Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Benchmarking foundation models for hyperspectral image classification: Application to cereal crop type mapping

Walid Elbarz, Mohamed Bourriz, Hicham Hajji, Hamd Ait Abdelali, François Bourzeix

TL;DR

This study tackles cross-region hyperspectral cereal crop mapping using foundation models. It benchmarks three hyperspectral foundation models—HyperSigma, DOFA, and SpectralEarth-ViT—on pixel-level cereal vs non-cereal classification, with a simple upsampling-based decoder and training on EnMAP data and testing on PRISMA data across two Moroccan regions. SpectralEarth-ViT dominates the results with OA around 93.5% and cereal F1 around 90%, while a compact SpectralEarth variant trained from scratch also shows strong performance, illustrating architecture's role in generalization. The ablation study indicates that moderate patch sizes (e.g., 3×3) balance spatial context and leakage, and larger patches can cause leakage and background noise; the work emphasizes training protocols and cross-sensor evaluation to guide future hyperspectral crop mapping.

Abstract

Foundation models are transforming Earth observation, but their potential for hyperspectral crop mapping remains underexplored. This study benchmarks three foundation models for cereal crop mapping using hyperspectral imagery: HyperSigma, DOFA, and Vision Transformers pre-trained on the SpectralEarth dataset (a large multitemporal hyperspectral archive). Models were fine-tuned on manually labeled data from a training region and evaluated on an independent test region. Performance was measured with overall accuracy (OA), average accuracy (AA), and F1-score. HyperSigma achieved an OA of 34.5% (+/- 1.8%), DOFA reached 62.6% (+/- 3.5%), and the SpectralEarth model achieved an OA of 93.5% (+/- 0.8%). A compact SpectralEarth variant trained from scratch achieved 91%, highlighting the importance of model architecture for strong generalization across geographic regions and sensor platforms. These results provide a systematic evaluation of foundation models for operational hyperspectral crop mapping and outline directions for future model development.

Benchmarking foundation models for hyperspectral image classification: Application to cereal crop type mapping

TL;DR

This study tackles cross-region hyperspectral cereal crop mapping using foundation models. It benchmarks three hyperspectral foundation models—HyperSigma, DOFA, and SpectralEarth-ViT—on pixel-level cereal vs non-cereal classification, with a simple upsampling-based decoder and training on EnMAP data and testing on PRISMA data across two Moroccan regions. SpectralEarth-ViT dominates the results with OA around 93.5% and cereal F1 around 90%, while a compact SpectralEarth variant trained from scratch also shows strong performance, illustrating architecture's role in generalization. The ablation study indicates that moderate patch sizes (e.g., 3×3) balance spatial context and leakage, and larger patches can cause leakage and background noise; the work emphasizes training protocols and cross-sensor evaluation to guide future hyperspectral crop mapping.

Abstract

Foundation models are transforming Earth observation, but their potential for hyperspectral crop mapping remains underexplored. This study benchmarks three foundation models for cereal crop mapping using hyperspectral imagery: HyperSigma, DOFA, and Vision Transformers pre-trained on the SpectralEarth dataset (a large multitemporal hyperspectral archive). Models were fine-tuned on manually labeled data from a training region and evaluated on an independent test region. Performance was measured with overall accuracy (OA), average accuracy (AA), and F1-score. HyperSigma achieved an OA of 34.5% (+/- 1.8%), DOFA reached 62.6% (+/- 3.5%), and the SpectralEarth model achieved an OA of 93.5% (+/- 0.8%). A compact SpectralEarth variant trained from scratch achieved 91%, highlighting the importance of model architecture for strong generalization across geographic regions and sensor platforms. These results provide a systematic evaluation of foundation models for operational hyperspectral crop mapping and outline directions for future model development.

Paper Structure

This paper contains 12 sections, 6 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Study area
  4. Hyperspectral data
  5. Ground truth data
  6. Foundation models
  7. Models selection
  8. Fine-tuning Methodology
  9. Results
  10. Ablation study
  11. Conclusion
  12. Acknowledgments

Figures (6)

  • Figure 1: Study area with the training and testing ROIs.
  • Figure 2: Benchmark’s modeling pipeline.
  • Figure 3: Comparison of SpectralEarth, DOFA, and HyperSigma on the local data using OA, AA, Kappa and F1 score (cereal) metrics.
  • Figure 4: Visual comparison of model predictions on a subset of the test area. The reference image (a) is shown alongside predictions from SpectralEarth (b), DOFA (c), and HyperSigma (d). the cyan color denotes class 1 (cereal) predicted by the model
  • Figure 5: SpectralEarth Nano ViT’s validation accuracy using different patch sizes.
  • ...and 1 more figures