Calibration transfer across multiple hyperspectral imaging-based plant phenotyping systems: I – Spectral space adjustment

Abstract Hyperspectral Imaging is one of the most popular technologies in plant phenotyping. Various kinds of hyperspectral imaging systems have been developed in the past years. However, there are always significant differences in sensors and imaging environmental conditions between different facilities, which makes it difficult to share image processing algorithms and plant features prediction models. Calibration transfer between the imaging systems is critically important. The white referencing calibration has been proved effective in removing the major lighting intensity differences, but significant spectral differences between systems still exist. In this study, we considered the calibration transfer as a spectral adjustment process and compared four different methods (DS, PDS, DPDS and SST) for adjusting the spectra. To perform the calibration transfer, maize plants were imaged using push-broom style, top-view VNIR hyperspectral camera in two greenhouse phenotyping facilities (inter-facility) and within the same facility (intra-facility) under different imaging conditions. The suggested spectral adjustment methods were tested using master partial least squares regression (PLSR) based calibration models developed for predicting the relative water content (RWC) and nitrogen content (N) of maize plants. Results showed that spectral space transformation (SST) decreased the RMSEv from 9.450% and 10.636% to 3.280% and 2.424% for the predicted RWC of intra and inter-facility transfer, respectively, while the corresponding measures achieved by direct standardization (DS) were 3.412% and 3.105%, respectively. In case of N predictions, similar results were observed for DS and SST. The other two methods (PDS and DPDS) reduced the prediction error for RWC and N, but their performance was not on par with DS and SST. The results indicated that the proposed methods, especially DS and SST were able to alleviate the perturbations inherited in the spectra and thus can help to avoid the need for time-consuming, labor-intensive and costly full recalibration process that arose due to the intra or inter-facility variations.