Data fusion of spectral, thermal and canopy height parameters for improved yield prediction of drought stressed spring barley

Abstract

Abstract Yield modelling based on visible and near infrared spectral information is extensively used in proximal and remote sensing for yield prediction of crops. Distance and thermal information contain independent information on canopy growth, plant structure and the physiological status. In a four-years′ study hyperspectral, distance and thermal high-throughput measurements were obtained from different sets of drought stressed spring barley cultivars. All possible binary, normalized spectral indices as well as thirteen spectral indices found by others to be related to biomass, tissue chlorophyll content, water status or chlorophyll fluorescence were calculated from hyperspectral data and tested for their correlation with grain yield. Data were analysed by multiple linear regression and partial least square regression models, that were calibrated and cross-validated for yield prediction. Overall partial least square models improved yield prediction (R2 = 0.57; RMSEC = 0.63) compared to multiple linear regression models (R2 = 0.46; RMSEC = 0.74) in the model calibration. In cross-validation, both methods yielded similar results (PLSR: R2 = 0.41, RMSEV = 0.74; MLR: R2 = 0.40, RMSEV = 0.78). The spectral indices R780/R550, R760/R730, R780/R700, the spectral water index R900/R970 and laser and ultrasonic distance parameters contributed favourably to grain yield prediction, whereas the thermal based crop water stress index and the red edge inflection point contributed little to the improvement of yield models. Using only more uniform modern cultivars decreased the model performance compared to calibrations done with a set of more diverse cultivars. The partial least square models based on data fusion improved yield prediction (R2 = 0.62; RMSEC = 0.59) compared to the partial least square models based only on hyperspectral data (R2 = 0.48; RMSEC = 0.69) in the model calibration. This improvement was confirmed by cross-validation (data fusion: R2 = 0.39, RMSEV = 0.76; hyperspectral data only: R2 = 0.32, RMSEV = 0.79). Thus, a combination of spectral multiband and distance sensing improved the performance in yield prediction compared to using only hyperspectral sensing.