An efficient model transfer approach to suppress biological variation in elastic modulus and firmness regression models using hyperspectral data

Abstract

The goal of the current work is to reduce biological variation of agricultural produce through an efficient model transfer approach with low labeling cost. The core step of the proposed algorithm is to use a selection strategy for choosing informative samples into the calibration set, and subsequently eliminating uninformative/weakly informative/interference samples out of the calibration set. Results on blueberry hyperspectral data validate that the proposed move-in/move-out model transfer approach can effectively diminish biological variability. The performance of the transferred model for elastic modulus prediction (Bluecrop → M2) tends to be stable with Rp and RMSEp (RPD) of 0.742 and 0.177 g/mm2/% (0.860) after the 64 samples are replaced. The transferred model of firmness (Bluecrop → M2) can predict the independent M2 samples with Rp and RMSEp (RPD) of 0.712 and 72.0 g (1.017) at the 15th cycle. For these two mechanical properties, the model transfer approach can lessen heavy annotation work when transferring the initial Bluecrop model to M2. The second model transfer does not perform well. Similar results are found in other model transfer orders such as M2 → Duke → Bluecrop. Good stability is shown during the first model transfer by results on stability analysis. Moreover, results of the comparison experiment indicate that this approach can give a better transfer result than the move-in approach. This algorithm can be extended to large scale applications that cost less to reduce biological variability. (Source codes are freely available to non-commercial users at: https://figshare.com/articles/Move-in_move-out_model_transfer_algorithm-Matlab_toolbox/6989090).