Related Experiment Video
Updated: Oct 19, 2025
08:47
Author Spotlight: UAV Remote Sensing for Efficient Invasive Plant Biomass Estimation
Published on: February 9, 2024
1.7K
Estimates of Maize Plant Density from UAV RGB Images Using Faster-RCNN Detection Model: Impact of the Spatial
K Velumani1,2, R Lopez-Lozano2, S Madec3
1Hiphen SAS, 120 Rue Jean Dausset, Agroparc, Bâtiment Technicité, 84140 Avignon, France.
Plant Phenomics (Washington, D.C.)
|September 22, 2021
Summary
High-resolution UAV imagery is crucial for early maize plant detection. The Faster-RCNN algorithm performs best with high-resolution data, but mixed resolutions and GAN-based super-resolution show promise for improving low-resolution image analysis.
Area of Science:
- Agricultural Science
- Remote Sensing
- Computer Vision
Background:
- Early-stage plant density is critical for crop yield prediction and management.
- Traditional methods for plant counting are labor-intensive and less accurate.
- Unmanned Aerial Vehicle (UAV) RGB imagery offers a high-throughput alternative for field monitoring.
Purpose of the Study:
- To investigate the impact of image Ground Sampling Distance (GSD) on maize plant detection accuracy.
- To evaluate the Faster-RCNN object detection algorithm's performance across different image resolutions.
- To assess the effectiveness of resolution-mixing and super-resolution techniques for maize plant counting.
Main Methods:
- Collected high-resolution (≈0.3 cm GSD) UAV RGB imagery over six diverse field sites.
- Trained and validated the Faster-RCNN model using native high-resolution and downsampled low-resolution (≈0.6 cm GSD) images.
- Evaluated model performance using metrics like root-mean-square error (RMSE).
- Applied a Generative Adversarial Network (GAN)-based super-resolution method to enhance native low-resolution images.
Main Results:
- Faster-RCNN achieved excellent detection (rRMSE = 0.08) with native high-resolution images.
- Training on synthetic low-resolution images yielded good results (rRMSE = 0.11) on similar synthetic data.
- Significant performance drops occurred when models trained on one resolution were applied to another.
- Mixed-resolution training improved performance on both high- and synthetic low-resolution images, but native low-resolution images remained problematic (rRMSE = 0.48).
- GAN-based super-resolution improved low-resolution image detection (rRMSE = 0.22) over bicubic upsampling, but still lagged behind high-resolution results.
Conclusions:
- Image resolution significantly impacts UAV-based early-stage maize plant detection accuracy.
- Matching training and validation image resolutions is crucial for optimal Faster-RCNN performance.
- While GAN-based super-resolution offers improvements, high-resolution imagery remains superior for reliable early-stage plant detection.
