Predicting leaf nitrogen content in olive trees using hyperspectral data for precision agriculture

Olive orchard is one of the main crops in the Mediterranean basin and, particularly, in Spain, with 56% of European production. In semi-arid regions, nitrogen (N) is the main limiting factor of olive trees after water and its quantification is essential to carry out accurate fertilization planning. In the present study, N status of an olive orchard located in Carmonita (southwest Spain) was analysed using hyperspectral data. Reflectance data were recorded with a high precision spectro-radiometer through the full spectrum (350-2500 nm). Different vegetation indices (VI), combining two or three wavelengths, and partial least squares regression (PLSR) models were developed, and the prediction capabilities were compared. Different pre-processing (smoothing, SM; standard normal variate, SNV; first and second derivative) were applied to analyse the influence of the noise generated by the spectro-radiometer measurements when computing the determination coefficient between leaf N content (LNC) and spectra data. Results showed that second derivative combined with SNV pre-processing produced the best determination coefficients. The wavelengths most sensitive to N variation used to perform VI were selected from the visible and the short-wave infrared spectrum regions, which relate to chlorophylla + band N absorption features. DCNI and TCARI showed the best fittings for the LNC prediction (R-2 = 0.72, R-cv(2) = 0.71; and R-2 = 0.64, R-cv(2) = 0.63, respectively). PLSR models yielded higher accuracy than the models based on VI (R-2 = 0.98, R-cv(2) = 0.56), although the large difference between calibration and cross-validation showed more uncertainty in the PLSR models.

Automated summary

This study analyzed the nitrogen status of an olive orchard in Carmonita, Spain using hyperspectral data. The results showed that the second derivative combined with standard normal variate preprocessing produced the best determination coefficients, and partial least squares regression models yielded higher accuracy compared to models based on vegetation indices.