Comparative analysis of vegetation indices, non-parametric and physical retrieval methods for monitoring nitrogen in wheat using UAV-based multispectral imagery

Unmanned Aerial Vehicles (UAV)-based remote sensing offers great possibilities to acquire in a fast and convenient way field data for precision agriculture applications. The UAV-based multispectral images with five wavebands (490, 550, 671, 700, 800 nm) were obtained at five growth stages from consecutive two years' wheat field experiments with different combinations in variety, N application rate and planting density. In this paper, we compared systematically leaf nitrogen content (LNC) estimation accuracy and processing speed of a multitude of vegetation indices (VIs), non-parametric and physically-based model retrieval methods. With regard to vegetation indices, most possible band combinations with existing five bands and a linear regression fitting function have been evaluated. The best performing index was optimized three-band combination 2.5(R
800
-R
700
)/(R
800
+6R
700
-7.5R
490
) according to EVI (Enhanced vegetation index) for LNC with 10-fold cross-validation determination (R2) of 0.73. This method shows especially fast processing speed (0.03 s). As for non-parametric methods, 14 common regression algorithms have been evaluated. Among these algorithms, Random Forest [TreeBagger] is the best performing method with R
²
of 0.79 for LNC. This method has the advantage of making use of the full optical spectrum as well as flexible, nonlinear fitting. Additionally, the model is trained and validated relatively fast (2.3 s). Compared with the front two methods, it remains a challenge to estimate the canopy nitrogen through inversion of a PROSAIL based radiative transfer model (RTM). After the generation of a look-up table (LUT), a number of cost functions and regularization options were evaluated for inversion of LCC (leaf chlorophyll content). Inversion of nitrogen indirectly relying on inversion of LCC based on the empirical linear relationship between LCC and LNC. Although this method offered per-pixel estimation, generation of a look-up table and image processing took considerably more time. Besides, the validation performed less reliable. To sum up, Random Forest [TreeBagger] provides fast and accurate estimation of nitrogen using UAV-acquired multispectral imagery which will be good basis for remote sensing of canopy nitrogen status in a wide range of crop.