Prediction of yield and quality in medicinal plant Ligusticum chuanxiong Hort. using uncrewed aerial vehicle multispectral measurement

Yunfan Li, Wu Chen, Hongmei Jia, Xi Chen, Jin-Niu Xing, Wei-Ping Gao, Zhuyun YanPeerJ 2025 · DOI

TL;DR

This study leverages Uncrewed Aerial Vehicle (UAV) multispectral imagery to predict the yield and quality of the medicinal plant Ligusticum chuanxiong (chuanxiong rhizoma) before harvest. By developing machine learning models, researchers can now accurately estimate key metrics like rhizome weight and active compound content, offering a more efficient and reliable alternative to traditional methods.

What Problem Does This Paper Address?

The quality and yield of medicinal plants are crucial for their effective use and economic value. However, traditional methods for estimating these factors are often time-consuming, labor-intensive, and lack the precision needed for timely harvest and processing decisions. This can lead to suboptimal resource allocation and potential compromises in the final quality of medicinal materials. The paper addresses the need for a rapid, non-destructive, and accurate method to assess both the quantity (yield) and quality of Ligusticum chuanxiong before it's harvested, enabling better management of the entire production chain.

Methods

The core of this research lies in applying machine learning techniques to multispectral data captured by UAVs. The process involved:

  • UAV Multispectral Imaging: Data was collected using a DJI Phantom 4 Multispectral UAV equipped with five monochromatic sensors. Flights were conducted at two different time points before harvest.
  • Image Processing: Raw multispectral images underwent geometric and radiometric correction, followed by orthographic mosaicking and segmentation to isolate the L. chuanxiong canopy.
  • Vegetation Index (VI) Extraction: A suite of 23 characteristic variables, primarily vegetation indices (VIs) derived from the multispectral bands, were calculated for regions of interest corresponding to field samples.
  • Feature Selection: Pearson correlation analysis was used to identify the most relevant VIs for predicting yield (fresh weight, dry weight) and active component content (ferulic acid, Z-ligustilide, senkyunolide A). Variables with high correlation and potential collinearity were carefully selected.
  • Machine Learning Models: Four ensemble machine learning algorithms were employed: Random Forest (RF), Adaptive Boosting (AdaBoost), Gradient Boosting Decision Tree (GBDT), and Extreme Gradient Boosting (XGBoost). These were used to build regression models for yield and active component prediction, and classification models for quality assessment.
  • Quality Discriminant Models: Models were developed to classify L. chuanxiong based on its ferulic acid content, adhering to the standard set by the Pharmacopoeia of the People's Republic of China (2020). Borderline SMOTE was used to handle class imbalance.
  • Model Evaluation: Model performance was assessed using metrics like R-squared (R²), Normalized Root Mean Squared Error (NRMSE), Mean Absolute Percentage Error (MAPE), Accuracy, Precision, Recall, F1-Score, and Area Under the Curve (AUC). Independent validation sets were used to test the models' generalization capabilities.

Datasets

The study utilized multispectral imagery collected from Ligusticum chuanxiong fields in Shiyang Town and Aoping Town, China.

  • Training and Testing Data: 125 samples from six different planting fields in Shiyang Town were used for training and testing the models.
  • Independent Validation Data: 24 samples from eight different planting fields in Aoping Town were collected and used as an independent validation set to assess the models' real-world applicability and universality.

The paper does not introduce new benchmark datasets but rather uses collected field data for model development and validation.

Key Insights and Ideas

  • UAVs as a Powerful Tool for Medicinal Plants: This research demonstrates that UAV-based multispectral imaging is not just for common crops but is highly effective for specialized medicinal plants like L. chuanxiong.
  • Sensitivity of Bands Varies with Growth Stage: The study highlights that different spectral bands (e.g., red edge, near-infrared, blue) are more sensitive to yield and active component prediction at different growth stages, likely due to physiological changes in the plant.
  • Model Performance Varies by Target Variable: No single machine learning algorithm was universally best for all predictions. XGBoost excelled in yield prediction, while AdaBoost and RF showed promise for active component prediction, and XGBoost/RF were strong for quality discrimination. This suggests a need for tailored model selection.
  • Importance of Independent Validation: The study emphasizes the critical role of independent validation to confirm model universality and identify potential over-fitting, showing that performance can decrease when applied to new spatial data.
  • Feasibility of Quality Discrimination: The successful construction of quality discriminant models based on ferulic acid content validates the use of VIs for non-destructively assessing medicinal plant quality against established standards.

