Determining the effectiveness of the sample split approach on supervised classification model on maize yield prediction

Abstract

The study was motivated by the challenges of yield volatility in developing countries, which in turn affects people's livelihoods and slows economic development. Maize is a staple food in Tanzania, consumed widely across both rural and urban areas. It is vital for national food security, providing a major share of daily caloric intake. Economically, it supports millions of smallholder farmers through subsistence and income. This study aimed to predict maize yield in Tanzania using discriminant supervised classification model.  Data were collected using a structured questionnaire from 421 smallholder farmers in the Mbozi and Mvomero districts in Tanzania. Data analysis was performed using R programming 4.2.3. The results showed 0.867 classifier accuracy on the training sample, indicating a likelihood of the studied units being classified as low-yielding producers, with 13.3 percent of the expected cost of misclassification. Using the sample spilt approach, the study results on out-of-sample discovered the highest probability of farmers was classified as below average with 0.873 model performance and 12.7 expected costs of misclassification. Out of 100 cases (small farmers), 13 are misclassified, slightly fewer than what has been correctly classified. Applying the sample division approach, out of the 100 cases, 12.7 are misclassified. The classification model results indicated that the out-of-sample improves the model accuracy compared to the training sample, suggesting the intervention in resource allocation in terms of subsidies, training programs, and access to better seeds and fertilizers to the producers below the average.