Liu et al. (2026) Cross-regional estimation of leaf chlorophyll and soil moisture content in drip-irrigated citrus orchards using UAV data and transfer learning

Identification

• Journal: Agricultural Water Management
• Year: 2026
• Date: 2026-03-31
• Authors: Quanshan Liu, Mingjun Wang, NingBo Cui, Shunsheng Zheng, Zongjun Wu, Shouzheng Jiang, Zhihui Wang, Daozhi Gong, Lu Zhao, Liwen Xing, Guoyu Zhu
• DOI: 10.1016/j.agwat.2026.110317

Research Groups

• State Key Laboratory of Hydraulics and Mountain River Engineering & College of Water Resource and Hydropower, Sichuan University, Chengdu 610065, China
• Institute of Environment and Sustainable Development in Agriculture, Chinese Academy of Agricultural Sciences, Beijing 100081, China

Short Summary

This study developed a transfer learning framework using UAV multispectral data to achieve cross-regional estimation of leaf chlorophyll content (LCC) and soil moisture content (SMC) in drip-irrigated citrus orchards. The Fine-tuning strategy, combined with the CNN-LSTM-Attention-XGBoost (CLA-X) model, significantly enhanced estimation accuracy in a new region, demonstrating a viable framework for precision water and fertilizer management.

Objective

• To develop and compare three attention mechanism-based deep learning algorithms (CNN-LSTM-Attention (CLA), CNN-LSTM-Attention-Adaboost (CLA-A), and CNN-LSTM-Attention-XGBoost (CLA-X)) for citrus LCC and SMC estimation.
• To evaluate the cross-regional generalization performance of these attention mechanism-based deep learning models from a source site to a target site.
• To compare two unsupervised (Domain-Adversarial Neural Network (DANN), Transfer Component Analysis (TCA)) and two semi-supervised (Fine-tuning, Multi-task Learning (MTL)) transfer learning methods to improve cross-site adaptability.
• To interpret the most accurate model using SHapley Additive exPlanations (SHAP) to identify key vegetation indices and explore their contributions to LCC and SMC estimation.

Study Configuration

• Spatial Scale: Two experimental citrus orchards in Sichuan Province, Southwest China.
  • Source domain: Renshou site (Meishan City, 29°55′N, 104°16′E, elevation 437 m). Soil type: purple soil, predominantly silt loam (16.7% clay, 68.5% silt, 14.8% sand), bulk density 1450 kg m⁻³. Tree spacing: 2.5 m × 5.0 m, canopy height: 2.5–3.0 m.
  • Target domain: Xuyong site (Luzhou City, 28°04′N, 105°23′E, elevation 539 m). Soil type: middle yellow earths, bulk density 1260 kg m⁻³. Tree spacing: 2.0 m × 5.0 m, canopy height: 1.5–2.2 m.
  • Sampling points: 62 in Renshou, 66 in Xuyong.
  • Soil moisture content (SMC) measured at 10 cm, 20 cm, and 40 cm depths.
• Temporal Scale:
  • Renshou site: 14 dates across 2022 and 2023 (April 20, May 27, June 27, August 7, October 17, November 15, December 18 in 2022; April 15, May 5, June 24, August 30, October 14, November 10, December 28 in 2023). Total 744 samples.
  • Xuyong site: 3 dates in 2024 (April 15, June 30, November 1). Total 198 samples.

Methodology and Data

• Models used:
  • Attention mechanism-based deep learning models: CNN-LSTM-Attention (CLA), CNN-LSTM-Attention-Adaboost (CLA-A), and CNN-LSTM-Attention-XGBoost (CLA-X).
  • Transfer learning strategies: Fine-tuning (semi-supervised), Multi-task Learning (MTL, semi-supervised), Domain-Adversarial Neural Network (DANN, unsupervised), Transfer Component Analysis (TCA, unsupervised).
  • Interpretability: SHapley Additive exPlanations (SHAP).
• Data sources:
  • UAV Multispectral Data:
    • Renshou: DJI Phantom 4 Multispectral UAV (Blue, Green, Red, Red Edge, Near-Infrared bands). Flight altitude 60 m, ground sampling distance (GSD) approximately 5 cm.
    • Xuyong: DJI Matrice 600 Pro UAV with dual MicaSense RedEdge-MX sensor (10 spectral bands across visible, red-edge, and near-infrared regions). Flight altitude 60 m, GSD approximately 5 cm.
  • Field Measurements:
    • Soil Moisture Content (SMC): Gravimetric oven-dry method at 10 cm, 20 cm, and 40 cm depths.
    • Leaf Chlorophyll Content (LCC): SPAD 502Plus chlorophyll meter readings, converted to LCC (g m⁻²) using a calibration model.
  • Derived Features: Multispectral Indices (MI), Structure Indices (SI) (e.g., Canopy Height, Canopy Elevation Fluctuation Rate (CERR)), Texture Features (TF) (Gray-level co-occurrence matrix (GLCM)), and RGB Indices.

