Petar Dimitrov1, Ilina Kamenova1, Eugenia Roumenina1, Lachezar Filchev1, Iliana Ilieva1, Georgi Jelev1, Alexander Gikov1, Martin Banov2, Veneta Krasteva2, Viktor Kolchakov2, Milena Kercheva2, Emil Dimitrov2, Nevena Miteva2
1 Bulgarian Academy of Sciences, Space Research and Technology Institute, Department of Remote Sensing and GIS, 1113 Sofia, Bulgaria
2 Agricultural Academy, Institute of Soil Science, Agrotechnologies and Plant Protection “Nikola Poushkarov”, 1331 Sofia, Bulgaria
Dimitrov, P., Kamenova, I., Roumenina, E., Filchev, L., Ilieva, I., Jelev, G., Gikov, A., Banov, M., Krasteva, V., Kolchakov, V., Kercheva, M., Dimitrov. E., & Miteva. N. (2019) Estimation of biophysical and biochemical variables of winter wheat through Sentinel-2 vegetation indices. Bulgarian Journal of Agricultural Science, 25(5), 819–832
Traditionally, the growth and physiological status of winter wheat (Triticum aestivum L.) is monitored in the field by measuring different biophysical and biochemical variables such as Above Ground Biomass (AGB), Nitrogen content (N), N uptake, Leaf Area Index (LAI), Fraction of vegetation Cover (fCover), Canopy Chlorophyll Content (CCC), and fraction of Absorbed Photosynthetically Active Radiation (fAPAR). The objective of this study was to investigate the possibility of estimating these crop variables through statistical regression modelling and spectral vegetation indices derived by the Sentinel-2 satellites. Field data were collected over two growing seasons, 2016/2017 and 2017/2018, in test fields around Knezha, northern Bulgaria. A combination of spectral data from Sentinel-2 images and field spectroscopy obtained through the first growing season was used for model calibration and cross-validation. The models were further validated with Sentinel-2 image data from the second growing season. The accuracy of the models varied widely across crop variables. According to the cross-validation, the relative RMSE was below 25% for fAPAR, fCover, and fresh AGB, with particularly good result for fAPAR (13%). For N content and dry AGB the error was between 25% and 30%. The accuracy was low for CCC, LAI, and N uptake (error between 30% and 43%). The models’ performance was worse when they were applied to the data from the second growing season, resulting in relative RMSE which were 3-8% higher in the general case. The cross-validation results suggested that the variety-specific models are more accurate than the generally calibrated models for most crop variables. The accuracy obtained in this study for the prediction of fAPAR, fCover and AGBf through VIs is promising. Future studies and incorporation of new field data will be needed to better account for variety, season, and site variations in the modelled relationships and to improve their generalisation potential.