Abstract
This paper introduces a novel approach for detecting shadows and changes in man-made objects using background modeling in high-resolution aerial imagery. Traditional methods often fail in these contexts because they assume natural backgrounds follow a Gaussian distribution, whereas large-scale man-made structures create complex, multi-modal distributions. By redefining both natural terrain and buildings as the "background" and identifying shadows as the statistical anomalies, the proposed method achieves robust shadow detection. This is followed by shadow compensation and radiometric normalization to accurately detect structural changes between multi-temporal images.
The Challenge of High-Resolution Imagery
In low-resolution satellite data, buildings are often small enough to be treated as statistical outliers within a predominantly natural landscape. However, as image resolution increases (e.g., 1:8000 scale), buildings occupy significant portions of the frame. This shift breaks the standard assumption that image pixels follow a simple Gaussian distribution.
As shown in Figure 4, the presence of buildings creates multiple peaks in the image histogram, making it impossible to isolate them using simple thresholding or basic anomaly detection. Furthermore, shadows cast by these buildings change based on the sun's position, often leading to false positives in change detection algorithms that mistake moving shadows for new construction.
Methodology
The proposed pipeline addresses these complexities through a multi-stage process:
- Image Partitioning and Classification: The image is divided into Areas of Interest (AOI). Within these areas, Vector Quantization (VQ) is employed to classify pixels into distinct groups such as vegetation, rooftops, and roads.
- Shadows as Statistical Anomalies: By grouping pixels into these classes, the researchers can once again assume a Gaussian distribution within each specific class. Because shadows are consistently darker and occupy a relatively small area, they become the true statistical anomalies within their respective background classes.
- Thresholding and Compensation: A Chi-square ($\chi^2$) distribution is used to set a rigorous threshold for shadow identification. Once detected, shadows are "compensated" by replacing them with the mean brightness value of their background class, effectively creating a shadow-free image for analysis.
- Radiometric Normalization and Differencing: To compare images across different years and scales, linear regression is used to align brightness levels. Finally, image differencing is performed on the compensated images to isolate actual structural changes.
Experimental Results
The method was tested using aerial photographs of an urban area taken in 2000 (1:8000 scale) and 2002 (1:15000 scale). This dataset provided a significant challenge due to the differences in both time and capture scale.
The results demonstrated that the "shadow-as-proxy" approach successfully identified new building construction while ignoring false changes caused by shifting shadows. As seen in Figure 3, the algorithm effectively isolates relevant structural features from the complex background, even when comparing images of significantly different scales.
Conclusion
By flipping the traditional perspective and treating shadows as the primary detection target, this research provides a robust solution for change detection in high-resolution imagery. The use of shadow compensation and radiometric normalization ensures that the algorithm remains focused on actual physical changes in the urban landscape, overcoming the limitations of traditional Gaussian background models.
Summary updated