Drones and AI team up to spot missing maize seedlings early
image: (a) Data deduplication. The blue waypoints represent the original waypoints, and the yellow waypoints represent the selected waypoints. (b) Image cropping. (c) Data annotation. The red boxes indicate the missing seedlings, and the green boxes indicate the maize seedlings. (d) Sliding window sampling.
Credit: The authors
This method not only boosts detection accuracy but also pinpoints missing seedlings with sub-meter geographic precision, allowing for timely replanting and smarter field management.
Maize is a global staple, and early-stage seedling emergence determines final yield. Gaps caused by non-germination or seedling loss can disrupt uniformity and limit productivity. While UAV (unmanned aerial vehicle) remote sensing has revolutionized non-destructive crop monitoring, most studies focus on later growth stages, missing the critical window for early intervention. Moreover, traditional deep learning requires labor-intensive annotation of large datasets, making timely deployment difficult. Semi-supervised learning, which combines small amounts of labeled data with abundant unlabeled data, offers a cost-effective alternative for improving detection performance in complex agricultural environments.
A study (DOI: 10.1016/j.plaphe.2025.100011) published in Plant Phenomics on 22 February 2025 by Fei Liu’s team, Zhejiang University, developes a UAV-based, semi-supervised detection pipeline that accurately identifies early-stage maize seedlings and missing plants, enabling precise geolocation and timely field management to enhance crop uniformity and yield.
To assess the effectiveness of early-stage maize seedling monitoring, researchers applied a fully supervised object detection model using a labeled UAV image dataset. This model achieved impressive performance, detecting seedlings with 92.08% precision, 90.50% recall, and 92.44% AP50, and detecting missing seedlings with slightly lower metrics—92.04% precision, 84.38% recall, and 90.02% AP50—likely due to class imbalance in the training data. To further improve performance and reduce annotation costs, the team introduced semi-supervised learning using varying amounts of unlabeled data. When 9,000–10,000 unlabeled samples were added to the training set, the model's detection metrics increased by up to 2.69% in precision, 2.65% in recall, and 2.35% in AP50, indicating that unlabeled data significantly enhances learning. The model’s confidence in both seedling and missing seedling predictions also increased with the amount of unlabeled data, peaking and stabilizing around 9K samples. Building on these results, the researchers implemented a post-processing optimization step that slightly adjusted the seedling emergence rate in a sample plot from 66.03% to 65.34%, confirming the model’s accuracy and boundary prediction effectiveness. Finally, they applied a single-image geolocation method to translate detected missing seedling coordinates into geographic positions. This method achieved an average positional deviation of 0.462 meters, well within two times the maize plant spacing of 0.25 meters. Despite slight directional biases caused by field elevation variation, the approach proved feasible for practical field management. Together, these methods form a robust and scalable pipeline for early maize seedling monitoring and geospatial feedback, offering timely and precise insights to guide replanting and improve crop uniformity.
This UAV and AI-powered pipeline provides a timely, scalable, and precise solution for early-stage maize monitoring. Farmers and agronomists can now identify and replant missing seedlings early in the season, ensuring more uniform fields and improved yields. The combination of semi-supervised learning and drone imagery significantly reduces manual annotation burdens while enhancing model performance, making it ideal for real-world agricultural scenarios. Beyond maize, the framework offers a blueprint for early seedling detection in other crops, especially under tight time and resource constraints.
###
References
DOI
Original Source URL
https://doi.org/10.1016/j.plaphe.2025.100011
Funding information
This work was supported by National Key R&D Program of China (2023YFD2000203) and Science and Technology Department of Zhejiang Province (2022C02034).
About Plant Phenomics
Science Partner Journal Plant Phenomics is an online-only Open Access journal published in affiliation with the State Key Laboratory of Crop Genetics & Germplasm Enhancement, Nanjing Agricultural University (NAU) and distributed by the American Association for the Advancement of Science (AAAS). Like all partners participating in the Science Partner Journal program, Plant Phenomics is editorially independent from the Science family of journals. Editorial decisions and scientific activities pursued by the journal's Editorial Board are made independently, based on scientific merit and adhering to the highest standards for accurate and ethical promotion of science. These decisions and activities are in no way influenced by the financial support of NAU, NAU administration, or any other institutions and sponsors. The Editorial Board is solely responsible for all content published in the journal. To learn more about the Science Partner Journal program, visit the SPJ program homepage.