Estimation of soybean yield parameters under lodging conditions using RGB information from unmanned aerial vehicles

Dong Bai, Delin Li, Chaosen Zhao, Zixu Wang, Mingchao Shao, Bingfu Guo, Yadong Liu, Qi Wang, Jindong Li, Shiyu Guo, Ruizhen Wang, Yinghui Li, Lijuan Qiu, Xiuliang JinFrontiers in Plant Science 2022 · DOI

TL;DR

This study leverages Unmanned Aerial Vehicle (UAV) RGB imagery and machine learning to estimate soybean yield parameters, even under challenging lodging conditions. By analyzing a large dataset of over 1,600 cultivars and fusing various data types (vegetation indices, texture, canopy cover, crop height), they found that the flowering stage is optimal for prediction, and Deep Neural Networks (DNNs) offer the highest accuracy.

What Problem Does This Paper Address?

Accurate and early estimation of crop yields is crucial for agricultural planning and food security. Traditional methods are often labor-intensive and can be influenced by environmental factors. While Unmanned Aerial Vehicles (UAVs) offer a promising solution for efficient data collection, previous studies have often been limited by small numbers of cultivars and controlled experimental settings, making them less applicable to real-world production scenarios. Furthermore, the impact of lodging (plants falling over) on yield estimation, a common issue in agriculture, has not been thoroughly investigated using these advanced techniques. This paper aims to bridge this gap by developing a robust method for estimating soybean yield parameters using readily available RGB data from UAVs, specifically addressing the complexities introduced by lodging.

Methods

The researchers employed a multi-faceted approach:

  • UAV Data Acquisition: RGB images were collected from a DJI Phantom 4 drone at 17 different time points throughout the soybean growing season.
  • Data Feature Extraction: From these images, they extracted several key features:
  • Vegetation Indices (VIs): A suite of 24 VIs were calculated to capture plant health and vigor.
  • Texture Information: Using Gray-Level Co-occurrence Matrix (GLCM), they extracted nine texture indices across RGB bands, further processed to capture mean, min, max, SD, and CV values, resulting in 135 texture variables.
  • Canopy Cover: Extracted using the excess green index (EGI).
  • Crop Height (CH): Derived from a digital surface model (DSM) and a bare ground digital elevation model (DEM).
  • Manual Data Collection: Crucially, they also collected manual data on yield parameters (grain number of seeds per plant and grain weight per plant) and lodging severity (classified on a 1-5 scale).
  • Feature Selection: The Boruta algorithm was used to identify the most important features for predicting yield, helping to reduce redundancy and improve model performance.
  • Machine Learning Models: Five regression algorithms were applied to estimate yield parameters:
  • Partial Least Squares Regression (PLSR)
    • Random Forest Regression (RFR)
    • Logistic Regression (Logistic)
    • Support Vector Machine Regression (SVM)
    • Deep Neural Network (DNN)
  • Model Evaluation: Models were trained on 70% of the data and validated on the remaining 30%, using metrics like R², Root Mean Square Error (RMSE), and relative RMSE (rRMSE).

Datasets

The study utilized data collected from 1,615 soybean cultivars and 190 control plots in Nanchang, Jiangxi Province, China, during the 2020 growing season. This represents a significantly larger and more diverse dataset than typically used in previous UAV-based yield estimation studies. The data was collected over 17 time points, capturing various growth stages.

Key Insights and Ideas

  • Flowering Stage is Key: The study strongly suggests that the flowering stage (around 48 days after sowing) is the most opportune time for estimating soybean yield parameters. This aligns with the physiological peak of soybean growth and development.
  • Data Fusion is Powerful: Combining multiple data sources (vegetation indices, texture, canopy cover, crop height) significantly improves yield estimation accuracy compared to using single data types.
  • Texture Matters: Texture information, often overlooked, proved to be highly valuable, particularly for estimating grain weight per plant. This suggests that subtle variations in canopy structure and appearance carry significant predictive power.
  • Lodging's Impact: Lodging significantly affects yield parameters, especially the grain number of seeds per plant, highlighting the importance of accounting for this common agricultural challenge.
  • DNNs Shine with More Data: While simpler models performed adequately with fewer features, DNNs demonstrated superior performance as the number of input indices increased, underscoring their capacity to handle complex, high-dimensional data.

