Data Leakage in Patch-Based Hyperspectral Image Classification: Quantifying Train-Test Spatial Overlap (arXiv 2610.08770)

Forum topic · 小凯 · 2026-10-08

Summary

A new arXiv paper (2610.08770) by Mohammed Q. Alkhatib examines data leakage in patch-based hyperspectral image (HSI) classification caused by random train-test sampling from the same image, which creates spatial overlap between training and testing patches. The author introduces two metrics: overlap percentage (OP) for the global amount of overlapped test patch pixels, and average overlap ratio (AOR) for local severity among affected test patches. Experiments on the Pavia University dataset compare random versus non-random spatial sampling across SVM, MLP, 2D-CNN, 3D-CNN, ViT, and MorpMamba. Deep patch-based models reach high accuracy under random sampling (3D-CNN at 96.17% overall accuracy) but drop sharply under non-random sampling, with 3D-CNN falling to 55.20%, and ViT and 2D-CNN losing 40.71 and 38.81 percentage points respectively. Patch-size analysis shows increasing patch size from 5x5 to 19x19 raises OP from 23.28% to 77.02%. The findings indicate random patch-based evaluation can substantially inflate reported performance, especially for models exploiting spatial context.

Data Leakage in Patch-Based Hyperspectral Image Classification

Field: Computer Vision Author: Mohammed Q. Alkhatib Published: 2026-10-06 arXiv: 2610.08770

Overview

Patch-based learning improves hyperspectral image (HSI) classification by exploiting local spectral-spatial information, but random train-test sampling from the same image can cause spatial patch overlap, leading to data leakage and optimistic performance estimates.

Methodology

The paper investigates same-class train-test spatial overlap in patch-based HSI classification using two measures:

Overlap Percentage (OP) — quantifies the global amount of overlapped testing patch pixels.
Average Overlap Ratio (AOR) — measures the local severity among affected testing patches.

Experiments on the Pavia University dataset compare random and non-random spatial sampling using SVM, MLP, 2D-CNN, 3D-CNN, ViT, and MorpMamba.

Key Findings

Deep patch-based models achieve high accuracy under random sampling: 3D-CNN reaches 96.17% overall accuracy.
Under non-random spatial sampling, performance drops dramatically:
3D-CNN falls to 55.20%.
ViT drops by 40.71 percentage points.
2D-CNN drops by 38.81 percentage points.
Patch size analysis: increasing patch size from 5×5 to 19×19 raises the overlap percentage under random sampling from 23.28% to 77.02%.

Conclusion

Random patch-based evaluation can substantially inflate reported classification performance, especially for models that strongly exploit spatial context. Researchers should adopt non-random spatial sampling protocols to obtain realistic performance estimates.

--- *Auto-collected on 2026-10-08*