IMAGE QUATY ASSESSMENT FOR VHR REMOTE SENSING IMAGE CLASSIFICATION

Zhipeng Li, Li Shen, Linmei Wu2016 · DOI

Abstract

Remote sensing images are essential for characterizing land use and land cover. With the increasing availability of very high resolution (VHR) imagery, automated classification has become a primary tool for information extraction. However, VHR images present challenges: high within-class variance and low between-class variance. This study investigates whether standard image quality metrics can objectively predict the classification performance of VHR data. Our findings suggest that while these metrics are useful for detecting severe degradation, they are poor predictors of classification accuracy in natural, high-quality images.

Introduction

In the domain of Very High Resolution (VHR) remote sensing, the abundance of detail is a "double-edged sword." While 2-meter resolution images are visually rich, they complicate automated classification. High resolution often increases within-class variance (e.g., diverse rooftop materials) and decreases between-class variance (e.g., roads vs. parking lots).

Engineers often assume that superior image quality—defined by sharpness, contrast, or signal-to-noise ratios—automatically yields better classification. However, most quality metrics were designed for human visual aesthetics rather than machine learning tasks. This research explores the relationship between standard image quality indices and "classification ability," measured by the Kappa coefficient.

Methodology

The study utilized Worldview satellite imagery of Zhengzhou, China (2-meter resolution), split into 20 rural and 20 urban scenarios. Additionally, 200 simulated images were created by adding noise and blur to study information loss.

The researchers evaluated quality using two categories of metrics:

  1. Direct Metrics (Image Content): Spectral measures like Standard Deviation (SD) and Average Gradient (AG); and spatial/texture measures like Entropy (ENT), Contrast (CON), and Homogeneity (HOM).
  2. Indirect Metrics (Information Loss): Peak Signal-to-Noise Ratio (PSNR) and Structural Similarity (SSIM).

These were tested across different scene complexities (urban vs. rural), semantic hierarchies (broad vs. specific classes), and algorithms (SVM-based pixel analysis vs. Object-Oriented analysis).

Key Results and Discussion

The experimental results revealed a significant disconnect between standard quality metrics and classification success. Regression analysis showed consistently low $R^2$ values, indicating that global metrics like Standard Deviation or Entropy cannot reliably predict a specific image's Kappa score.

Relationship between Kappa coefficient and Standard Deviation (SD) for urban and rural areas.

As shown in the scatter plots, there is no clear linear trend between spectral quality (SD) or texture metrics (ASM and ENT) and the resulting classification accuracy. This mismatch occurs because standard quality indices are "global" averages, whereas classification is a "local" task dependent on specific feature boundaries.

Kappa coefficient vs. Angular Second Moment (ASM) and Entropy (ENT).

Other critical findings include:

  • Scene Complexity: The inherent difficulty of the landscape (urban vs. rural) impacts accuracy far more than minor fluctuations in pixel-level quality.
  • Algorithm Choice: In complex urban environments, Object-Oriented algorithms significantly outperformed pixel-based methods, suggesting that processing strategy is more vital than raw image metrics.
  • The Degradation Threshold: While metrics fail to predict performance on clear images, they are excellent at identifying the "breaking point" where noise and blur cause a predictable collapse in accuracy.

Conclusion

Current global image quality indices are largely divorced from the semantic requirements of automated classification. To better serve the remote sensing community, future research should focus on developing "information-level" indices. These new metrics should describe local information content and the distinguishability of specific features, moving beyond human-centric aesthetics toward machine-centric utility.

Summary updated

Abstract

The data from remote sensing images are widely used for characterizing land use and land cover at present. With the increasing availability of very high resolution (VHR) remote sensing images, the remote sensing image classification becomes more and more important for information extraction. The VHR remote sensing images are rich in details, but high within-class variance as well as low between-class variance make the classification of ground cover a difficult task. What’s more, some related studies show that the quality of VHR remote sensing images also has a great influence on the ability of the automatic image classification. Therefore, the research that how to select the appropriate VHR remote sensing images to meet the application of classification is of great significance. In this context, the factors of VHR remote sensing image classification ability are discussed and some indices are selected for describing the image quality and the image classification ability objectively. Then, we explore the relationship of the indices of image quality and image classification ability under a specific classification framework. The results of the experiments show that these image quality indices are not effective for indicating the image classification ability directly. However, according to the image quality metrics, we can still propose some suggestion for the application of classification.
BibTeX
@inproceedings{li2016image,
  author = {Zhipeng Li and Li Shen and Linmei Wu},
  title = {{IMAGE QUATY ASSESSMENT FOR VHR REMOTE SENSING IMAGE CLASSIFICATION}},
  booktitle = {˜The},
  year = {2016},
  doi = {10.5194/isprsarchives-xli-b7-11-2016},
}
Figure