9  Conclusion

In this thesis, I delved into the remote sensing of foliar nitrogen concentration in Californian almond trees. The various methodologies explored can be broadly categorized into three groups: hand-crafted features, machine learning, and radiative transfer models Historically, hand-crafted features, especially single vegetation indices, have been the most prevalently employed approach for modelling crop nitrogen based on reflectance data.

Nonetheless, the findings of this thesis reveal that vegetation indices fall short in comparison to the other examined techniques. Machine learning algorithms on the other hand typically perform well on datasets they have been trained on; however, they often lack generalizability due to their excessive adaptability and absence of prior domain knowledge utilization. Radiative transfer models offer a promising solution to these limitations. They provide a generalizable method for estimating plant characteristics through remote sensing, which can subsequently facilitate the prediction of crop nitrogen.

Radiative transfer models typically present the challenge model inversion. Solving the model inversion can be accomplished through mathematical optimization, utilization of look-up tables, or the application of machine learning techniques. In this thesis, I have demonstrated that employing a multicubic polynomial surrogate model for the cost function can dramatically enhance the predictions derived from a look-up table almost to the level of the best-performing machine learning models.

The slightly superior predictions generated by machine learning models suggest that there is some room for improvement regarding the radiative transfer models. Possible enhancements include examining chlorophyll a and b independently at the leaf-level or improving the representation of the three-dimensional structure of crops at the canopy-level.

Regarding spectral data, the span of the covered range proves to be more crucial than the spectral resolution. Specifically, the short-wave-infrared region appear to be particularly vital for improving the precision of crop nitrogen estimation.

A considerable body of research on remote sensing of crop nitrogen has been conducted in conjunction with nitrogen fertilization trials. The artificially heightened range of the measured nitrogen concentrations may inadvertently result in overestimating the accuracy of various models, particularly when numerous observations with low foliar nitrogen concentration are present in the dataset. Consequently, further research aimed at addressing the issue of nitrogen leaching through remote sensing ought to focus on experimental settings that more closely resemble standard practices in commercial agriculture.