Background-Aware Domain Adaptation for Plant Counting.

Area of Science:

• Computer Vision
• Machine Learning
• Agricultural Technology

Background:

• Deep learning models excel at plant counting but struggle with domain gaps, where training and testing data differ.
• Unsupervised domain adaptation (UDA) uses unlabeled target data to improve model performance, but UDA methods are underexplored in plant counting.
• Existing UDA methods show limitations in addressing plant counting challenges.

Purpose of the Study:

• To evaluate existing unsupervised domain adaptation (UDA) methods for plant counting.
• To propose a novel background-aware domain adaptation (BADA) module to address the limitations of current UDA techniques in plant counting.
• To enhance the robustness and accuracy of deep learning-based plant counting models across diverse environments.

Main Methods:

• Evaluated standard feature-level and image-level UDA methods on plant counting tasks.
• Developed a novel background-aware domain adaptation (BADA) module integrated into object counting models.
• Combined BADA with adversarial training strategies for improved robustness.

Main Results:

• The proposed BADA module significantly improved cross-domain plant counting performance, particularly by reducing background counting errors.
• BADA achieved the lowest Mean Absolute Error (MAE) in 6 out of 7 diverse domain adaptation settings.
• Ablation studies and visualizations confirmed the effectiveness and utility of the BADA module.

Conclusions:

• The background-aware domain adaptation (BADA) module effectively mitigates domain gap issues in deep learning-based plant counting.
• BADA offers a flexible and powerful solution for improving plant counting accuracy in real-world, varied conditions.
• The method demonstrates strong potential for enhancing agricultural monitoring and automation through more reliable plant counting.