Machine learning approaches for grain seed quality assessment: a comparative study of maize seed samples in Malawi

The study assessed machine and deep learning algorithms’ ability to predict and classify the quality of maize grain seed for increased agricultural output. It relied on a dataset of 2460 maize seed samples examined by a KEPHIS ISTA-accredited seed testing facility. The K-NN and Logistic Regression algorithms performed the best in predicting and classifying seed samples, with 100% accuracy, precision, recall, and fi-score. The algorithms found that 46.2% of the grain maize seed was correctly classified as poor-quality seed due to improper handling, and poses a danger to productivity and food security for smallholder farmers. The Deep Learning Convolutional Neural Network presented a 92% accuracy with slight fluctuations, mainly due to the simple and structured nature of the data, which was not in a grid-like or time series format. The study therefore recommends using K-Nearest Neighbor and/or Logistic Regression for grain seed classification when presented with well-structured agricultural data. Still, it also suggests expanding the methodology to other agricultural commodities and implementing seed management measures to prevent low-quality seed distribution. This includes training traders on how to maintain ISTA-required levels of germination, purity, and moisture content in their stores. The study highlights the significance of high-quality seeds for smallholder farmers to improve production and food security.