A multi-sensor approach for predicting biomass of extensively managed grassland

Abstract

Leaf area index (LAI), ultrasonic sward height (USH) and common vegetation indices (VI) derived by spectral radiometric reflection data were collected on an experimental field site with three sward types comprising a pure stand of reed canary grass (Phalaris aruninacea), a legume grass mixture and a diversity mixture with thirty-six species in an extensive two cut management system. Sensor measurements and biomass samplings of 0.25m2 subplots were conducted biweekly between May and October in 2009 and 2010. Different combinations of the sensor response values were used in multiple regression analysis to improve biomass (BM) predictions compared to exclusive sensors. Wavelength bands for sensor specific NDVI-type vegetation indices were selected from the hyperspectral data and evaluated for the biomass prediction as exclusive indices or in combination with LAI and USH. In the set of tested parameters, ultrasonic sward height was the best to predict biomass in single sensor approaches (R2 0.73–0.76). Inclusion of LAI improved the model performance and reduced the prediction accuracy by up to 30% for complex swards, while inclusion of vegetation indices resulted only in minor improvements compared to exclusive USH. LAI acted complementary to USH in a combined prediction model, correcting for overestimations of biomass in high swards. Prediction models using exclusive LAI were barely suited to predict biomass accurately (R2 0.36–0.44) but improved significantly when combined with waveband selected VIs (R2<0.8). Combining all three sensors did not significantly improve the model performance.