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:
- 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).
- 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.
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.
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