基于Stacking集成学习的夏玉米覆盖度估测模型研究

Estimation of Summer Corn Fractional Vegetation Coverage Based on Stacking Ensemble Learning

以基于无人机多光谱影像提取的夏玉米植被指数作为特征变量,利用皮尔森相关系数结合随机森林反向验证权重的方法进行特征选择,去除冗余特征。以随机森林、梯度提升树、支持向量机和岭回归作为初级学习器,以岭回归作为次级学习器,建立基于Stacking集成学习的夏玉米覆盖度估测模型,并通过5折交叉验证进一步提升模型泛化能力,采用随机搜索和网格搜索结合的方法对模型超参数进行优化,使用4种回归指标进行模型精度评价,并利用次年数据验证其鲁棒性。结果表明,与单一模型以及决策树、Xgboost、Adaboost、Bagging集成框架相比,Stacking集成学习模型具有更高的精度和更强的鲁棒性,R^(2)为0.9509,比单一模型平均提升0.0369,比其他集成模型平均提升0.0417;Stacking集成学习模型RMSE、MAE和MAPE分别为0.0432、0.0330和5.01%,各指标分别比单一模型平均降低0.0138、0.0130和2.14个百分点,分别比其他集成模型平均降低0.0185、0.0126和2.15个百分点。本研究为夏玉米覆盖度估测提供了新的方法。

Based on the UAV multi-spectral image,the summer corn vegetation index was extracted as a feature variable,and the Pearson correlation coefficient combined with the random forest algorithm was used to reverse the verification weight method for feature selection and redundant features were removed.Random forest,gradient boosting tree,support vector machine and ridge regression were used as the primary learner,and ridge regression was used as the secondary learner to establish a summer corn coverage estimation model based on Stacking ensemble learning,and 5-fold cross-validation was used to further improve model generalization ability,a combination of random search and grid search was used to optimize model hyper parameters,four regression indicators were used for model accuracy evaluation,and the following year's data was used to verify its robustness.The experimental results showed that compared with a single model and decision tree,Xgboost,Adaboost,and Bagging integrated framework,the Stacking integrated learning model had higher accuracy and stronger robustness.The R^(2) was 0.9509,which was an average improvement of 0.0369 than that of the single model.Compared with other integrated models,the average increase was 0.0417;RMSE,MAE and MAPE were 0.0432,0.0330 and 5.01%,respectively,which were 0.0138,0.0130 and 2.14 percentage points lower than that of the single model,and 0.0185,0.0126 and 2.15 percentage points lower than that of the other integrated models.The research result provided a method and effective support for the estimation of summer corn coverage.