vary with ambient air temperature (Cihacek & Kerby, 1991; Sauchelli, 1964). Research was also conducted to establish a relation between chlorophyll and leaf nitrogen (Hong et al., 1997). The correlation between the chlorophyll and nitrogen content of rice leaves ranged between 0·90 and 0·94 (Hong et al., 1997). Leaf nitrogen was also determined using leaf canopy reflectance (Flowers et al., 2000; Reeves et al., 1993; Scharf & Lory, 2000; Viets & Hageman, 1971). The coefficient of determination between plant nitrogen spectral index and nitrogen uptake at different spectral bands (i.e. red at 761 nm and near infrared (NIR) at 780 nm) was found to be 0·67 ( Stone et al., 1996). Stone et al. (1996) found a correlation coefficient of 0·87 at 550 nm, which was higher than those at 430 and 680 nm wavelengths. Leaf chlorophyll and nitrogen of maize was determined using airborne hyper-spectral and infrared remote sensing ( Beatty et al., 2000). Beatty et al. (2000) found a more prominent variation of spectral response in the green and NIR bands. Leaf nitrogen of maize crop was also determined using aerial images ( Blackmer & White, 1998). A correlation coefficient of 0·74 was found between spectral reflectance (green at 550 nm and red at 710 nm) and leaf nitrogen. In a similar study, aerial images were used to compare the variability of leaf nitrogen deficiencies ( Blackmer et al., 1996). The red band image was reported to provide a higher correlation with nitrogen deficiencies. Variation of reflectance, transmittance, and absorption spectra of normal and nitrogen deficient maize leaves was separated by Al-Abbas et al. (1974). The potential of determining plant chlorophyll at different spectral bands of visible or NIR spectrum has previously been examined (Huggins & Pan, 1993; Zhuang & Bemard, 1990; Sauchelli, 1964; Stone et al., 1996; Blackmer & White, 1998; Blackmer et al., 1996). A strong relationship was also reported between plant chlorophyll and leaf nitrogen (Huggins & Pan, 1993; Stone et al., 1996). Nitrogen, when consumed by plants in the form of nitrate, converted to ammonia and subsequently to glutamine (Black, 1999). The glutamine is a key enzyme for the development of protein in the leaves (Black, 1999) and, thus, for the quantification of chlorophyll in plants. In addition, prior studies reveal that the chlorophyll affects the spectral properties of the light reflected from leaves (Beatty et al., 2000; Flowers et al., 2000). Therefore, spectral properties of plants in the form of images may contain information to predict leaf nitrogen. Hence, it was hypothesised that image information in the form of statistical and textural features can be used to predict leaf nitrogen in field conditions. The application of artificial intelligence techniques, such as neural networks (NN) is appropriate to develop prediction models, such as prediction of leaf nitrogen in field conditions (Sérélé et al., 2000; O’Neal et al., 2002; Tumbo et al., 2002; Yang et al., 1996; Pachepsky et al., 1996; Drummond et al., 2002). The use of this technique enables more complex data to be analysed, compared to other methods (e.g. statistical techniques), particularly when the feature space is complex and all data do not follow the same distribution pattern ( Benediktsson et al., 1990; Schalkoff, 1992; Atkinson & Tatnall, 1997). Moreover, the NN technique incorporates a priori knowledge and realistic physical constraints into the analysis ( Brown & Harris, 1994; Foody, 1995). Previous studies described the use of various NN architectures, such as multilayer perception ( Thai & Shewfelt, 1991; Foody, 1995; Tumbo et al., 2002; Pachepsky et al., 1996) and radial basis function ( Huang, 1999; Chen et al., 1991; Behloul et al., 2002). The selection of specific NN architecture depends on the nature of the problem and data type. This study focuses on the development of various NN models using two NN architectures; a multi-layer perceptron and a radial basis function. Moreover, the performance of NN models was evaluated based on simultaneous comparison of multiple parameters. In addition, an algorithm was developed to select acceptable models out of several NN models based on user-defined threshold criteria and then to determine an optimal NN model. Weighted coefficients were given to each performance parameter based on their significance. The optimal values of each weighted coefficient were determined using two optimisation techniques: (i) genetic algorithm; and (ii) linear programming.

Multi-spectral band images were processed in image-processing software to extract pertinent statistical and textural features for predicting leaf nitrogen content of maize under field conditions. Two additional images; normalised difference vegetation index (NDVI), and green vegetation index (GVI) were also derived using an appropriate transformation technique. These image features were used as inputs to develop 20 neural network (NN) prediction models based on multi-layer perceptron and radial basis function architectures. An algorithm based on the Manhattan and Euclidean metric measure was developed and the associated weighted coefficients were optimised using the genetic algorithm and linear programming. The generalised algorithm could select an optimum NN model based on simultaneous comparison of multiple parameters. For this study, the algorithm compared five parameters simultaneously to select first, acceptable NN models and then, an optimum NN model. The radial basis function NN model with five textural features derived from the GVI image was found to be the optimum model. This model provided an average prediction accuracy of 92·1%. It satisfied the user-defined threshold values for all five performance parameters. The study also showed that the textural image features provided better accuracies for predicting leaf nitrogen content than those provided by the simple statistical image features. The radial basis function NN provided higher performance in predicting leaf nitrogen as compared to the multi-layer perceptron NN architecture. When a NN prediction model is applied it is difficult to find the ideal performance. The simultaneous comparison of multiple performance parameters is advantageous for real-world applications. Moreover, comparing the performance of a given NN model with respect to user-defined threshold values of multiple parameters is practical. In this study, an algorithm was proposed and implemented that incorporated aforementioned hypothesis. The developed algorithm selected an optimum model based on the user-defined performance criteria. Further, weighted coefficients were applied to each performance parameter and these coefficients were optimised within the specified domain using two popular optimisation techniques, i.e. genetic algorithm (GA) and linear programming (LP). The algorithm was designed to be modified to use for other applications and can use any number of parameters.