Athar et al. (2025) Phenology-aware in-season crop yield estimation through UAV multispectral imagery and deep neural networks

Identification

• Journal: Computers and Electronics in Agriculture
• Year: 2025
• Date: 2025-11-17
• Authors: Usama Athar, Muhammad Ali, Zuhair Zafar, Zahid Mahmood, Karsten Berns, Farid Bourennani, Muhammad Moazam Fraz
• DOI: 10.1016/j.compag.2025.111210

Research Groups

• National University of Sciences and Technology (NUST), Islamabad, Pakistan
• National Agricultural Research Centre, Islamabad, Pakistan
• Rheinland-Pfälzische Technische Universität, Kaiserslautern, Germany
• University of Jeddah, Jeddah, Saudi Arabia

Short Summary

This study introduces a novel phenology-aware framework for in-season crop yield estimation using high-resolution UAV multispectral imagery and deep neural networks, demonstrating significantly improved accuracy (R² of 0.89) by integrating temporal phenological features with structural head metrics.

Objective

• To develop a scalable, non-destructive, and phenology-aware in-season crop yield estimation pipeline that outperforms traditional spectral index-based methods using UAV multispectral imagery and deep neural networks.

Study Configuration

• Spatial Scale: Plot-level, high-resolution imagery for crop fields.
• Temporal Scale: In-season monitoring across 18 distinct timestamps.

Methodology and Data

• Models used:
  • Spatial Phenology Attention and Feature Cross (SPARC) Network (for automated phenological stage mapping)
  • U-Net model (fine-tuned for wheat head segmentation)
  • SAM 2 (used for generating masks for U-Net fine-tuning)
  • Multiple regression models (for yield estimation)
  • Gradient Boosting Regression (achieved best performance)
• Data sources:
  • UAV-based multispectral imagery
  • Processed reflectance maps
  • Vegetation indices (VIs)
  • Canopy height models (CHMs)
  • Fractional cover maps
  • Plot-level phenological and morphological features

Main Results

• The integration of temporal phenological features with structural head metrics significantly improved in-season yield estimation accuracy.
• Gradient Boosting Regression achieved the highest estimation accuracy with an R² of 0.89.
• The proposed approach enhances the granularity and timeliness of in-season yield estimations.
• The framework enables scalable, non-destructive crop monitoring solutions, providing practical information for farmers and breeders.

Contributions

• Presentation of a novel in-season crop yield estimation framework that is phenology-aware and utilizes high-resolution UAV multispectral imagery and deep neural networks.
• Introduction of an integrated pipeline for automated phenological stage mapping using a custom SPARC Network, canopy structure modeling, and wheat head segmentation via a U-Net model fine-tuned with SAM 2 generated masks.
• Demonstration that combining temporal phenological features with structural head metrics significantly improves yield estimation accuracy compared to spectral index-based methods.
• Development of a scalable and non-destructive solution for crop monitoring, offering improved granularity and timeliness in yield predictions.

Funding

[No funding information was provided in the excerpt.]

Citation

@article{Athar2025Phenologyaware,
  author = {Athar, Usama and Ali, Muhammad and Zafar, Zuhair and Mahmood, Zahid and Berns, Karsten and Bourennani, Farid and Fraz, Muhammad Moazam},
  title = {Phenology-aware in-season crop yield estimation through UAV multispectral imagery and deep neural networks},
  journal = {Computers and Electronics in Agriculture},
  year = {2025},
  doi = {10.1016/j.compag.2025.111210},
  url = {https://doi.org/10.1016/j.compag.2025.111210}
}