Overcoming Small Minirhizotron Datasets Using Transfer Learning

Weihuang Xu, Guohao Yu, Alina Zare, Brendan Zurweller, Diane Rowland, Joel Reyes-Cabrera, Felix B. Fritschi, Roser Matamala, Thomas E. Juenger

Abstract

Minirhizotron technology is essential for non-invasive, in-situ study of plant root development. However, the automated analysis of these images—specifically pixel-level segmentation—is hindered by the scarcity of annotated training data. This paper explores the use of U-Net architectures and transfer learning to overcome this "small data" challenge. By pre-training a model on a moderate-sized peanut root dataset (17,550 images) and fine-tuning it on a tiny switchgrass dataset (28 images), we achieved over 99% segmentation accuracy. Our results demonstrate that domain-specific transfer learning significantly outperforms general-purpose pre-training on ImageNet for specialized agricultural tasks.

Introduction: The Challenge of Root Phenotyping

Understanding root systems is critical for improving crop yields and addressing climate change. Minirhizotron systems facilitate this by capturing high-resolution color imagery of roots through clear tubes in the soil. Despite the wealth of data these systems generate, analysis remains a bottleneck. Manual tracing of roots is an error-prone and "extremely tedious" process, making it difficult to generate the large-scale labeled datasets typically required for deep learning.

Raw switchgrass root imagery illustrates the difficulty of segmenting thin, low-contrast roots from complex soil backgrounds.

Methodology

The researchers utilized the U-Net architecture, a convolutional neural network designed for fast and precise semantic segmentation. To address the lack of data for specific species like switchgrass, they investigated several transfer learning strategies:

  • Scratch Training: Random initialization.
  • ImageNet Pre-training: Initializing the encoder with weights from the massive, general-purpose ImageNet dataset.
  • Domain-Specific Pre-training: Initializing the network with weights learned from a related peanut root dataset.

A key finding was that pre-training both the encoder and the decoder (the P-EnDe model) on domain-specific data provided the most stable and accurate results.

Datasets and Ground Truth

The study leveraged two distinct datasets to test transfer learning efficacy:

  1. Peanut Root Dataset: 17,550 images (760 x 580 pixels) with manual ground truth masks.
  2. Switchgrass Root Dataset: A much smaller set of 28 high-resolution images (2160 x 1560 pixels).
Examples of raw imagery and manually labeled ground truth (GT) for both the peanut and switchgrass datasets.

Key Results and Insights

The experiments yielded several critical insights for automated plant phenotyping:

  • Domain Relevance Over Scale: Features learned from peanut roots were far more effective for switchgrass segmentation than those from ImageNet, despite ImageNet being much larger. This highlights the value of "discipline-specific" pre-training.
  • High Accuracy with Minimal Data: The peanut-pre-trained model achieved 99% accuracy on the switchgrass dataset using only 21 training images for fine-tuning.
  • Robustness to Labeling Errors: The deep learning models demonstrated an ability to "correct" minor inconsistencies in the manual ground truth, such as small gaps in the labeled root segments.
  • Superior Convergence: Domain-specific models converged faster and with less variance than those trained from scratch or with ImageNet weights.

Conclusion

Transfer learning from related, moderate-sized datasets offers a powerful solution to the data scarcity problem in agricultural imaging. By integrating these high-accuracy models into annotation tools, researchers can significantly reduce the manual labor required for root analysis, paving the way for larger-scale studies and new scientific discoveries in plant development.

Summary updated

Abstract

Minirhizotron technology is widely used for studying the development of roots. Such systems collect visible-wavelength color imagery of plant roots in-situ by scanning an imaging system within a clear tube driven into the soil. Automated analysis of root systems could facilitate new scientific discoveries that would be critical to address the world's pressing food, resource, and climate issues. A key component of automated analysis of plant roots from imagery is the automated pixel-level segmentation of roots from their surrounding soil. Supervised learning techniques appear to be an appropriate tool for the challenge due to varying local soil and root conditions, however, lack of enough annotated training data is a major limitation due to the error-prone and time-consuming manually labeling process. In this paper, we investigate the use of deep neural networks based on the U-net architecture for automated, precise pixel-wise root segmentation in minirhizotron imagery. We compiled two minirhizotron image datasets to accomplish this study: one with 17,550 peanut root images and another with 28 switchgrass root images. Both datasets were paired with manually labeled ground truth masks. We trained three neural networks with different architectures on the larger peanut root dataset to explore the effect of the neural network depth on segmentation performance. To tackle the more limited switchgrass root dataset, we showed that models initialized with features pre-trained on the peanut dataset and then fine-tuned on the switchgrass dataset can improve segmentation performance significantly. We obtained 99\% segmentation accuracy in switchgrass imagery using only 21 training images. We also observed that features pre-trained on a closely related but relatively moderate size dataset like our peanut dataset are more effective than features pre-trained on the large but unrelated ImageNet dataset.
BibTeX
@article{xu2020overcoming,
  author = {Weihuang Xu and Guohao Yu and Alina Zare and Brendan Zurweller and Diane Rowland and Joel Reyes-Cabrera and Felix B. Fritschi and Roser Matamala and Thomas E. Juenger},
  title = {{Overcoming Small Minirhizotron Datasets Using Transfer Learning}},
  journal = {Comput. Electron. Agric.},
  year = {2020},
  doi = {10.1016/j.compag.2020.105466},
}
Figure