Zhao et al. (2026) Physics-guided deep learning improves UAV multispectral estimation of strawberry canopy LAI and SPAD

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-09-12
• Authors: ShuangWen Zhao, Tintin Ting, Qiao Pan
• DOI: 10.1038/s41598-026-69278-9

Research Groups

• Agricultural Experimental Base of Liaodong University (Dandong, China)
• Department of Computer Science and Engineering (Liaodong University)

Short Summary

This study proposes a physics-guided deep learning framework, DeepRT, to jointly estimate leaf area index (LAI) and canopy chlorophyll status represented by SPAD from UAV multispectral imagery in greenhouse conditions. The framework combines PROSAIL-generated spectral priors with modality-specific learning from calibrated reflectance, vegetation indices, and texture descriptors.

Objective

• Investigate the feasibility of using physics-guided deep learning to improve UAV-based estimation of LAI and SPAD in greenhouses.
• Develop a robust and reliable framework for joint retrieval of LAI and SPAD under varying illumination conditions.

Study Configuration

• Spatial Scale: Greenhouse scale (40 m x 8 m) with 10 fixed sampling units, each covering approximately 1.2 m².
• Temporal Scale: 34 UAV campaigns conducted over a period of 3 months (November 2024 to January 2025).

Methodology and Data

• Models used: DeepRT framework combining PROSAIL-generated spectral priors with modality-specific learning from calibrated reflectance, vegetation indices, and texture descriptors.
• Data sources: UAV multispectral imagery (P4M bands) acquired over the greenhouse area, along with reference LAI and SPAD measurements obtained through destructive sampling.

Main Results

• DeepRT achieved high accuracy in estimating LAI (0.920 ± 0.010) and SPAD (0.840 ± 0.020) on the independent test set.
• The framework showed improved performance compared to other baseline methods, including PROSAIL+Transformer and XGBoost.

Contributions

• This study demonstrates the effectiveness of physics-guided deep learning in improving UAV-based estimation of LAI and SPAD in greenhouses.
• The proposed DeepRT framework provides a robust and reliable approach for joint retrieval of LAI and SPAD under varying illumination conditions.

Funding

• This research was funded by the National Natural Science Foundation of China (Grant No. 52177043) and the Liaoning Provincial Department of Education (Grant No. L2021LHJZT001).

Citation

@article{Zhao2026Physicsguided,
  author = {Zhao, ShuangWen and Ting, Tintin and Liu, Xiaoxuan and Pan, Qiao},
  title = {Physics-guided deep learning improves UAV multispectral estimation of strawberry canopy LAI and SPAD},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-69278-9},
  url = {https://doi.org/10.1038/s41598-026-69278-9}
}