Spectral monitoring of wheat leaf nitrogen content based on canopy structure information compensation

Abstract

Nitrogen is an important nutrient element for crop growth. A timely understanding of plant nitrogen information helps to adopt appropriate agricultural production management to maintain high yield and quality in wheat production. Crop spectral monitoring technology can obtain the leaf nitrogen content (LNC) information of wheat quickly and nondestructively. However, optical radiation interacts with the atmosphere, the canopy, and the soil before being captured by the sensor, and the capability to intercept, reflect, and transmit the radiation is different for different canopy structures. In wheat LNC monitoring, differences in target canopy structures will lead to changes in canopy reflectance, which can affect the monitoring accuracy. In this study, RGB and depth images were used to obtain wheat canopy structure indices. By analyzing the correlations between wheat LNC, canopy structure indices, and spectral reflectance, the indices that had large influence on spectral reflectance were screened; the change dynamics of these canopy structure indices under different internal and external factors were compared, and factors with greater impact were selected as the basis for grouping. On this basis, this study used the spectral indices RVI (660, 815) and RVI (730, 815) as fixed-effect variables and wheat population grouping variables as random-effect variables to construct linear mixed models of wheat LNC. In addition, the spectral indices and canopy structure indices were used as input parameters and wheat population grouping variables as output parameters to construct a random forest classifier for wheat canopy types. When monitoring the target population, we first predicted the classification of the unknown canopy by the classifier. The classification results and spectral indices were then input into the linear mixed models to realize the prediction of wheat LNC. The R2 of the prediction for wheat LNC with RVI (660, 815) and RVI (730, 815) increased from 0.57 and 0.71 to 0.76 and 0.80, respectively, and the RRMSE decreased from 20.86% and 17.33% to 15.58% and 14.20%, respectively. This study used digital image information to compensate for spectral information, broadened the amount of information used for the remote sensing monitoring of farmland, and effectively improved the accuracy and universality of wheat LNC monitoring.