DeepSeedling: deep convolutional network and Kalman filter for plant seedling detection and counting in the field

• Main Authors: Yu Jiang, Changying Li, Andrew H. Paterson, Jon S. Robertson
• Format: Article
• Language: English
• Published: BMC 2019-11-01
• Series: Plant Methods
• Subjects: Faster RCNN; Object detection; Video tracking; Cotton; Population density
• Online Access: http://link.springer.com/article/10.1186/s13007-019-0528-3

Abstract Background Plant population density is an important factor for agricultural production systems due to its substantial influence on crop yield and quality. Traditionally, plant population density is estimated by using either field assessment or a germination-test-based approach. These approaches can be laborious and inaccurate. Recent advances in deep learning provide new tools to solve challenging computer vision tasks such as object detection, which can be used for detecting and counting plant seedlings in the field. The goal of this study was to develop a deep-learning-based approach to count plant seedlings in the field. Results Overall, the final detection model achieved F1 scores of 0.727 (at $$IOU_{all}$$ IOUall ) and 0.969 (at $$IOU_{0.5}$$ IOU0.5 ) on the $$Seedling_{All}$$ SeedlingAll testing set in which images had large variations, indicating the efficacy of the Faster RCNN model with the Inception ResNet v2 feature extractor for seedling detection. Ablation experiments showed that training data complexity substantially affected model generalizability, transfer learning efficiency, and detection performance improvements due to increased training sample size. Generally, the seedling counts by the developed method were highly correlated ($$R^2$$ R2 = 0.98) with that found through human field assessment for 75 test videos collected in multiple locations during multiple years, indicating the accuracy of the developed approach. Further experiments showed that the counting accuracy was largely affected by the detection accuracy: the developed approach provided good counting performance for unknown datasets as long as detection models were well generalized to those datasets. Conclusion The developed deep-learning-based approach can accurately count plant seedlings in the field. Seedling detection models trained in this study and the annotated images can be used by the research community and the cotton industry to further the development of solutions for seedling detection and counting.