MTDNGP: A Crop Phenotype Prediction Method Based on Channel Attention and Multi-Tower Convolution

Crop phenotype prediction is of significant importance in precision breeding, and how to accurately predict phenotypes from genotype information remains a considerable challenge. Traditional statistical methods often fail to adequately capture the nonlinear relationship between genotype and phenotype, and the accuracy of prediction is difficult to meet the needs. Deep learning methods with powerful feature learning capabilities have been gradually applied to crop phenotype prediction. However, existing phenotype prediction methods have certain deficiencies in recognizing the importance of key features. To address this problem, we propose MTDNGP, a deep learning model based on channel attention mechanism and multi-tower convolution, which aims to improve the accuracy and reliability of crop phenotype prediction. MTDNGP uses a parallel multi-tower convolution architecture to extract features from multiple dimensions, while combines the channel attention mechanism with the weighting of the key features to improve the model's ability to learn the complex genetic patterns from genomic data. In addition, the model employs batch normalization and dropout strategies to further improve stability and generalization. Cross-validation experiments on soybean and wheat datasets demonstrate that MTDNGP significantly outperforms other traditional statistical, machine learning, and deep learning methods on multiple traits. This validates the superiority of MTDNGP in genomic selection and provides an effective tool for accurate prediction and improvement of crop phenotypes.