??AI for Maize: UAV-Assisted Detection of Seedling Quality ??

Assessing the quality of maize seedlings, particularly their emergence uniformity, missing seedling rate, and repeated seeding rate, is crucial for optimising field management and evaluating germplasm.

Traditional methods are time-consuming, labour-intensive, and subjective, while satellite imagery lacks the necessary resolution for accurate assessment.

Proposed Solution: The researchers propose a method that leverages the strengths of YOLOv8, a deep learning object detection algorithm, and Voronoi diagrams, a spatial analysis tool, to evaluate seedling quality.        

Methodology

1. Data Acquisition and Preprocessing: High-resolution UAV images of maize fields were acquired at different growth stages (V3 and V5) and under varying weather conditions. The images were preprocessed (stitched, orthorectified, cropped) and annotated to train and test the YOLOv8 model.
2. Maize Seedling Detection: A YOLOv8 model was trained to accurately detect and locate individual maize seedlings in the preprocessed images, even in complex field environments with overlapping plants and shadows.
3. Voronoi Diagram Construction and Optimisation: The detected seedling coordinates were used to construct Voronoi diagrams, which partition the field into polygons representing the growth space of each seedling. An outer envelope interpolation method was introduced to optimise the Voronoi segmentation by mitigating edge effects that could lead to inaccurate estimations.

In the figure above: (A) Simulation of absolute uniform spatial distribution of maize seedlings. (B) Simulation of spatial distribution of maize seedlings by the grid perturbation method, σ = 0.2. (C) Grid missing method to simulate the spatial distribution of maize seedlings, with missing proportion P1 = 0.3. (D) Grid repeating method to simulate the spatial distribution of maize seedlings, with repeating proportion of seedlings P2 = 0.06, and with an offset range of 0.3 times the plant spacing. The blue dashed lines represent a grid with a row-to-column spacing ratio of 2:3, the green dots represent the simulated maize positions, and the red solid lines are the Voronoi diagram with the green dots as control points.

In the figure above: (A) Coordinates of maize seedlings. (B) Convex polygon containing all maize seedlings. (C) Expanded convex polygon with inserted virtual points. (D) Constructed Delaunay triangulation. (E) Circumcenters of each triangle calculated. (F) Voronoi diagram constructed by connecting all circumcenters. (Green dots represent maize seedlings, red dots are inserted virtual points, orange dots are circumcenters of triangles, blue dashed lines represent the Delaunay triangulation, the solid blue line represents the minimum convex polygon that includes all the maize seedlings, and orange solid lines represent the Voronoi diagram).

Seedling Quality Indicator Extraction

Three indicators were extracted from the Voronoi diagrams:

• Voronoi Polygon Uniformity Index (VPUI): This novel index quantifies the uniformity of seedling spatial distribution based on the coefficient of variation of Voronoi polygon areas, indicating the evenness of resource allocation and potential intraspecific competition.
• Missing Seedling Rate: Calculated by identifying abnormally large Voronoi polygons, indicating areas where seedlings failed to emerge.
• Repeated Seeding Rate: Calculated by identifying abnormally small Voronoi polygons, indicating areas with multiple seedlings planted too closely together.

Results

• YOLOv8 Performance: The YOLOv8 model achieved high accuracy (over 95%) in detecting maize seedlings, outperforming YOLOv5 and Faster R-CNN models, particularly in scenarios with overlapping plants and shadows.
• Accuracy of Voronoi-Based Estimation: Both in simulated and real-world scenarios, the Voronoi method demonstrated high accuracy in estimating missing and repeated seedling rates, surpassing the accuracy of traditional plant spacing methods.
• Impact of Edge Expansion: The outer envelope interpolation method effectively reduced the negative impact of edge effects on estimation accuracy, particularly for small and fragmented plots.

Discussion

Advantages of VPUI: The study highlights the superior performance of VPUI as an indicator of seedling uniformity compared to traditional plant spacing metrics. It also suggests its potential for characterising intraspecific competition.

Factors Affecting Estimation Accuracy: The authors acknowledge the influence of factors like YOLOv8 detection accuracy and image acquisition timing on the overall accuracy of the proposed method.

Importance of Multiple Observations: The study emphasises the need for multiple observations throughout the emergence stage to account for variations in seedling emergence time and obtain a comprehensive assessment of seedling quality.

In the figure above: (A) Classification results without virtual points added. (B) Classification results with virtual points added. (Red Voronoi units represent missing seedlings, uncolored Voronoi units represent normal units, and blue Voronoi units represent multiple seedlings).

In the figure above: (A) Scatter plot of the predicted missing rates versus ground truth missing rates. (B) Scatter plot of the predicted multiple seedling rates versus ground truth multiple seedling rates.

Conclusions

This research presents a robust and efficient method for evaluating maize seedling quality by integrating UAV imagery, deep learning, and Voronoi spatial analysis.

• The proposed method outperforms traditional approaches in terms of accuracy and provides valuable insights into seedling uniformity, missing seedling rates, and repeated seeding rates.
• The study acknowledges limitations and suggests future research directions, such as optimising target detection models and conducting multi-temporal analyses for a comprehensive understanding of seedling quality dynamics.

This method holds significant potential for improving maize planting management practices and aiding in germplasm evaluation.

It offers a scalable and objective tool for assessing seedling quality, ultimately contributing to increased agricultural production efficiency and reduced labour costs.

Reference

Ren, L.; Li, C.; Yang, G.; Zhao, D.; Zhang, C.; Xu, B.; Feng, H.; Chen, Z.; Lin, Z.; Yang, H. The Detection of Maize Seedling Quality from UAV Images Based on Deep Learning and Voronoi Diagram Algorithms. Remote Sens. 2024, 16, 3548. https://doi.org/10.3390/rs16193548

AI in Agriculture Podcast: Maize Seedling Quality Detection via UAV and Deep Learning

What's on

If you'd like to receive the regular 'AI in Agriculture' newsletter in your inbox, simply add your email to my mailing list.

Join over 8,742 readers who enjoy weekly updates on AI advancements in agriculture!

To get more information about wheat phenotyping capabilities with mobile - ask Petiole Pro