Estimation of yield and quality of legume and grass mixtures using partial least squares and support vector machine analysis of spectral data

• Published in: Computers and electronics in agriculture 2019-07, Vol.162, p.246-253
• Main Authors: Zhou, Zhenjiang, Morel, Julien, Parsons, David, Kucheryavskiy, Sergey V., Gustavsson, Anne-Maj
• Format: Article
• Language: English
• Subjects: Agricultural Science; Calibration; Canopies; Clover; Commercialization; Datasets; Dry matter yield; Forage crop; Grass; Grasses; Hyperspectral reflectance; Jordbruksvetenskap; Least squares; Legumes; Nitrogen uptake; Nutritive value; Partial least squares; Red and white clover; Reflectance; Spectra; Spectral reflectance; Support vector machine; Support vector machines
• ISSN: 0168-1699; 1872-7107; 1872-7107
• DOI: 10.1016/j.compag.2019.03.038

•Yara-N oblique sensor successfully estimated N uptake in all developmental stages.•Yara-N oblique sensor can satisfactorily estimate forage yield and crude protein concentration.•SVM outperformed PLS for the estimation of forage dry matter yield and nutitional quality. The project aim was to estimate N uptake (Nup), dry matter yield (DMY) and crude protein concentration (CP) of forage crops both during typical harvest times and at a very early developmental stage. Canopy spectral reflectance of legume and grass mixtures was measured in Sweden using a commercialized radiometer (400–1000 nm range). In total, 377 plant samples were tested in-situ in different grass and legume mixtures (6 grass species and 2 clover species) across two years, two locations and five N rates. Two mathematical methods, namely partial least squares (PLS) and support vector machine (SVM) were used to build prediction models between Nup, DMY and CP, and canopy spectral reflectance. Of the total 377 samples, 251 were randomly selected and used for calibration, and the remaining 126 samples were used as an independent dataset for validation. Results showed that the performance of SVM was better than PLS (based on mean absolute error (MAE) for both calibration and validation datasets) for the estimation of all investigated variables. Results for the validation set showed that the MAEs of PLS and SVM for Nup estimation were 17 and 9.2 kg/ha, respectively. The MAEs of PLS and SVM for DMY estimation were 587 and 283 kg/ha, respectively. The MAEs of PLS and SVM for CP estimation were 2.8 and 1.8%, respectively. In addition, a subsample, which corresponded to an early developmental stage, was analysed separately with PLS and SVM as for the whole dataset. Results showed that SVM was better than PLS for the estimation of all investigated variables. The high performance of SVM to estimate legume and grass mixture N uptake and dry matter yield could provide support for varying management decisions including fertilization and timing of harvest.