This study assesses the potential of using multispectral images collected by an unmanned aerial system (UAS) on machine learning (ML) frameworks to estimate cereal rye (Secale cereal L.) biomass. Multispectral images and ground-truth cereal rye biomass data were collected from 15 farmers’ fields up to three times between March and May in northwest Ohio. Images were processed to derive 13 vegetation indices (VIs). Out of 13 VIs, six optimal sets of VIs, including excess green (ExG), normalized green red difference index (NGRDI), soil adjusted vegetation index (SAVI), blue green ratio (B_G_ratio), red-edge triangular vegetation index (RTVI), and normalized difference red-edge (NDRE) were selected using the variance inflation factor (VIF) based feature selection approach. Six regression models including a multiple linear regression (MLR), elastic net (ENET), multivariate adaptive regression splines (MARS), support vector machine (SVM), random forest (RF), and extreme gradient boosting (XGB) were investigated for estimation of cereal rye biomass based on the VIs. For most of the models, the six selected VIs performed better than or similar to the full set of 13 VIs with R² ranging from 0.24 to 0.59 and RMSE ranging from 83.13 to 91.89 g/m² during 10-fold cross-validation. During independent accuracy assessment with the selected set of VIs, XGB exhibited the highest R² (0.67) and lowest RMSE (83.13 g/m²) and MAE (48.13 g/m²) followed by RF and ENET. For all the models, the agreement between observed and predicted biomass was high for biomass less than or equal to 200 g/m² but decreased for biomass greater than 200 g/m². When field-collected structural features were integrated with the selected VIs, the models showed improved performance, with R² and RMSE of the models reaching up to 0.82 and 61.67 g/m² respectively. Among the six VIs, SAVI showed the strongest impact on the model prediction for the best-performing RF and XGB regression models. The findings of this study demonstrate the potential of precisely estimating and mapping cereal rye biomass based on UAS-captured multispectral images. Timely information on cover crop growth can facilitate numerous decision-making processes, including planning the planting operations, and management of nutrients, weeds, and soil moisture to improve agronomic and environmental outcomes.

本研究评估了在机器学习 (ML) 框架上使用无人机系统 (UAS) 收集的多光谱图像来估计谷物黑麦 (Secale Creek L.) 生物量的潜力。 3 月至 5 月期间，从俄亥俄州西北部 15 个农民的田地收集了多达 3 次的多光谱图像和地面实况谷物黑麦生物量数据。图像经过处理后得出 13 个植被指数 (VI)。 13个VI中，有6组最优VI，包括过量绿（ExG）、归一化绿红差指数（NGRDI）、土壤调节植被指数（SAVI）、蓝绿比（B_G_ratio）、红边三角植被指数（RTVI） ）和归一化差异红边（NDRE）是使用基于方差膨胀因子（VIF）的特征选择方法来选择的。研究了六种回归模型进行估计，包括多元线性回归（MLR）、弹性网络（ENET）、多元自适应回归样条（MARS）、支持向量机（SVM）、随机森林（RF）和极限梯度提升（XGB）基于 VI 的谷物黑麦生物量。对于大多数模型，选定的 6 个 VI 的性能优于或类似于全套 13 个 VI，其中 R ² 范围为 0.24 至 0.59，RMSE 范围为 83.13 至 91.89 g/m ² (0.67) 和最低的 RMSE (83.13 g/m ² ) 和 MAE (48.13 g/m < b4> ) 其次是 RF 和 ENET。对于所有模型，对于小于或等于 200 g/m ² 的生物量，观测到的生物量与预测的生物量之间的一致性较高，但对于大于 200 g/m ² 的生物量，观测到的生物量与预测的生物量之间的一致性较低。 当现场收集的结构特征与选定的 VI 集成时，模型表现出更好的性能，模型的 R ² 和 RMSE 分别达到 0.82 和 61.67 g/m ² 。在六个 VI 中，SAVI 对性能最佳的 RF 和 XGB 回归模型的模型预测影响最大。这项研究的结果证明了根据无人机捕获的多光谱图像精确估计和绘制谷物黑麦生物量的潜力。有关覆盖作物生长的及时信息可以促进众多决策过程，包括规划种植作业以及养分、杂草和土壤湿度的管理，以改善农艺和环境成果。