Abstract
High-spatial resolution remotely sensed (HSRRS) imagery provides intricate details of ground objects, yet traditional spectral-based classification often struggles with high intra-class and low inter-class variance. This paper introduces the Pixel Shape Index (PSI), a novel spatial descriptor that captures local geometric structures by analyzing spectral similarity in multiple radial directions. By fusing PSI-derived shape features with traditional spectral data using Support Vector Machines (SVM), the proposed method significantly enhances classification accuracy in complex urban environments.
Introduction
The advent of high-resolution sensors like QuickBird has revolutionized remote sensing, offering sub-meter to few-meter spatial resolutions. However, this increased detail introduces "noise" for traditional pixel-based classifiers. In urban scenes, spectrally identical materials (e.g., concrete) may represent distinct objects like roads or building roofs. Conversely, a single object class may exhibit high spectral diversity due to shadows and varying materials. To overcome these limitations, incorporating spatial context—specifically the shape and structural characteristics of objects—is essential.
Methodology: The Pixel Shape Index (PSI)
The core contribution of this research is the Pixel Shape Index (PSI). Unlike traditional texture measures that rely on fixed-window statistics (such as the Gray-Level Co-occurrence Matrix), PSI is an adaptive, direction-based descriptor.
Directional Probing
For each pixel, the algorithm probes outward in $D$ directions (typically $D=20$). A probe extends along a line as long as the spectral similarity between the center pixel and the neighbor remains within a threshold $T_1$. To prevent infinite extension in homogeneous areas like large water bodies, a maximum length threshold $T_2$ is applied.
Feature Extraction
The resulting lengths of these directional probes form a shape profile for each pixel. From this profile, three primary indices are derived:
- $PSI_1$ (Total Length): The sum of all directional lengths, representing the overall "openness" or size of the local feature.
- $PSI_2$ (Maximum Length): Captures elongated structures, such as roads or narrow corridors.
- $PSI_3$ (Minimum Length): Identifies compact or constrained features, such as small buildings or isolated trees.
SVM Fusion
These three PSI features are concatenated with the original spectral bands (e.g., RGB and Near-Infrared). This high-dimensional feature vector is then processed by a Support Vector Machine (SVM) classifier. SVM was selected for its robustness in handling non-linear decision boundaries and its efficiency with relatively small training datasets compared to other machine learning models.
Experimental Results
The method was tested on QuickBird imagery of Beijing, China (2.44m resolution). The study compared the PSI-SVM approach against spectral-only classification and other spatial feature methods like Wavelet Transforms and GLCM.
Accuracy Improvements
The inclusion of PSI features led to a dramatic increase in classification performance:
- Overall Accuracy (OA) Boost: In a complex urban scene, the OA jumped from 60.5% (spectral-only) to 87.6% (spectral + PSI).
- Class-Specific Gains: The most significant improvement was observed in the "Buildings" class. Spectral-only methods achieved a mere 15.3% accuracy due to confusion with roads and bare land. With PSI, building detection accuracy soared to 88.9%.
Parameter Sensitivity
The research highlighted that the spatial threshold $T_2$ serves as a critical scale parameter. Larger $T_2$ values are effective for identifying large-scale urban blocks, while smaller values are necessary for resolving fine-grained details like narrow alleys.
Discussion and Conclusion
The Pixel Shape Index provides a computationally efficient way to encode the anisotropic nature of urban environments. By capturing the "reach" of a pixel in various directions, it effectively distinguishes between objects that share spectral signatures but differ in geometry, such as elongated roads versus rectangular buildings.
While highly effective, the authors note that future improvements could involve multi-scale PSI implementations to capture objects of vastly different sizes within a single pass. Nevertheless, the fusion of spectral and shape features via SVM represents a significant step forward in the automated interpretation of high-resolution satellite imagery.
Summary updated