Intelligent approaches to wheat grain classification using neural networks in the agricultural sector

¹ Reshetnev Siberian State University of Science and Technology, 660037 Krasnoyarsk, Russia
² Bauman Moscow State Technical University, 105005 Moscow, Russia
³ Moscow Timiryazev Agricultural Academy, Russian State Agrarian University, 127550 Moscow, Russia

^* Corresponding author: ankoz9@yandex.ru

Abstract

This work is dedicated to conducting a comprehensive analysis of a wheat dataset to identify significant attributes for accurate grain classification. The initial dataset contains various parameters of wheat grains, such as length, perimeter, area, compactness, and asymmetry coefficient. The focus of the study is on analyzing the relationships between these attributes and their impact on the classification target field. Initially, data normalization was performed to eliminate the influence of scale differences between variables. This was followed by a correlation analysis, which revealed several key relationships between attributes. Specifically, it was found that the asymmetry coefficient has a moderate positive correlation with the classification target, while the attributes area, compactness, perimeter, and width exhibit a moderate negative correlation. Length, on the other hand, shows a weak negative correlation with the target attribute. To gain a deeper understanding of the data structure, Kohonen Self-Organizing Maps were used, which helped to identify three clusters. The analysis revealed that the most significant attributes for clustering are compactness, width, perimeter, and area, while the asymmetry coefficient was found to be the least significant. In conclusion, two classification models were built and evaluated. The first model included all attributes from the dataset and demonstrated an accuracy of 0.97. The second model used a subset of attributes excluding the asymmetry coefficient and showed a slightly higher accuracy of 0.98. These results confirm that excluding less significant attributes can lead to a minor but noticeable improvement in model accuracy. Overall, the work highlights the importance of selecting the right attributes to enhance the effectiveness of classification models.