Modeling For Rapeseed's Leaf Nitrogen Nutrient Diagnosis Based On Multifractal Detrended Moving Average Analysis

Nutrient diagnosis is the base of fertilization guidance. Leaves, as the one of most important organ of crop to absorb nutrients, can reflect the crop growth conditions through the digital images based on mathematics and computers. Once the crop is detected nutrient deficiency, their color, shape and texture feature of leaf will be changed. However Current studies are focused on the influence of leaf color and shape characteristics, few studies on the texture information. As an inherent nature and characteristics of leaf, the leafs texture remains relatively stable in the crop growth than color and shape. That is because the color and shape are more susceptible to the environment influence. So the texture feature of leaves is a good choice to be used for expressing nutritional status. Fractal, especially multi-fractal, as a effective tools to describe the texture feature of images, which is widely used in the filed of image processing. In order to identify and diagnose the texture features of rape leaf under different nitrogen levels, eleven kinds of generalized Hurst index and other six kinds of related multifractal characteristic parameters of the rape leaf images were calculated by using multifractal detrended moving average analysis (MF-DMA) with three key position parameters, namely,?=0,0.5 and 1, respectively. When we has extracted the texture information of crop, fisher's linear discriminant, extreme learning machine, support vector machine, BP-NN, random forests and K-nearest neighbor algorithm were proposed as the models for rapeseed's leaf nitrogen nutrient diagnosis based on multifractal detrended moving average analysis. By applying different combinations of characteristic parameters, the nitrogen nutrition diagnosis and recognition were conducted for the base leaf, central leaf and top leaf, respectively. The results showed that the performance of diagnosis with position parameter of ?=0 is better than that with 0.5 and 1. In addition, the best diagnose accuracy came from the base leaf and the central leaf, which demonstrated that the base leaf and the central leaf were more sensitive to the nitrogen deficiency than the top leaf. By diagnosing for the nitrogen deficiency and nitrogen moderately of the three parts of the mixed rape leaf samples, it showed that support vector machines and kernel method (SVMKM) and the random forest were the best two methods to obtain the accuracy, by which the best recognition accuracy rate reached to 95.81% and 96.63%, respectively. It expressed that our model possessed good effectiveness.