Liu et al. (2026) Estimation of soybean phenotypic parameters across growth stages using UAV-based multi-source feature fusion and XGBoost

Identification

• Journal: Climate smart agriculture.
• Year: 2026
• Date: 2026-01-18
• Authors: Zhimin Liu, Dawei Ding, Abdoul Kader Mounkaila Hamani, Weiguang Zhai, Jia Tian, Yuhong Wang, Lexuan Zhang, Guangshuai Wang, Yadan Du
• DOI: 10.1016/j.csag.2026.100098

Research Groups

• Northwest Agriculture and Forestry University, Yangling, Shanxi Province, China
• Institute of Farmland Irrigation, Chinese Academy of Agricultural Sciences, National Field Comprehensive Observation Research Station of Agricultural Ecosystems in Shangqiu, Xinxiang, Henan Province, China
• Shandong Agricultural University, Taian, Shandong Province, China

Short Summary

This study developed an integrated framework using UAV-based multi-source feature fusion and the XGBoost algorithm to accurately estimate soybean Leaf Area Index (LAI) and Above-Ground Biomass (AGB) across growth stages. It demonstrated superior performance over single-feature models and other algorithms, identifying optimal observation windows and agronomic practices for enhanced precision agriculture.

Objective

• Systematically analyze and compare the performance and potential of machine learning algorithms (SVR, RF, XGBoost) in inverting soybean LAI and AGB.
• Construct inversion models for soybean growth parameters based on multi-source feature fusion (vegetation indices and texture features) and evaluate their dynamic monitoring performance and stability throughout the entire growth period.
• Clarify the regulatory effects of different irrigation methods and planting densities on key phenotypic parameters of soybeans.

Study Configuration

• Spatial Scale: Field experiments conducted in Datian Town, Dongfang City, Hainan Province, China (108°37′E−108°42′E, 19°05′N–19°10′N). The experimental design included 12 plots, each 19.2 m² (6 m length × 3.2 m width), with a uniform row spacing of 0.4 m.
• Temporal Scale: Data collected throughout 2025 across two soybean growing seasons (Season A: sown February 21, harvested June 10; Season B: sown March 22, harvested July 11). Measurements were taken at five key growth stages: seedling (R1), branching (R2), initial pod (R3), full pod (R4), and seed-filling (R5).

Methodology and Data

• Models used: Support Vector Regression (SVR), Random Forest (RF), and eXtreme Gradient Boosting (XGBoost) machine learning algorithms.
• Data sources:
  • UAV Remote Sensing: Multispectral images acquired using a DJI M300 RTK quadrotor UAV equipped with a Specvision multispectral camera (18 spectral bands, 456–776 nm, 15 nm bandwidth per band). Flight altitude was 30 m, flight speed 5.6 m s⁻¹, with 80% forward and 70% side overlap.
  • Ground Truth Data: Leaf Area Index (LAI) measured using a LAI-2000 plant canopy analyzer. Above-Ground Biomass (AGB) measured by destructive sampling and oven-drying method. Leaf Chlorophyll Content (LCC) measured with an SPAD-502 chlorophyll meter.
  • Feature Extraction: 11 vegetation indices (V) and 8 texture features (T) (Mean, Variance, Homogeneity, Contrast, Dissimilarity, Entropy, Second Moment, Correlation) extracted from the near-infrared band using the Gray Level Co-occurrence Matrix (GLCM) with a 3 × 3 pixel window.
  • Experimental Treatments: Split-plot design with two irrigation methods (drip irrigation, micro-sprinkler irrigation) and two planting densities (210,000 plants ha⁻¹, 270,000 plants ha⁻¹). Compound fertilizer (N:P₂O₅:K₂O = 1:1:1) applied at 600 kg ha⁻¹. Irrigation depth was 0.2 m, with an upper limit of 75% field capacity.

Main Results

• Multi-source feature fusion (vegetation indices + texture features) models consistently outperformed single-feature models, showing a 9% increase in R² and a 6.3% decrease in RMSE for LAI estimation.
• The XGBoost algorithm demonstrated superior performance across all growth stages for both LAI and AGB estimation, achieving average R² values of 0.673 for LAI and 0.671 for AGB, with corresponding average RMSE values of 0.117 and 79.751 kg ha⁻¹.
• The full pod stage (R4) was identified as the optimal remote sensing observation window, where the best models achieved peak R² values of 0.846 for LAI and 0.731 for AGB, with RMSE values of 0.131 and 81.01 kg ha⁻¹, respectively.
• Drip irrigation combined with high planting density (270,000 plants ha⁻¹) significantly (P < 0.05) increased soybean LAI and AGB compared to other irrigation and density treatments.

Contributions

• Introduces and validates an integrated framework for dynamic, full-growth-period monitoring of soybean LAI and AGB by fusing UAV-based multispectral vegetation indices and texture features with the XGBoost machine learning algorithm.
• Provides a systematic comparison of the performance of SVR, RF, and XGBoost algorithms, highlighting XGBoost's superior accuracy and robustness for soybean phenotypic parameter inversion across diverse growth stages.
• Identifies the full pod stage (R4) as the optimal phenological window for high-accuracy remote sensing-based estimation of soybean LAI and AGB.
• Quantifies the positive regulatory effects of drip irrigation combined with high planting density on soybean LAI and AGB, offering insights for optimizing cultivation practices in tropical regions.
• Delivers a robust, high-throughput technical solution for crop phenotyping, contributing to the advancement of data-driven smart agriculture and precision cultivation management.

Funding

• China Postdoctoral Science Foundation (2024M753568)
• Open Research Fund of State Key Laboratory of Efficient Utilization of Agricultural Water Resources (SKLAWR-2024-11)
• Chinese Agrosystem Long-Term Observation Network (CALTON-SQ)
• Collaborative Innovation Center of Soybean Biotechnology and Nutrition Efficiency, Henan

Citation

@article{Liu2026Estimation,
  author = {Liu, Zhimin and Ding, Dawei and Hamani, Abdoul Kader Mounkaila and Zhai, Weiguang and Tian, Jia and Wang, Yuhong and Zhang, Lexuan and Wang, Guangshuai and Du, Yadan},
  title = {Estimation of soybean phenotypic parameters across growth stages using UAV-based multi-source feature fusion and XGBoost},
  journal = {Climate smart agriculture.},
  year = {2026},
  doi = {10.1016/j.csag.2026.100098},
  url = {https://doi.org/10.1016/j.csag.2026.100098}
}