Key Results

  • Yield Prediction: The best prediction models for fresh weight achieved NRMSE of 23.76% and MAPE of 14.75% on the independent validation set (RF model). For dry weight, the best model (AdaBoost) achieved NRMSE of 34.65% and MAPE of 21.73%.
  • Active Component Prediction:
    • Ferulic Acid: NRMSE = 13.35%, MAPE = 10.25% (AdaBoost model).
  • Z-ligustilide: NRMSE = 34.35%, MAPE = 23.40% (RF model).
  • Senkyunolide A: NRMSE = 45.26%, MAPE = 25.48% (XGBoost model).
  • Quality Discrimination: XGBoost and RF models achieved the best performance on the independent validation set for quality discrimination, with Accuracy = 0.7083 and AUC = 0.7214.

Limitations and Future Work

  • Spatial Differences: The study notes that while the validation data was collected around the same time as the training data, differences in soil and ecological factors between the two locations likely contributed to lower prediction accuracy on the independent validation set. The predicted values were also generally lower than the true values, possibly due to these spatial differences.
  • Model Universality: While the models showed a certain degree of universality, performance did decrease on the independent validation set, suggesting that models might need recalibration for different geographical regions or environmental conditions.
  • Future Work: The authors suggest that UAV multispectral technology has significant potential for non-destructive prediction of medicinal plant yield and quality. This can provide a valuable reference for production, processing, transportation, and storage, ultimately ensuring better quality and more efficient resource utilization. Further research could explore incorporating more advanced spectral bands or fusion techniques to improve model accuracy and robustness across diverse conditions.

Summary updated

Abstract

Accurate predicting the yield and quality of medicinal materials before harvest can effectively guide post-harvest process, including processing and storage, thereby ensuring the final quality of medicinal materials. Currently, traditional experimental methods for yield and quality estimation are inadequate to offer reliable guidance for harvesting and processing of medicinal plan. Uncrewed aerial vehicle (UAV) multispectral can quickly and accurately estimate the yield and quality of field crops. Based on the UAV multispectral data of <i>Ligusticum chuanxiong</i> Hort. obtained about half a month before and near harvest, this study predicted the rhizome yield and the content of active components such as ferulic acid, Z-ligustilide and senkyunolide A. Additionally, the quality discriminant models of chuanxiong rhizoma were constructed according to the ferulic acid content index stipulated in Pharmacopoeia of the People's Republic of China (2020). The results performed on the independent validation set show that the best prediction effects of fresh weight and dry weight of rhizome were NRMSE = 23.76%, MAPE = 14.75% and NRMSE = 34.65%, MAPE = 21.73%, respectively. And the best predictive effects of ferulic acid, Z-ligustilide and senkyunolide A were as follows: NRMSE = 13.35%, MAPE = 10.25%; NRMSE = 34.35%, MAPE = 23.40%; and NRMSE = 45.26%, MAPE = 25.48%. Furthermore, the quality discriminant models XGBoost and AdaBoost had effective performances (Accuracy = 0.7083, AUC = 0.7214). These results suggest that UAV multispectral can be effectively employed to predict both yield and quality before harvest, thereby guiding the harvest and processing of <i>L</i>. <i>chuanxiong</i>.
BibTeX
@inproceedings{li2025prediction,
  author = {Yunfan Li and Wu Chen and Hongmei Jia and Xi Chen and Jin-Niu Xing and Wei-Ping Gao and Zhuyun Yan},
  title = {{Prediction of yield and quality in medicinal plant <i>Ligusticum chuanxiong</i> Hort. using uncrewed aerial vehicle multispectral measurement}},
  booktitle = {PeerJ},
  year = {2025},
  doi = {10.7717/peerj.19264},
}
Figure