Abstract
This paper evaluates the efficacy of Active Learning (AL) in reducing the manual labeling burden for image-based plant phenotyping. By comparing various AL strategies—including uncertainty-based and diversity-based methods—across soybean stress and weed classification tasks, the study demonstrates that AL consistently outperforms random sampling. Specifically, uncertainty-based methods like Entropy sampling achieve higher accuracy with significantly smaller labeling budgets by prioritizing "informative" images, effectively addressing the labeling bottleneck in biological data annotation.
Introduction
Deep learning has revolutionized image-based plant phenotyping, enabling automated identification of diseases and weed species. However, these models require vast amounts of expert-annotated data, which is expensive and time-consuming to produce in the biological sciences. This "labeling bottleneck" often hinders the deployment of robust models. This research investigates whether Active Learning (AL)—a framework where the model iteratively selects the most informative samples for annotation—can bypass the need for massive, randomly labeled datasets.
Methodology
The researchers implemented a pool-based active learning cycle. Starting with a small labeled seed set, an acquisition function scans an unlabeled pool to select a batch of 1,000 samples for expert review. The model is then retrained with the expanded dataset, and the cycle repeats. Four AL strategies were compared against a Random Sampling baseline:
- Least Confidence: Targets images where the model’s highest class probability is lowest.
- Entropy: Selects images with the most uniform probability distribution across classes.
- Deep Bayesian Active Learning (DBAL): Uses Monte Carlo dropout to estimate model uncertainty via version disagreement.
- Coreset: A diversity-based approach selecting a representative geometric subset of the feature space.
Experimental Datasets
The study utilized two distinct datasets to test the robustness of AL across different environments:
- Soybean Stress Dataset: 16,573 RGB images of soybean leaves across 9 classes, including 8 types of stress and healthy leaves, collected under controlled indoor conditions.
- Weed Species Dataset: 17,509 RGB images across 9 classes (8 weed species and a "weed-free" class), collected in diverse field conditions with varying backgrounds and lighting.
Key Insights
- "Hard Class" Prioritization: AL methods automatically identify difficult classes. In the soybean dataset, AL prioritized samples from visually similar bacterial diseases (e.g., Bacterial Blight and Bacterial Pustule) over "easier" healthy leaf samples.
- Uncertainty vs. Diversity: In controlled environments (soybean dataset), uncertainty-based methods (Entropy, Least Confidence) outperformed the diversity-based Coreset method. This suggests that for "clean" data, targeting confusion is more effective than ensuring geometric diversity.
- Field Data Resilience: AL showed even greater gains on the noisy weed dataset. While random sampling wastes budget on redundant or simple field images, AL focuses on informative edge cases amidst the clutter.
- The "Dip" Phenomenon: A temporary accuracy dip was observed in the soybean dataset around the 2,000-sample mark, likely due to class imbalance introduced by AL before the model reached stability.
Results
AL-based acquisition strategies consistently achieved higher mean accuracy than random sampling for any fixed labeling budget across both datasets. Entropy sampling emerged as the top performer, though Least Confidence and DBAL were highly competitive. Crucially, AL methods reached high performance plateaus much faster than random sampling, demonstrating significant efficiency gains in expert time.
Limitations and Future Work
While effective, AL can introduce class imbalance, potentially requiring adjusted loss functions. Future research could explore combining AL with unsupervised representation learning to improve initial sample identification. Additionally, developing incremental training methods would reduce the computational cost of retraining models from scratch in each AL cycle.
Summary updated