[Paper Review] Active learning for efficient annotation in precision agriculture: a use-case on crop-weed semantic segmentation

Bart M. van Marrewijk, Charbel Dandjinou+6 more|arXiv (Cornell University)|Apr 3, 2024

Smart Agriculture and AIAgricultural and Biological Sciences3 citations

TL;DR

This study evaluates active learning with Bayesian Active Learning by Disagreement (BALD) and its stochastic variant PowerBALD for crop-weed semantic segmentation in precision agriculture, using Monte-Carlo dropout to estimate model uncertainty. PowerBALD outperformed random sampling in model accuracy with lower variance, though differences were not statistically significant due to high class imbalance and image redundancy in agricultural datasets.

ABSTRACT

Optimizing deep learning models requires large amounts of annotated images, a process that is both time-intensive and costly. Especially for semantic segmentation models in which every pixel must be annotated. A potential strategy to mitigate annotation effort is active learning. Active learning facilitates the identification and selection of the most informative images from a large unlabelled pool. The underlying premise is that these selected images can improve the model's performance faster than random selection to reduce annotation effort. While active learning has demonstrated promising results on benchmark datasets like Cityscapes, its performance in the agricultural domain remains largely unexplored. This study addresses this research gap by conducting a comparative study of three active learning-based acquisition functions: Bayesian Active Learning by Disagreement (BALD), stochastic-based BALD (PowerBALD), and Random. The acquisition functions were tested on two agricultural datasets: Sugarbeet and Corn-Weed, both containing three semantic classes: background, crop and weed. Our results indicated that active learning, especially PowerBALD, yields a higher performance than Random sampling on both datasets. But due to the relatively large standard deviations, the differences observed were minimal; this was partly caused by high image redundancy and imbalanced classes. Specifically, more than 89\% of the pixels belonged to the background class on both datasets. The absence of significant results on both datasets indicates that further research is required for applying active learning on agricultural datasets, especially if they contain a high-class imbalance and redundant images. Recommendations and insights are provided in this paper to potentially resolve such issues.

Motivation & Objective

• To assess the effectiveness of uncertainty-based active learning in reducing annotation effort for semantic segmentation in precision agriculture.
• To investigate whether established acquisition functions like BALD and PowerBALD generalize well to agricultural datasets with high class imbalance and image redundancy.
• To compare active learning strategies (BALD, PowerBALD, Random) on real-world crop-weed datasets to identify performance gains and limitations.
• To provide actionable insights for improving active learning in agricultural computer vision by addressing data imbalance and redundancy issues.

Proposed method

• Employed Monte-Carlo dropout to estimate epistemic uncertainty in semantic segmentation models, enabling uncertainty-based image selection.
• Applied three acquisition functions: Bayesian Active Learning by Disagreement (BALD), stochastic variant PowerBALD, and random sampling for comparison.
• Trained models on pre-labeled data and iteratively selected the most uncertain images using uncertainty scores derived from multiple forward passes.
• Evaluated performance on two agricultural datasets—Sugarbeet and Corn-Weed—each with three classes: background, crop, and weed.
• Used pixel-wise segmentation accuracy as the primary metric to compare model performance across acquisition strategies.
• Conducted multiple repetitions to assess stability and variability of results, especially under high redundancy and class imbalance.

Figure 1 : FCHarDNet network architecture [ 13 ] with dropout (DO). The numbers in the figure indicate the number of strides for a given block.

Experimental results

Research questions

• RQ1 Can uncertainty-based active learning methods like BALD and PowerBALD improve model performance with fewer annotated images in agricultural semantic segmentation?
• RQ2 How do BALD and PowerBALD compare to random sampling in terms of segmentation accuracy and training stability on real-world crop-weed datasets?
• RQ3 To what extent do class imbalance and image redundancy in agricultural datasets limit the effectiveness of active learning?
• RQ4 Does the performance of active learning vary between different agricultural crop-weed scenarios, such as sugarbeet and corn-weed?
• RQ5 What are the key challenges in applying active learning to agricultural datasets, and how can they be mitigated?

Key findings

• PowerBALD achieved higher mean segmentation accuracy than random sampling on both the Sugarbeet and Corn-Weed datasets, indicating improved learning efficiency.
• PowerBALD demonstrated lower variance across repeated experiments compared to random sampling, suggesting greater sampling stability.
• Despite better trends, no statistically significant performance differences were observed (p ≤ 0.05), primarily due to high standard deviations.
• Over 89% of pixels in both datasets belonged to the background class, indicating severe class imbalance that likely masked performance gains.
• Image redundancy across the datasets reduced the informativeness of selected samples, diminishing the advantage of uncertainty-based selection.
• Experiments on Cityscapes showed that both BALD and PowerBALD outperformed random sampling, confirming their potential when applied to less redundant, balanced datasets.

(a)