Crop nitrogen trait estimations have been used for decades in the frame of precision agriculture and phenotyping researches. They are crucial information towards a sustainable agriculture and efficient use of resources. Remote sensing approaches are currently accurate tools for nitrogen trait estimations. They are usually quantified through a parametric regression between remote sensing data and the ground truth. For instance, chlorophyll or nitrogen concentration are accurately estimated using features like vegetation indices. However, those models tend to simplify the observed scene in averaging the features. Thus, they encompass potential other information as the distribution of the concerned trait and tend to under exploit the data.

In this study, we have modelled organ scale nitrogen maps from multi-spectral camera array images mounted on a ground-based platform. Two wheat fertilisation trials were imaged for two years. Reference nitrogen concentration measurements were performed in the laboratory, by taking care to separate the flag leaves from the other leaves. As we had a single or two nitrogen measurements per image, the following idea was to exploit the machine learning algorithm capacity to generalise and find patterns. Therefore, we drew several small leaf patches over the images of different nitrogen input objects and growth stages. Patches on the flag leaves and on the other leaves were associated with the corresponding nitrogen measurement. Three machine learning algorithms were trained and tested using features from the six bands of these patches. The arising patterns, created by the nitrogen changes over the season and nitrogen objects, were used by the model to generate nitrogen maps at organ scale. It also provided the nitrogen concentration distribution of the observed scene. The k-nearest neighbours model was able to predict the leaf and the flag leaf nitrogen concentration with a relative RMSE of 0.15 and 0.11, respectively, using the average and the percentile 80 of the map, respectively. Moreover, the nitrogen map can be fused with a height map, for example measured by stereovision, to study the distribution of nitrogen concentration as a function of the height of the leaf in the canopy.