Main Results

• In the source domain (Renshou), the CLA-X model achieved the best performance for LCC estimation (R² ranging from 0.827 to 0.897, RMSE from 0.185 to 0.230 g m⁻²) and SMC estimation across all depths (SMC10: R²=0.830–0.979, RMSE=0.686–1.848 %; SMC20: R²=0.830–0.991, RMSE=0.441–1.848 %; SMC40: R²=0.810–0.988, RMSE=0.523–1.956 %).
• Direct transfer of models from the source to the target domain showed significant performance degradation for both LCC (R² ranging from 0.158 to 0.308) and SMC (R² generally below 0.20 across all depths).
• Semi-supervised transfer learning strategies significantly enhanced estimation accuracy in the target domain. Fine-tuning consistently achieved the highest prediction accuracy, with the Fine-tuned CLA-X model yielding R²=0.751 and RMSE=0.426 g m⁻² for LCC.
• For SMC, Fine-tuning also provided the largest improvements, with ΔR² of 0.130–0.145 for SMC10 and 0.195–0.226 for SMC20. SMC40 showed limited improvement across all transfer strategies.
• Unsupervised transfer learning strategies (DANN, TCA) showed limited improvements (ΔR² < 0.10 for LCC), indicating their reduced effectiveness without labeled target samples.
• SHAP interpretation revealed region-specific changes in dominant predictors for LCC: REVI, GNDVI, and NDYI were dominant in the source domain, while NOG became more influential after transfer, suggesting stronger effects of canopy structure and background reflectance.
• For SMC, CERR, NDVI, and RVI2 consistently contributed most across both domains and depths, highlighting the robust coupling between canopy architecture and soil water status.

Contributions

• Developed an integrated framework combining attention-based deep learning, transfer learning, and SHAP interpretation for cross-regional LCC and SMC estimation in citrus orchards.
• Demonstrated the superior performance of the CNN-LSTM-Attention-XGBoost (CLA-X) model for estimating LCC and SMC in the source domain.
• Systematically evaluated and compared semi-supervised and unsupervised transfer learning strategies, identifying Fine-tuning as the most effective method for improving cross-regional adaptability under data-scarce target domain conditions.
• Provided novel insights into the interpretability of deep learning models through SHAP analysis, revealing how dominant vegetation indices shift or remain stable across different regions and soil depths, linking data-driven predictions to underlying biophysical mechanisms.
• Proposed a viable and efficient framework that provides a crucial decision-making basis for citrus water and fertilizer management, enabling precise irrigation and nitrogen application.

Funding

• National Key Research and Development Program of China (2021YFD1600803)
• National Natural Science Foundation of China (52279041, 52309057)
• Science and Technology Program of Sichuan Province (2023YFN0024, 2024ZHCG0101, 2024YFHZ0200, 2024YFHZ0217)
• Training Program of the Innovation Guidance and Scientific and Technological Enterprise of Yunnan Province (202304BT090019)

Citation

@article{Liu2026Crossregional,
  author = {Liu, Quanshan and Wang, Mingjun and Cui, NingBo and Zheng, Shunsheng and Wu, Zongjun and Jiang, Shouzheng and Wang, Zhihui and Gong, Daozhi and Zhao, Lu and Xing, Liwen and Zhu, Guoyu},
  title = {Cross-regional estimation of leaf chlorophyll and soil moisture content in drip-irrigated citrus orchards using UAV data and transfer learning},
  journal = {Agricultural Water Management},
  year = {2026},
  doi = {10.1016/j.agwat.2026.110317},
  url = {https://doi.org/10.1016/j.agwat.2026.110317}
}