Data-Centric Machine Learning for Earth Observation: Necessary and Sufficient Features

Hiba Najjar, Marlon Nuske, Andreas DengelarXiv 2024 · DOI · arXiv

Abstract

Earth Observation (EO) research has traditionally focused on model-centric improvements, often at the cost of extreme data redundancy and computational inefficiency. This paper proposes a data-centric shift, utilizing the Remove and Retrain (ROAR) framework to identify necessary and sufficient features across multimodal EO datasets. We demonstrate that many state-of-the-art models maintain baseline performance with as little as 20% of their original input features. Our findings highlight significant redundancies in spectral and temporal data, suggesting a path toward more efficient and deployable EO models.

Introduction

The prevailing paradigm in Earth Observation has long been "more data is better." By integrating high-resolution satellite imagery, meteorological data, and ground sensors over extended time series, researchers have built increasingly complex models. However, this model-centric approach often ignores the inherent redundancies in multimodal data, leading to high computational overhead and environmental impact.

This study addresses the data-centric challenge: determining which features are truly necessary for performance and identifying the minimal sufficient set required to reach baseline accuracy. By pruning redundant information, we can create more efficient, generalizable models that are easier to deploy in resource-constrained environments.

Methodology

The researchers employed a modified version of the RemOve And Retrain (ROAR) framework. Unlike standard feature importance methods that merely mask data with zeros, ROAR involves a rigorous three-step process:

  1. Feature Ranking: Features are ranked using attribution methods, comparing Shapley Value Sampling (SVS)—a perturbation-based method—against Guided Backprop (GB)—a gradient-based method.
  2. Feature Removal: The top (or bottom) $k\%$ of features are physically removed from the dataset.
  3. Model Retraining: The model is retrained from scratch on the reduced dataset.

This ensures the model learns to solve the task using only the available features, rather than simply learning to ignore missing values. The study evaluated several architectures, with the Temporal Convolutional Neural Network (TempCNN) emerging as the top performer across all tasks, achieving an 87.8% accuracy on crop classification and high $R^2$ scores for regression tasks.

Key Results and Discussion

Experiments across three distinct datasets—CropHarvest (classification), CropYield (regression), and China PM2.5 (regression)—reveal massive data redundancy.

Performance curves for CropHarvest, CropYield, and China PM2.5 datasets under the ROAR framework. The top row shows the impact of removing the most important features (demonstrating necessity), while the bottom row shows the impact of removing the least important features (demonstrating sufficiency).

Massive Redundancy and Sufficiency

As shown in the figure above, a significant portion of the input data can be discarded with minimal impact on performance. In the CropYield task, the model recovered its baseline performance even after 80% of the spectral bands were removed. For the China PM2.5 task, the model maintained its baseline $R^2$ score after 65% of features were deleted, relying primarily on key meteorological variables like wind speed, direction, and humidity.

SVS vs. Guided Backprop

A critical finding is the superiority of Shapley Value Sampling (SVS) over gradient-based methods. In the ROAR curves (top row), removing features identified as "important" by SVS led to a significantly sharper decline in performance compared to those identified by GB. This indicates that SVS provides more faithful feature rankings for identifying necessary information, whereas GB often fails to capture the true impact of feature removal.

Temporal and Spectral Insights

The study observed that information in EO tasks is often temporally distributed. Performance curves for temporal pruning showed moderate slopes, suggesting that predictive signals are spread across the time series rather than being concentrated in a single "critical" moment. However, specific tasks showed unique patterns:

  • The "Sufficient" Band: For crop yield, a single band—the first short-wave infrared (SWIR) band from Sentinel-2—was sufficient to achieve baseline performance on its own.
  • Recency Bias in Pollution: In the PM2.5 task, the model relied almost entirely on data from the last two days of the 7-day window, suggesting that older data provided little to no predictive value for current air quality.

Conclusion

By shifting focus from model complexity to data efficiency, this research demonstrates that EO models can be significantly streamlined. Identifying necessary and sufficient features not only reduces computational costs but also provides insights into the physical processes underlying EO tasks. While the approach relies on computationally expensive XAI methods like SVS, the resulting efficiency gains are substantial. Future work will explore the universality of these features across different architectures and the scalability of this approach to global foundation models.

Summary updated

BibTeX
@misc{najjar2024data,
  author = {Hiba Najjar and Marlon Nuske and Andreas Dengel},
  title = {{Data-Centric Machine Learning for Earth Observation: Necessary and Sufficient Features}},
  year = {2024},
  eprint = {2408.11384},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.11384},
}
Figure