AI-based UAV pest and disease detection: Time for a reset?

Area of Science:

• Agricultural Science
• Computer Science
• Remote Sensing

Background:

• Uncrewed aerial vehicles (UAVs) combined with artificial intelligence (AI), machine learning (ML), and deep learning (DL) are increasingly used for crop pest and disease detection.
• The practical robustness and generalizability of these AI-driven remote sensing models in real-world agricultural settings are not well-established.

Purpose of the Study:

• To conduct a meta-analysis of UAV-based pest and disease detection studies.
• To evaluate dataset construction and model validation methodologies in existing research.
• To identify limitations and provide recommendations for improving model robustness and applicability.

Main Methods:

• Meta-analysis of 121 UAV-based studies published between 2018 and 2024.
• Examination of dataset construction, including the use of independent test datasets.
• Assessment of model validation practices and field-level transferability.

Main Results:

• 89% of studies lacked truly independent test datasets, leading to inflated performance metrics.
• Only 11% of studies evaluated models on independent fields, with limited successful transferability.
• Significant methodological limitations were identified in current validation practices.

Conclusions:

• Current validation practices for UAV-based AI models in agriculture are insufficient.
• Overestimated performance and poor generalizability hinder real-world adoption.
• Recommendations are provided to enhance robustness, reproducibility, and practical relevance for improved field applicability.