Research map / Image generation / Autoregressive and latent generation
Data Augmentation Approaches for Satellite Imagery
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Compares natural-image data augmentation methods on three classification and three regression satellite tasks and proposes three satellite-specific augmentations (Sat-CutMix, Sat-SlideMix, Sat-Trivial) that perform best.
Problem
Data augmentation for satellite imagery is typically limited to flips and rotations, and it is unclear whether natural-image augmentation methods transfer well to satellite data with distinct spectral and topological properties and small labeled datasets.
Approach
Adapts CutMix to regression and multi-sample generation (Sat-CutMix), introduces a label-preserving roll-based translation (Sat-SlideMix), and extends TrivialAugment with satellite-specific transforms like random erasing, saturation, and Gaussian noise (Sat-Trivial).
Key results
- Sat-SlideMix had the best average rank (2.17) across six tasks, outperforming all other methods on three tasks and significantly on elevation.
- Sat-CutMix (average rank 2.83) had the highest performance on forest cover and significantly outperformed CutMix on UC Merced.
- Sat-Trivial (average rank 3.17) significantly outperformed TrivialAugment across all tasks; geometric transforms were beneficial while color augmentations were generally not.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use Sat-SlideMix augmentation | ResNet-18 · Elevation | MAE | Helped |
| Use Sat-CutMix augmentation | ResNet-18 · Forest cover | MAE | Helped |
| Use Sat-Trivial instead of TrivialAugment | ResNet-18 · All six tasks | accuracy or MAE | Helped |
| Use geometric augmentation instead of color augmentation | ResNet-18 · Satellite classification and regression tasks | accuracy or MAE | Helped |
| Apply color augmentations | ResNet-18 · Satellite tasks other than elevation | accuracy or MAE | Hurt |
| Add small translation and shear to flip-and-rotate augmentation (ablation) | Geometric · Satellite tasks except Brazilian Coffee Scenes | accuracy or MAE | Helped |
| Remove shear from the geometric augmentation set (ablation) | Geometric · Forest cover | MAE | Hurt |
| Remove shear from the geometric augmentation set (ablation) | Geometric · Tasks other than forest cover | accuracy or MAE | No clear effect |
| Increase the maximum shear magnitude (ablation) | Geometric · Brazilian Coffee Scenes, EuroSAT, and nighttime lights | accuracy or MAE | Hurt |
| Increase the maximum shear magnitude (ablation) | Geometric · Elevation | MAE | Helped |
| Increase the maximum shear magnitude (ablation) | Geometric · Other two tasks | accuracy or MAE | No clear effect |
| Use Sat-CutMix instead of CutMix | ResNet-18 · UC Merced Land Use | accuracy | Helped |
Surprising findings
- Color augmentation sets were generally harmful for satellite classification tasks, contrary to common natural-image practice.
- Natural-image automated policies (AutoAugment, RandAugment, TrivialAugment) gave mixed or degraded results, showing they do not transfer directly to satellite data.
Limitations
- Evaluated only with ResNet-18 and small downsampled datasets.
- Sat-CutMix and CutMix are not applicable to multispectral images in the same way; many baseline augmentations cannot be applied to EuroSAT's 13 bands.
- Proposed methods not evaluated on segmentation tasks, which the authors note as possible extension.
Experimental setup
Datasets: UC Merced Land Use Brazilian Coffee Scenes EuroSAT Forest cover (Rolf et al. 2021) Nighttime light intensity (Rolf et al. 2021) Elevation (Rolf et al. 2021)
Benchmarks: UC Merced Land Use Brazilian Coffee Scenes EuroSAT Forest cover Nighttime light intensity Elevation
Metrics: accuracy precision recall mean absolute error mean squared error R2 percent improvement over no augmentation
Baselines: AutoAugment (ImageNet) AutoAugment (CIFAR10) AutoAugment (SVHN) RandAugment TrivialAugment MoCo v2 aug-plus CutMix Flip & Rotate Geometric Color
Builds on: CutMix TrivialAugment AutoAugment RandAugment MoCo v2
Closest papers by meaning
- Rethinking Transformers Pre-training for Multi-Spectral Satellite Imagery (CVPR 2024 · similarity 0.85)
- Spectrally Distilled Representations Aligned with Instruction-Augmented LLMs for Satellite Imagery (CVPR 2026 · similarity 0.84)
- SatSynth: Augmenting Image-Mask Pairs through Diffusion Models for Aerial Semantic Segmentation (CVPR 2024 · similarity 0.83)
- Effective Data Augmentation With Diffusion Models (ICLR 2024 · similarity 0.83)
- Dual-Granularity Semantic Guided Sparse Routing Diffusion Model for General Pansharpening (CVPR 2025 · similarity 0.82)
- SatCLIP: Global, General-Purpose Location Embeddings with Satellite Imagery (AAAI 2025 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.