Multiple instance regression for the estimation of leaf nutrient content in olive trees using multispectral data taken with UAVs.

The rational fertilisation of olive trees, based on adding exclusively the nutrients that are actually needed, is important from both the economic and environmental sustainability points of view. This paper employs UAV-obtained multispectral data collected from five different orchards located in Southern Spain to build a set of models for the prediction of the leaf nutrient content of olive trees using Support Vector Regression. The paper shows the convenience of addressing the problem as a Multiple Instance Regression, and compares two strategies of data aggregation and different choices of feature vectors derived from the raw multispectral data. The models provided good results for N, P and K (r 2 = 0.76, r 2 = 0.87 and r 2 = 0.91, respectively for the Hojiblanca model, and r 2 = 0.79, r 2 = 0.80 and r 2 = 0.80 for the Picual model). The rest of nutrients studied also offered good results for both the Picual and Hojiblanca models, ranging from r 2 = 0.69 for B to r 2 = 0.93 for Cu. The results indicate a robust performance of the models and a potential for improvement with the addition of more data, along with an advantage of considering individual models for each cultivar variety. Overall, these results are very promising for the estimation of the leaf nutrient content of olives trees and the detection of spatial variability in the fertilisation needs of orchards. [Display omitted] • Leaf nutrient content of olive trees is predicted from UAV multispectral data. • Multiple Instance Regression is applied and its benefits discussed. • Models provide good results for N, P and K in the Hojiblanca and Picual models. • The rest of nutrients also show acceptable results. • Results are promising for detecting spatial variability in fertilisation needs. [ABSTRACT FROM AUTHOR]