Feasibility study of BP neural network and probabilistic neural network for nitrogen nutrition diagnosis of rice images

Objectives Identifying key indicators for image and establishing models for processing of massive image data rapidly were two main steps in achieving nitrogen nutrition diagnosis. This project screened the sensitive period and location of rice nitrogen nutrition diagnosis, optimized the image processing technical parameters, and compared the reliability between two methods, BP neural network and probabilistic neural network, in nutrient diagnosis, which provided ideas and methods for determining the nutritional status of crops and inverting the growth process quickly and accurately by using computer vision virtual technology.

Methods In this study, super hybrid rice ‘LYP9’ was used as experimental crop to set up four kinds of rice cultivation experiments at different fertilization levels (equivalent to N 0, 210, 300 and 390 kg/hm²), the image data of a total of 1920 groups of the first leaves and the second leaves, the third leaves from crop top, and their corresponding sheaths were obtained by scanning with a scanner during the panicle differentiation stage and the full heading stage. Nineteen rice indexes were obtained. The diagnosis models of rice nitrogen nutrition on standardized nineteen rice characteristic indexes obtained by image processing were respectively established by applying BP neural network and probabilistic neural network. Based on the models, rice nitrogen nutrition diagnosis and identification were carried out.

Results 1) The accuracy of overall recognition of rice at the panicle differentiation stage was higher than that of the full heading stage; the image data of the third leaves was the most reliable among the three parts of leaves; 2) For the panicle differentiation stage and the full heading stage, the overall recognition accuracy of the nineteen rice characteristic indexes obtained by image processing was higher by using the BP neural network, compared with the probabilistic neural network. The accuracy of diagnosis and identification of rice nitrogen nutrition in the nineteen rice characteristic indexes of the third leaves at the panicle differentiation stage processed through the BP neural network was up to 90%. The accuracy of diagnosis and identification of rice nitrogen nutrition in the nineteen rice characteristic indexes of the second leaves and the third leaves at the panicle differentiation stage processed through the probabilistic neural network was up to 82%.

Conclusions The leaf characteristics of the third leaves in the panicle differentiation stage are the most distinguishable, and it is easy to diagnose and identify nitrogen nutrition, which can be used as an effective period and location for nitrogen nutrition diagnosis. The components of six color space, RGB and HSI, can best reflect nitrogen nutrition status. Regarding the recognition effect, the BP neural network is higher than the probabilistic neural network, and its overall recognition accuracy is 90%.