Prediction models of starch content in fresh cassava roots for a tapioca starch manufacturer in Thailand

This paper involves an application of prediction models to study quality of incoming raw materials of a tapioca starch manufacturer in Thailand. The objectives are to estimate starch content of fresh cassava roots and to identify significant factors that affect starch content in cassava roots. Three prediction models, including multiple regression, artificial neural network (ANN), and hybrid deep belief network (HDBN), are implemented. Input data were collected from 242 farmers from 49 different sub-districts in Nakhon Ratchasima province in the Northeast of Thailand, who supply fresh cassava roots to the manufacturing plant. Potential factors are classified into four categories: farmers' demographics, cultivation activities, harvesting activities, and logistics activities, a total of 38 variables. Regression models, ANNs with one hidden layer, and HDBNs were constructed for starch content prediction. Prediction performances were evaluated using the root mean square error (RMSE) and mean absolute percentage errors (MAPE), which were 2.44 percent of starch content and 7.283% for the best regression model; 2.41 and 7.055% for the best ANN, and 2.35 and 6.226% for the best HDBN, respectively. The results indicate that HDBN outperforms the other two models in terms of prediction performance. The final regression model and the best ANN are primarily used to identify seven important factors that can potentially describe starch content. These include harvest age, planting density, growing season, farm location, type of soil, cassava variety, and weed control method.