Research map / Label-efficient and robust learning / Self-supervised and contrastive
Accounting for Spatial Variability with the Histogram of Oriented Gradients Based Masking Improves Performance of Masked Autoencoder over Hyperspectral Satellite Imagery (Student Abstract)
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.
Proposes HOGMAE, a masked autoencoder masking strategy that biases mask selection using HOG-based spatial variability of PCA-reduced hyperspectral patches, improving reconstruction over random masking on EnMAP imagery.
Problem
Random masking in masked autoencoders ignores the high-dimensional spatial and spectral variability of hyperspectral satellite imagery, leading to suboptimal reconstruction and representation learning.
Approach
Replace uniform random masking with a HOG-guided, preferential masking scheme: patches are split into high- and low-variability groups by HOG features computed on a PCA-reduced channel, and masking is biased toward low-variability patches while keeping the overall ratio fixed. Requires no separate pre-training to produce attention maps and no prior knowledge of bands.
Key results
- On unseen-region Test dataset 2, HOG masking reduces MAE from 0.0136 to 0.0112 (17.6% reduction) versus random masking.
- HOG masking raises PSNR from 34.22 to 36.08 (5.4% increase) and SSIM from 0.8466 to 0.8872 (4.8% increase) on Test dataset 2.
- On Test dataset 1, HOG masking improves MAE 0.0126 to 0.0106, PSNR 34.39 to 35.94, SSIM 0.8582 to 0.8899.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HOG-based masking instead of random masking | MAE · Test dataset 1 | MAE: 0.0126 -> 0.0106 | Helped |
| Use HOG-based masking instead of random masking | MAE · Test dataset 1 | PSNR: 34.39 -> 35.94 | Helped |
| Use HOG-based masking instead of random masking | MAE · Test dataset 1 | SSIM: 0.8582 -> 0.8899 | Helped |
| Use HOG-based masking instead of random masking | MAE · Test dataset 2 (unseen areas) | MAE: 0.0136 -> 0.0112 | Helped |
| Use HOG-based masking instead of random masking | MAE · Test dataset 2 (unseen areas) | PSNR: 34.22 -> 36.08 | Helped |
| Use HOG-based masking instead of random masking | MAE · Test dataset 2 (unseen areas) | SSIM: 0.8466 -> 0.8872 | Helped |
Limitations
- Evaluated only on a single hyperspectral dataset (EnMAP) with a single masking ratio of 75%.
- Only compared against random masking; no comparison with other MAE masking strategies such as attention-based or SpectralMAE in experiments.
- Short student abstract; details of the split ratio, local masking ratios, and model architecture are not fully specified.
Experimental setup
Datasets: EnMAP hyperspectral satellite images
Benchmarks: EnMAP test dataset 1 EnMAP test dataset 2 (unseen regions)
Metrics: MAE PSNR SSIM
Baselines: Random masking
Builds on: masked autoencoder Histogram of Oriented Gradients SpectralMAE
Closest papers by meaning
- S2MAE: A Spatial-Spectral Pretraining Foundation Model for Spectral Remote Sensing Data (CVPR 2024 · similarity 0.86)
- HyperFree: A Channel-adaptive and Tuning-free Foundation Model for Hyperspectral Remote Sensing Imagery (CVPR 2025 · similarity 0.83)
- Self-Guided Masked Autoencoder (NeurIPS 2024 · similarity 0.83)
- LESSViT: Robust Hyperspectral Representation Learning under Spectral Configuration Shift (NeurIPS 2026 · similarity 0.82)
- Asymmetric Masked Distillation for Pre-Training Small Foundation Models (CVPR 2024 · similarity 0.82)
- Learning Mask Invariant Mutual Information for Masked Image Modeling (ICLR 2025 · similarity 0.82)
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.