Robust classification of the nutrition state in crop plants by hyperspectral imaging and artificial neural networks

Abstract

Hyperspectral imaging of crop plants offers the means for a non-invasive, precise and high-throughput plant-phenotyping in plant research and precision agriculture. We already reported the successful separation of spectral signatures by means of unsupervised learning (e.g. clustering) of tobacco leaves grown from different genetic background and under different nutritional conditions [1,2]. In this contribution we evaluate supervised methods to predict the plant's nutrition state by classification and whether they are robust towards dominant sources of data variation like leaf age or intra-leaf pixel position which are irrelevant for the task at hand. Support Vector Machine (SVM)[3], Supervised Relevance Neural Gas (SRNG) [4], Generalized Relevance Learning Vector Quantization (GRLVQ) [5] and a Radial Basis Function (RBF) Network [6] adopted to perform relevance learning as well were tested. Leaf age snowed the largest impact on classification performance, where SVM and RBF produced robust results while SRNG and GRLVQ methods were reduced to near guessing level. Three cameras covering the VIS/SWIR range were tested and relevance of spectral bands towards nutrition prediction were calculated.