基于无人机影像与机器学习的柑橘产量估测研究

Citrus Yield Estimation by Integrating UAV Imagery and Machine Learning

摘要:

为了准确、快速地预测柑橘产量以准确指导果园生产管理，通过大疆多光谱版无人机获取柑橘果实成熟期的遥感影像数据，并从图像中提取了可见光和多光谱波段指数作为特征变量，采用极端梯度提升（eXtreme gradient boosting，XGB）、随机森林（Random forest，RF）以及支持向量机（Support vector machine，SVM）模型分别构建柑橘果实有无分类模型、果实数量和质量估测模型。结果表明：通过XGB模型对特征变量进行筛选分析，柑橘果实有无的分类中超红指数ExR最重要，而数量和质量的估测中改进超绿指数MExG最重要。组合建模中3个模型均在组合4的情况下精度较好。对于分类模型，最优模型为SVM模型（AUC为0.969，准确率为0.919），而对于数量和质量估测模型，最优模型为XGB模型（数量：R2=0.79，RMSE为466个；质量：R2=0.79，RMSE为19.51kg）。最后利用Shapley additive explanations（SHAP）方法揭示了植被指数特征在产量估测模型构建时的重要性，并阐明了SHAP值排在前四的特征交互影响。本研究结果可为无人机遥感在柑橘产量方面的研究提供应用参考和理论依据。

Abstract:

In order to accurately and rapidly predict citrus yield to precisely guide orchard production management, remote sensing image data of citrus fruit ripening stage was obtained by DJI multispectral version of UAV, and visible and multispectral band indices were extracted as feature variables from the images. The eXtreme gradient boosting (XGB), random forest (RF) and support vector machine (SVM) model were used to construct citrus fruit presence and absence classification model, fruit number and quality estimation model, respectively. The results showed that the excess red index was the most important in the classification of citrus fruit presence and absence while the modified excess green index was the most important in the estimation of number and quality through the screening analysis of feature variables by the XGB model. All three models in combination modeling had better accuracy in combination 4. For the classification model, the optimal model was the SVM model with AUC of 0.969 and accuracy of 0.919.While the XGB model was the best model for estimating both number and quality,with the number estimation model’s R2 value being 0.79 and RMSE being 466, and the quality estimation model’s R2 value being 0.79 and RMSE being 19.51kg. Finally, the Shapley additive explanations (SHAP) method was utilized to reveal the importance of vegetation index features in the construction of the yield estimation model and to elucidate the interaction effects of the features with the top four SHAP values. The research results can provide an application reference and theoretical basis for the research of UAV remote sensing in citrus yield.