Key Results

  • Optimal Estimation Time: The flowering stage (48 days after sowing) yielded the best results for estimating yield parameters.
  • Best Model Performance: The Deep Neural Network (DNN) model achieved the highest accuracy, with an R² of 0.66 and an rRMSE of 32.62% for training and R² of 0.50 and rRMSE of 43.71% for validation when estimating the grain number of seeds per plant using 63 indices. For grain weight per plant, the DNN achieved an R² of 0.64 (training) and 0.49 (validation) with all 63 indices.
  • Feature Importance: For grain weight per plant, texture indices were most important, while for grain number of seeds per plant, vegetation indices and lodging were more influential.
  • Data Fusion Benefits: Combining all screened indices (vegetation indices, texture, canopy cover, crop height, and lodging) led to the most accurate estimations across all models.

Limitations and Future Work

  • RGB Data Limitations: While RGB data is economical and practical, it inherently contains less spectral information than multispectral or hyperspectral data, leading to a "poor accuracy of estimation" in some instances.
  • Single Year Study: The study was conducted over a single growing season. Multi-year, multi-area validation is needed to assess model stability and generalizability.
  • Single Time Point Analysis: While the flowering stage was identified as optimal, analyzing data from multiple time points could further enhance estimation accuracy.
  • Lodging Quantification: The study acknowledges the difficulty in numerically quantifying the precise effect of lodging and suggests future work could explore more advanced methods to better model its impact.
  • Future Model Improvements: The authors suggest exploring long and short-term memory (LSTM) or improved Recurrent Neural Network (RNN) algorithms combined with time-series data for more robust yield prediction. They also highlight the need for more efficient methods to improve the stability of estimated yield models for a wider range of genotypes.

Summary updated

Abstract

The estimation of yield parameters based on early data is helpful for agricultural policymakers and food security. Developments in unmanned aerial vehicle (UAV) platforms and sensor technology help to estimate yields efficiency. Previous studies have been based on less cultivars (<10) and ideal experimental environments, it is not available in practical production. Therefore, the objective of this study was to estimate the yield parameters of soybean (Glycine max (L.) Merr.) under lodging conditions using RGB information. In this study, 17 time point data throughout the soybean growing season in Nanchang, Jiangxi Province, China, were collected, and the vegetation index, texture information, canopy cover, and crop height were obtained by UAV-image processing. After that, partial least squares regression (PLSR), logistic regression (Logistic), random forest regression (RFR), support vector machine regression (SVM), and deep learning neural network (DNN) were used to estimate the yield parameters. The results can be summarized as follows: (1) The most suitable time point to estimate the yield was flowering stage (48 days), which was when most of the soybean cultivars flowered. (2) The multiple data fusion improved the accuracy of estimating the yield parameters, and the texture information has a high potential to contribute to the estimation of yields, and (3) The DNN model showed the best accuracy of training (R<sup>2</sup>=0.66 rRMSE=32.62%) and validation (R<sup>2</sup>=0.50, rRMSE=43.71%) datasets. In conclusion, these results provide insights into both best estimate period selection and early yield estimation under lodging condition when using remote sensing.
BibTeX
@inproceedings{bai2022estimation,
  author = {Dong Bai and Delin Li and Chaosen Zhao and Zixu Wang and Mingchao Shao and Bingfu Guo and Yadong Liu and Qi Wang and Jindong Li and Shiyu Guo and Ruizhen Wang and Yinghui Li and Lijuan Qiu and Xiuliang Jin},
  title = {{Estimation of soybean yield parameters under lodging conditions using RGB information from unmanned aerial vehicles}},
  booktitle = {Frontiers in Plant Science},
  year = {2022},
  doi = {10.3389/fpls.2022.1012293},
}
Figure