Achieving prediction of starch in cassava (Manihot esculenta Crantz) by data fusion of Vis-NIR and Mid-NIR spectroscopy via machine learning

Abstract

Cassava is an important crop for both the food and bio-energy industries, serving as a primary carbohydrate source and a staple food. Knowledge of the starch content (SC) is a key parameter index its quality. Non-destructive measurement of SC is needed to track the same tubers to study behaviours of SC accumulation in cassava tubers, which benefit for breeders in discovering a good variety and farmer to indicating harvesting period. This paper involves the prediction of cassava tuber starch content (SC) using multiple near infrared (NIR) spectrometers, aiming to measure SC in fresh cassava tuber. This study applies both portable NIR spectrometers at 570–1031 nm and 860 – 1760 nm. The best results of the model which developed provided R2p and RPD were 0.69 and 1.80 for VIS-NIR region. Meanwhile, the Mid-NIR region provided poor performance with R2p and RPD of 0.46 and 1.36, respectively. Therefore, this study aims to assess whether it was better for predicting SC if the model was developed using combined selected significant wavelength from both VIS-NIR and Mid-NIR regions using machine learning. Genetics algorithm (GA) and variables important projection (VIP) were used for selecting significant wavelength. Selected wavelengths were combined and then generate calibration models via machine learning (ML). The results of the SC model developed using NIR spectra from selected wavelength with SV regression method had the highest performance, with R2c and RMSEc of 0.88 and 1.28%. Meanwhile, the R2p, RMSEp, and RPD of 0.74, 1.86% and 1.97, respectively. Then, the best calibration model was used to measure the unknown sample which corrected from a different harvest season. The external test set provided R2p, RMSEp were 0.88 and 1.38%. The results indicate that combination of NIR with different region can achieve.