Identification of maize kernel varieties based on interpretable ensemble algorithms

春光 毕 ^1,2 Xinhua Bi ² Jinjing Liu ^2* Hao Xie ^2* Shuo Zhang ^3* He Chen ^2* Mohan Wang ^4* Lei Shi ^1,2* Shaozhong Song ^5*

• ¹ Institute for the Smart Agriculture, Jilin Agricultural University, ChangChun, China
• ² College of Information Technology, Jilin Agricultural University, ChangChun, China
• ³ Changchun Humanities and Sciences College, Changchun, Hebei Province, China
• ⁴ Jilin Zhongnong Sunshine Data Company Limited, Changchun, Hebei Province, China
• ⁵ School of Data Science and Artificial Intelligence, Jilin Engineering Normal University, Changchun, China

Maize kernel variety identification is extremely important for reducing maize storage losses and ensuring food security. However, traditional single models show limitations when processing largescale multimodal data. The aim of this study is to construct an interpretable ensemble learning model for maize seed grain variety identification by improving differential evolutionary algorithm through multimodal data fusion. First, morphological and hyperspectral data of maize samples were extracted and preprocessed, and three methods were used to screen features, respectively. Subsequently, morphological data and hyperspectral data are fused, the base learner of the Stacking integration model was selected by diversity index and performance index, and the parameters of the base learner are optimized using a differential evolution algorithm that introduces multiple mutation strategies, dynamically adjusts mutation factors and recombination rates. Finally, Shapley Additive exPlanation is used for interpretable ensemble learning. The experimental results show that the accuracy of HDE-Stacking identification can reach 97.78%, in which the bands of 784 nm, 910 nm, 732 nm, 962 nm, and 666 nm have a positive impact on the identification results. This research provides a scientific basis for the efficient identification of corn kernels of different varieties, which helps to improve the accuracy and traceability of germplasm resource management. It also has important practical value in agricultural production, can effectively improve the efficiency of quality management, and provides technical support to ensure food security.