Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mitigating representation bias caused by missing pixels in methane plume detection

Julia Wąsala, Joannes D. Maasakkers, Ilse Aben, Rochelle Schneider, Holger Hoos, Mitra Baratchi

TL;DR

This work addresses representation bias in methane plume detection caused by MNAR missing pixels in TROPOMI data. It introduces two data-centric debiasing strategies: multiple imputations to decouple coverage from the plume label and a resampling scheme that enforces class balance within coverage bins during training. The combination of imputation and resampling reduces equalised-odds bias without harming balanced accuracy, precision, or recall, and improves plume detection in low-coverage images in operational scenarios, albeit with increased false positives relative to physics-based baselines. The study highlights the practical value of data-centric debiasing for remote sensing tasks and points to future work on integrating domain-specific features to further reduce false positives while maintaining improved bias metrics.

Abstract

Most satellite images have systematically missing pixels (i.e., missing data not at random (MNAR)) due to factors such as clouds. If not addressed, these missing pixels can lead to representation bias in automated feature extraction models. In this work, we show that spurious association between the label and the number of missing values in methane plume detection can cause the model to associate the coverage (i.e., the percentage of valid pixels in an image) with the label, subsequently under-detecting plumes in low-coverage images. We evaluate multiple imputation approaches to remove the dependence between the coverage and a label. Additionally, we propose a weighted resampling scheme during training that removes the association between the label and the coverage by enforcing class balance in each coverage bin. Our results show that both resampling and imputation can significantly reduce the representation bias without hurting balanced accuracy, precision, or recall. Finally, we evaluate the capability of the debiased models using these techniques in an operational scenario and demonstrate that the debiased models have a higher chance of detecting plumes in low-coverage images.

Mitigating representation bias caused by missing pixels in methane plume detection

TL;DR

This work addresses representation bias in methane plume detection caused by MNAR missing pixels in TROPOMI data. It introduces two data-centric debiasing strategies: multiple imputations to decouple coverage from the plume label and a resampling scheme that enforces class balance within coverage bins during training. The combination of imputation and resampling reduces equalised-odds bias without harming balanced accuracy, precision, or recall, and improves plume detection in low-coverage images in operational scenarios, albeit with increased false positives relative to physics-based baselines. The study highlights the practical value of data-centric debiasing for remote sensing tasks and points to future work on integrating domain-specific features to further reduce false positives while maintaining improved bias metrics.

Abstract

Most satellite images have systematically missing pixels (i.e., missing data not at random (MNAR)) due to factors such as clouds. If not addressed, these missing pixels can lead to representation bias in automated feature extraction models. In this work, we show that spurious association between the label and the number of missing values in methane plume detection can cause the model to associate the coverage (i.e., the percentage of valid pixels in an image) with the label, subsequently under-detecting plumes in low-coverage images. We evaluate multiple imputation approaches to remove the dependence between the coverage and a label. Additionally, we propose a weighted resampling scheme during training that removes the association between the label and the coverage by enforcing class balance in each coverage bin. Our results show that both resampling and imputation can significantly reduce the representation bias without hurting balanced accuracy, precision, or recall. Finally, we evaluate the capability of the debiased models using these techniques in an operational scenario and demonstrate that the debiased models have a higher chance of detecting plumes in low-coverage images.
Paper Structure (10 sections, 2 figures, 1 table)

This paper contains 10 sections, 2 figures, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data
  4. Methods
  5. Empirical Evaluation Setup
  6. Results
  7. Resampling and imputation reduce bias without hurting accuracy
  8. Less biased models flag more low-coverage images as plume
  9. Conclusion
  10. Acknowledgments.

Figures (2)

  • Figure 1: The distribution of images in the training, validation and testing partitions as a function of coverage (percentage of valid pixels per image). The partitions contain significantly more images of plumes with high coverage than low coverage. The y-axes are scaled independently across subplots to enable visual comparison.
  • Figure 2: (Left): Absolute percentage difference between average number of flags per grid cell (aggregated over architecture and imputation strategies) for resampled vs. standard (no resampling) networks. Dark purple cells indicate higher disagreement, though difference are small due to few average detections in those cells (Right, averaged over all models). Overall, the predictions show little difference between resampled and standard networks. Grey basemap/white background indicate no data or no difference.