融合Boruta-RFE与注意力机制的玉米地上生物量无人机遥感估算与单产预测方法

Maize Aboveground Biomass Estimation and Yield Prediction Using UAV Remote Sensing Imagery with Boruta-RFE and Lightweight Attention Mechanism

摘要:

为提高我国东北黑土区保护性耕作条件下地上玉米生物量估算与单产预测精度,以吉林省梨树县春玉米试验区为对象,利用无人机获取的可见光影像(Red-Green-Blue,RGB)、多光谱和热红外等多源影像,构建融合Boruta递归特征消除法(Boruta-recursive feature elimination,Boruta-RFE)、极端梯度提升算法(eXtreme gradientboosting,XGBoost)地上生物量估算与轻量注意力机制的单产预测方法。结果表明:不同生育期的地上生物量敏感特征存在明显阶段差异,热红外温度统计特征、多光谱纹理特征和RGB颜色指数均具有较高重要性;基于3类特征融合的XGBoost 模型在3个关键生育期均取得较好地上生物量估算效果,其中灌浆期验证集决定系数最高,成熟期验证集均方根误差(Root mean square error,RMSE)最低;仅采用无人机(Unmanned aerial vehicle,UAV)特征进行单产预测时,XGBoost综合表现最佳,验证集R2为0.700;进一步融合关键生育期地上生物量估算值后,各模型精度均提升,其中轻量注意力神经网络(Lightweight attention neural network,LANN) 最优,验证集R2和RMSE分别达到0.855和0.935t/hm2。研究表明,无人机多源影像与地上生物量信息融合可显著提升玉米单产预测精度。

Abstract:

Aiming to improve the accuracy of maize aboveground biomass estimation and yield prediction under conservation tillage in the black soil region of Northeast China, the research was conducted in a spring maize experimental area in Lishu County, Jilin Province. Based on RGB, multispectral, and thermal infrared images acquired by unmanned aerial vehicles (UAVs), a modeling framework integrating Boruta-RFE feature selection, stage-wise XGBoost aboveground biomass estimation, and a lightweight attention mechanism for yield prediction was developed. The results showed that the aboveground biomass-sensitive features exhibited clear stage-specific differences, and thermal infrared temperature statistics, multispectral texture features, and RGB color indices showed high importance. The XGBoost model based on the fusion of RGB, multispectral, and thermal infrared features achieved good aboveground biomass estimation performance at all three growth stages, with the highest validation R2 at the grain-filling stage and the lowest validation NRMSE at the maturity stage. When only UAV-derived features were used for yield prediction, XGBoost showed the best overall performance, with a validation R2 of 0.700. After further incorporating the predicted aboveground biomass values from the three stages, the prediction accuracy of all models improved, among which the lightweight attention-based neural network performed best, with validation R2 and RMSE reaching 0.855 and 0.935t/hm2, respectively. These results indicated that integrating UAV multi-source imagery with stage-wise aboveground biomass information can significantly improve maize yield prediction accuracy and provide technical support for maize growth monitoring and yield forecasting under conservation tillage conditions.