Application of feature selection for predicting leaf chlorophyll content in oats (Avena sativa L.) from hyperspectral imagery
¹Crop Research Institute, Division of Crop Management Systems, Drnovská 507/73, CZ161 06 Praha 6 Ruzyně, Czech Republic
²Czech University of Life Sciences Prague, Faculty of Engineering, Department of Agricultural Machines, Kamýcká 129, CZ165 00 Praha 6 Suchdol, Czech Republic
Feature selection can improve predictions generated by partial least squares models. In the context of hyperspectral imaging, it can also enable the development of affordable devices with specialized applications. The feasibility of feature selection for oat leaf chlorophyll estimation from hyperspectral imagery was assessed using a public domain dataset. A wrapper approach resulted in a simplistic model with poor predictive performance. The number of model inputs decreased from 94 to 3 bands when a filter approach based on the minimum redundancy, maximum relevance criterion was attempted. The filtering led to improved prediction quality, with the root mean square error decreasing from 0.17 to 0.16 g m-2 and R2 increasing from 0.57 to 0.62. Accurate predictions were obtained especially for low chlorophyll levels. The obtained model estimated leaf chlorophyll concentration from near infra-red reflectance, canopy darkness, and its blueness. The prediction robustness needs to be investigated, which can be done by employing an ensemble methodology and testing the model on a new dataset with improved ground-truth measurements and additional crop species.