| Abstract | Nitrogen (N) fertilization is key for maintaining high and stable yields. However, the nitrogen use efficiency (NUE) of winter wheat (Triticum aestivum L.) is only 30-50%. To ensure sufficient yield, some farmers overfertilize their fields. The excess N causes pollution and an increase in costs. For optimal fertilization, the plant N status should be monitored. Most of the existing methods are destructive and time-consuming. Multispectral imaging is a promising technology to estimate the N content of crops by using the leaf spectral reflectance.
This study aims to assess the efficiency of multispectral imaging in the determination of N content in winter wheat leaves. A field trial was set up with a split-plot design in four replications. N-treatments, (80 kg N/ha), 300kg NPK applied at seeding + 130 kg of Urea (46% N) and N+ included additional 120kg calcium ammonium nitrate (27%N) split into three applications, representing a main plot. The subplot was represented as ten different wheat varieties. During a heading stage (GS50) ten flag leaves have been collected per N-treatment x wheat variety x replication. Multispectral images in six different wavelengths were acquired, and different vegetation indexes were calculated for each leaf. Total leaf N was determined using the Kjeldahl method. Data were analysed using a general linear regression model with multispectral data as a predictor and N leaf content as a response. In addition, machine learning techniques were used: Multiple Linear Regression, Random Forest Algorithm, Partial Least Squared Regression and Support Vector Regression as well Lasso-Regression.
The Highest relationship for leaf N content was found with an Anthocyanin Index (R2=0,779), Excess Green Index, (R2= 0,726 ), and Chlorophyll Index (R2=0,639) for Linear Regression. The best model to predict leaf nitrogen concentration was the Multiple Linear Regression with R2= 0,867 but it did not show any significant independent variables. The rest of the machine learning algorithms showed similar results: the Random Forest Algorithm (R2= 0,762), Partial Least Squared Regression (R2= 0,763), and Support Vector Regression (R2= 0,747).
This study shows that N content in leaves can be estimated using multispectral imaging and that ML can help to deal with high-dimensional datasets. Nevertheless, further data must be provided to develop a more accurate model. |
