Comparing various machine learning algorithms for sugar prediction in chickpea using near-infrared spectroscopy

Chickpea is the third major pulse produced globally, with 11.6 million tonnes produced per annum (Merga and Haji, 2019). Sugar alcohols, inulin, starch are all prebiotic carbohydrates found in chickpeas (Johnson et al., 2020). Near-Infrared (NIR) spectroscopy is a non-destructive, versatile and powerful analytical technique.

Spectral data obtained from NIR spectroscopy requires application of various techniques to extract useful information from spectral data which is further used for building various models for prediction of physical or chemical components presents in agricultural crops. The main aim of this study is to apply various machine learning algorithms especially effective in predicting sugar concentration in chickpea. Sugar prediction models are developed using Linear Regression (LR), Artificial Neural Network (ANN), Random Forest (RF), Support Vector Regression (SVR) and Decision Tree Regression (DTR) algorithms. Performance of the models is evaluated using measures namely, Root Mean Square Error (RMSE), Residual Standard Error (RSE), Coefficient of Determination (R²) and Adjusted Coefficient of Determination (adjusted R²).

It was observed that, RF outperformed all other models in terms of accuracy for predicting sugar component from preprocessed spectra, with RMSE, RSE, R2 and adjusted R2 values of 0.054, 0.062, 0.954 and 0.937, respectively. The accuracy of the ANN model is similar to that of the RF, with minor differences in RMSE, RSE, R2 and adjusted R2, values of 0.057, 0.067, 0.952 and 0.935.