UNFOLD PRINCIPAL COMPONENT ANALYSIS AND FUNCTIONAL UNFOLD PRINCIPAL COMPONENT ANALYSIS FOR ONLINE PLANT STRESS DETECTION

Abstract

To be able to develop accurate plant-based irrigation scheduling tools, automatic and early detection of plant drought stress is of great importance. In this context, measurements of stem diameter variations are very promising as a source of information. These measurements are sensitive for drought stress, but also depend on changing microclimatic conditions. Specific data mining techniques, such as Unfold Principal Component Analysis (UPCA), have been developed to facilitate monitoring and diagnosing of such large-dimensional data sets. A UPCA model is used in this study to determine whether the measured stem diameter variations deviate from normal conditions due to drought stress. A newer technique, Functional Unfold Principal Component Analysis (FUPCA), combines functional data analysis with UPCA. The function parameters instead of the original data are then analysed by UPCA. The resulting FUPCA model is less complex and more robust compared to the original UPCA model. Moreover, FUPCA can handle days with missing data straightforwardly. The performances of UPCA and FUPCA models for online plant stress detection were investigated and compared to each other. Two pilot-scale setups were conducted: one with an herbaceous and one with a woody species. For both species, UPCA and FUPCA were shown to be applicable for stress detection. Both allowed successful detection days before visible symptoms appeared, while FUPCA exhibited a lesser parametric complexity.