Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Hiba Najjar, Marlon Nuske, Andreas Dengel

TL;DR

This paper investigates data-centric machine learning for Earth Observation by using feature attribution and incremental deletion to identify minimal, necessary, and sufficient feature subsets in temporal multimodal geospatial data. Itbenchmarks six attribution estimators and seven model architectures across three datasets (CropHarvest, CropYield, and China PM2.5) to reveal how data can be pruned without sacrificing baseline performance. The findings show that some datasets reach optimal accuracy with fewer than 20% of temporal instances, and in some cases a single spectral band suffices, highlighting substantial data usage efficiency and potential generalization gains. The study also discusses the faithfulness of attribution methods, noting limited but actionable improvements from ensemble variants, and points to directions for broader method comparisons and cross-model validation.

Abstract

The availability of temporal geospatial data in multiple modalities has been extensively leveraged to enhance the performance of machine learning models. While efforts on the design of adequate model architectures are approaching a level of saturation, focusing on a data-centric perspective can complement these efforts to achieve further enhancements in data usage efficiency and model generalization capacities. This work contributes to this direction. We leverage model explanation methods to identify the features crucial for the model to reach optimal performance and the smallest set of features sufficient to achieve this performance. We evaluate our approach on three temporal multimodal geospatial datasets and compare multiple model explanation techniques. Our results reveal that some datasets can reach their optimal accuracy with less than 20% of the temporal instances, while in other datasets, the time series of a single band from a single modality is sufficient.

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

TL;DR

This paper investigates data-centric machine learning for Earth Observation by using feature attribution and incremental deletion to identify minimal, necessary, and sufficient feature subsets in temporal multimodal geospatial data. Itbenchmarks six attribution estimators and seven model architectures across three datasets (CropHarvest, CropYield, and China PM2.5) to reveal how data can be pruned without sacrificing baseline performance. The findings show that some datasets reach optimal accuracy with fewer than 20% of temporal instances, and in some cases a single spectral band suffices, highlighting substantial data usage efficiency and potential generalization gains. The study also discusses the faithfulness of attribution methods, noting limited but actionable improvements from ensemble variants, and points to directions for broader method comparisons and cross-model validation.

Abstract

The availability of temporal geospatial data in multiple modalities has been extensively leveraged to enhance the performance of machine learning models. While efforts on the design of adequate model architectures are approaching a level of saturation, focusing on a data-centric perspective can complement these efforts to achieve further enhancements in data usage efficiency and model generalization capacities. This work contributes to this direction. We leverage model explanation methods to identify the features crucial for the model to reach optimal performance and the smallest set of features sufficient to achieve this performance. We evaluate our approach on three temporal multimodal geospatial datasets and compare multiple model explanation techniques. Our results reveal that some datasets can reach their optimal accuracy with less than 20% of the temporal instances, while in other datasets, the time series of a single band from a single modality is sufficient.
Paper Structure (18 sections, 4 figures, 1 table)

This paper contains 18 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Methodology
  3. Datasets
  4. CropHarvest.
  5. CropYield.
  6. China PM2.5.
  7. Modeling
  8. Attribution estimators
  9. Incremental Deletion
  10. Baseline Model
  11. Deletion Order
  12. Cycles
  13. Results
  14. Model selection
  15. Incremental deletion
  16. ...and 3 more sections

Figures (4)

  • Figure 1: Incremental band deletion results, comparing and attribution estimators. In the top row the most important bands are removed first, and the least important bands in the second row.
  • Figure 2: Incremental time-step deletion results, comparing and attribution estimators. In the top row the most important time-steps are removed first, and the least important time-steps in the second row.
  • Figure 3: Incremental band deletion results, comparing the ensemble-based estimators and , applied to and . In the top row the most important bands are removed first, and the least important bands in the second row.
  • Figure 4: Incremental time-step deletion results, comparing the ensemble-based estimators and , applied to and . In the top row the most important time-steps are removed first, and the least important time-steps in the second row.