Deriving phenology of barley with imaging hyperspectral remote sensing

The aim of this paper was to create a model that predicts the different phenological BBCH macro-stages of barley in laboratory on the plot scale and to transfer the most suitable model to the landscape scale. To characterise the phenology, eight vitality and phenology-related vegetation parameters like leaf area index (LAI), Chl-SPAD content, C-content, N-content, C/N-content, canopy chlorophyll content (CCC), gravimetric water content (GWC) and vegetation height at the same time as all imaging hyperspectral measurements (AISA-EAGLE, 395–973nm). These biochemical–biophysical vegetation parameters were investigated according to the different phenological macro-stages of barley. The predictive models were developed using four different types of vegetation indices (VI): (I) published VI’s, (II) reflectance VI’s as well as (III) VI(xy) formula combinations and (IV) a combination of all VI index types using the Library for Support Vector Machines (LibSVM) and tested with a recursive conditional correlation weighting selection algorithm (RCCW) to reduce the number of variables. To increase the performance of the model a 10-fold cross-validation was carried out for all statistical models. The GWC was found to be the most important variable for differentiating between the phenological macro-stages of barley. The most suitable model for predicting the phenological BBCH macro-stages was achieved by a model that combined all three kinds of VI’s: published VI’s, reflectance VI’s and formula combination VI’s with a classification accuracy of 84.80%. With the classification model for the reflectance VI’s Y=746nm and for the VI formula combinations Y=(527+612)nm and Y=(540+639)nm. The best predictive model was applied to the airborne AISA-EAGLE hyperspectral data to model the phenological macro-stages of barley at the landscape level. The classification error of the best predictive model of 12.80% as well as disturbance factors such as channels and areas with weeds or ruderal vegetation lead to misclassifications of BBCH macro-stages at the landscape level. By using One Sensor At Different Scales-Approach (OSADIS), sensor-specific differences in the model building and model transfer can be eliminated. The approach described in the paper for determining the phenology based on imaging hyperspectral RS data shows that in the process of plant phonological development a number of biochemical–biophysical vegetation traits in vegetation change, which can be thoroughly recorded with hyperspectral remote sensing technology. For this reason, hyperspectral RS constitutes an ideal, cost-effective and comparable approach, with whose help vegetation traits and changes can be quantified, which are key for ecological modelling.