Quantification of nitrogen stress on winter wheat by multispectral vision based on reflectance measurements and textural descriptors

Hand-held sensors (SPAD meter, N-Tester, …) are used for detecting the leaves nitrogen concentration (Nc) on the basis of an optical detection of the chlorophyll concentration. These devices are active sensors: an internal radiation source emits light and transmission through a leaf is measured in the red (650 nm) and in the near-infrared (920 nm) spectral regions. These devices present several drawbacks. The nitrogen concentration is gained by an indirect way through the chlorophyll concentration and the leaves have to be fixed in a defined position for the measurements. These drawbacks could be overcome by an imaging device that measures the canopy reflectance. In this context, the objective of the paper is to analyse the potential of different image parameters for estimating nitrogen concentration. The tests were carried out on parcels submitted to total nitrogen inputs of 180 kg N.ha-1 but with different fertilization modalities. Reference Nc measurements were obtained by the Kjeldahl method and a Hydro N-Tester (Yara). The developed imaging system comprised a CMos camera and a set of 22 interference filters ranging from 450 to 950 nm mounted on a wheel steered by a stepper motor. The image acquisition and the motor rotation were controlled by a program written in C++. The crop was imaged vertically at one meter height. The raw images presented 1280x1024 pixels covering an area of approximately 0.5x0.4 m and were recorded with a 12-bit luminance resolution. To deal with the natural irradiance variability of the scene, a white reference was used and the integration time was automatically adjusted for each image. The image treatment included the segmentation of Photosynthetically Active Leaves (PAL) by using Bayes theorem and the computation of the mean PAL reflectance after correction of background and illumination fluctuations. Nc was estimated on the basis of the 22 filters by Partial Least Square (PLS) method and by four filters selected by Best Subset Selection (BSS). In comparison with the Kjeldahl method, the estimation of Nc by the Hydro N-Tester, the PLS and the BSS (filters 600-80, 950-100, 650-40 and 450-80 nm) gave determination coefficient and standard error respectively equal to of 0.53, 0.29 %; 0.67, 0.21%; 0.56 and 0.25%. This indicated that the full multi-spectral approach gave significantly better Nc estimation than a portable device and suggested that a camera equipped with four filters would give similar results. In addition to these promising results, the Nc measurement were correlated to leaf area measurements and textural descriptors. For this purpose, image analysis based on Grey Level Co-occurrence Matrix (GLCM), Fourier transform and spatial autocorrelation were performed to characterize the nitrogen state of the crop