| Accurate and effective yield prediction is essential for wheat breeding, cultivation and field management. In this study, the multispectral and RGB images of winter wheat during the grain filling stage were collected from UAV, and 14 spectral traits and 28 morphological traits were extracted as feature variables. Ten machine learning methods, including linear regression, random forest and neural network, were used to construct wheat yield prediction models, and the differences between the models were compared to select the best one. Additionally, machine learning interpretability method SHAP was introduced to analyze the importance of the feature variables, in order to improve the prediction performance of the model. The results showed that among the 10 machine learning methods used, the BPNN model had the best prediction performance (r2=0.826, RMSE=0.094 t·hm-2). According to the feature importance ranking determined by SHAP, Anthocyanin Reflectance Index (ARI) and Three-Dimensional Canopy Volume (Volume) had the greatest impact on the prediction results, accounting for 45.48% of the total feature importance. After feature selection using SHAP, the BPNN model with the best performance was determined based on nine feature variables (r2=0.865, RMSE=0.075 t·hm-2). This improved the prediction accuracy compared to the BPNN model using all features and the pre-analysis method Pearson correlation analysis. Therefore, based on the optimal yield prediction model, SHAP mechanism can be used to select and analyze the importance of feature variables, so as to further improve the accuracy of wheat yield prediction. |
基本信息
