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

TL;DR

This study probes the practicality of active learning for pixel-wise semantic segmentation in precision agriculture by comparing BALD, PowerBALD, and Random acquisition on two crop–weed datasets. Using MC dropout with the FCHarDNet backbone and agricultural pre-training, PowerBALD demonstrates the strongest tendency to improve annotation efficiency, but gains are not consistently significant due to heavy background dominance and data redundancy. The work highlights critical challenges in applying uncertainty-based active learning to agricultural imagery and suggests avenues such as top-K pixel uncertainty and faster sampling methods to enhance practicality. Overall, the findings underscore both the potential and the limitations of active learning for reducing annotation effort in real-world agricultural segmentation tasks, guiding future methodological refinements.

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.

Figures (6)

• Figure 1: FCHarDNet network architecture chao2019fchardnet with dropout (DO). The numbers in the figure indicate the number of strides for a given block.
• Figure 2: Example image from the Sugarbeet dataset with (a) input image and (b) corresponding annotation, where red=sugarbeet, blue=weed (only a few pixels) and transparent=background. Similar visualization for the PhenoBench dataset (c, d).
• Figure 3: Example image from the Corn-Weed dataset with (a) input image and (b) corresponding annotation, where red=corn, purple=weed (only a few pixels) and transparent=background. Similar visualization for the Corn-Weed pre-trained dataset (c, d).
• Figure 4: Test performances (mIoU) as a function of the number of training images for BALD, PowerBALD, and Random on the Sugarbeet dataset. Subfigure (a) highlights the results without agricultural pre-training, and subfigure (b) with agricultural pre-training. In both subfigures, the solid colored lines are the mean performance values over the three repetitions. The colored areas around the lines represent the 95% confidence intervals around the means. The black dashed lines indicate the performance when FCHarDNet was trained on the entire training pool.
• Figure 5: Histogram of the uncertainty values when deploying BALD on a model trained (a) without agricultural pre-training and (b) with agricultural pre-training. In both (a) and (b), the blue histogram depicts the uncertainty values calculated on the 9287 images of the SugarBeet dataset. The orange histogram depicts the uncertainty values calculated on the 772 images of the Phenobench validation dataset. (c) Uncertainty values visualized on a pixel-level in an example image.