Abstract
Recognizing rice seedling growth stages to timely do field operations, such as temperature control, fertilizer, irrigation, cultivation, and disease control, is of great significance of crop management, provision of standard and well-nourished seedlings for mechanical transplanting, and increase of yield. Conventionally, rice seedling growth stage is performed manually by means of visual inspection, which is not only labor-intensive and time-consuming, but also subjective and inefficient on a large-scale field. The application of machine learning algorithms on UAV images offers a high-throughput and non-invasive alternative to manual observations. This paper presented automatic approaches to detect rice seedling of three critical stages, BBCH11, BBCH12, and BBCH13. Both traditional machine learning algorithms (HOG-SVM) and deep learning algorithms (CNNs) were investigated. EfficientnetB4 achieved the best performance with 99.47% accuracy.
Introduction
In commercial rice farming, seedlings are raised in nursery trays before being transplanted into paddies. The timing of this transplanting is narrow and critical; doing it too early or too late can drastically reduce grain yield. Farmers use the BBCH scale (a decimal system for plant development) to track growth, specifically focusing on stages BBCH11 through BBCH13, which correspond to the unfolding of the first three leaves.
Traditionally, monitoring these stages relies on manual visual inspection. On a large scale, this process is labor-intensive, slow, and prone to human subjectivity. While remote sensing via Unmanned Aerial Vehicles (UAVs) offers a potential solution, the visual similarity between these early growth stages makes automated classification a significant computer vision challenge. This study aims to develop a high-throughput method using drone imagery and machine learning to automatically detect these critical growth stages.
Methodology
The researchers compared two distinct pipelines for image classification: traditional machine learning using Histograms of Oriented Gradients (HOG) with Support Vector Machines (SVM), and modern deep learning using Convolutional Neural Networks (CNNs).
Data Acquisition and Experimental Design
Images were captured using a DJI Phantom 4 RTK drone flying vertically downward at a height of 3 meters. This provided a high ground sampling distance of 0.08 cm/pixel. The study site was located in Zhaoqing, Guangdong Province, China.
The dataset was designed to be comprehensive, accounting for various field conditions:
- Cultivars: Three different rice varieties (Guang8you2156, Huahang No57, and Huahang No51).
- Densities: Five different nursing seedling densities ranging from 35 to 120 g/tray.
- Temporal Range: Images were captured across different sowing dates in March 2021 to ensure representation of all three target growth stages.
Classification Models
- HOG-SVM: This traditional approach involved extracting texture and shape features using HOG, which counts gradient orientations in localized image portions. These features were then classified using an SVM optimized via grid search for hyperparameters.
- Deep Learning (CNNs): The team utilized transfer learning, fine-tuning models pre-trained on ImageNet. Architectures tested included VGG16, ResNet50, DenseNet121, and the EfficientNet family (B0-B7).
Results and Discussion
The experimental results demonstrated a clear superiority of deep learning approaches over traditional machine learning for this specific task.
Model Performance
- Top Performer: EfficientNetB4 was the most effective model, achieving a near-perfect accuracy of 99.47%, with corresponding high precision (99.53%) and F1 score (99.46%).
- Traditional ML: The best HOG-SVM model reached an accuracy of 84.9%. While useful, it lacks the reliability required for fully autonomous systems.
- Architecture Comparison: Within the CNN category, EfficientNetB4 outperformed DenseNet121 (99.06%), ResNet50 (98.97%), and VGG16 (94.84%).
Key Insights
- The "Adjacent Stage" Difficulty: Distinguishing between BBCH12 and BBCH13 proved to be the most significant challenge. Overlapping leaves and structural similarities made it difficult for HOG features to establish a clear boundary, leading to higher error rates in the SVM model.
- Resolution Sweet Spot: For HOG-SVM, performance peaked at an input size of 400x400 pixels. Smaller images lacked detail, while larger ones introduced noise that hindered the traditional classifier.
- Feature Engineering: Deep learning models succeeded because they automatically learn complex, hierarchical features. In contrast, hand-crafted HOG features may miss subtle biological nuances critical for distinguishing early growth stages.
Conclusion
This study confirms that UAV-based imagery combined with deep learning, specifically the EfficientNetB4 architecture, provides a highly accurate and efficient tool for detecting rice seedling growth stages. This high-throughput method offers a non-invasive alternative to manual observations, potentially revolutionizing large-scale crop management. Future work will focus on improving robustness against environmental factors such as varying light conditions and sensor angles.
Summary updated