Abstract
Building change detection is of great significance in high-resolution remote sensing applications. Multi-index learning, one of the state-of-the-art building change detection methods, still has drawbacks like the inability to find change types directly and the heavy computational consumption of the Morphological Building Index (MBI). This paper proposes a two-stage building change detection method to address these problems. In the first stage, a multi-scale filtering building index (MFBI) is calculated to detect building areas in each temporal with fast speed and moderate accuracy. In the second stage, images and the corresponding building maps are partitioned into grids. In each grid, the ratio of building areas in time T2 and time T1 is calculated. Each grid is classified into one of three change patterns: significantly increase (SI), significantly decrease (SD), and approximately unchanged (AU). Experiments indicate that the proposed method can detect building change types directly and outperform current multi-index learning methods.
Introduction
Detecting changes in urban environments using high-resolution satellite imagery is critical for urban planning and disaster response. However, existing state-of-the-art methods like the Morphological Building Index (MBI) face significant hurdles. MBI requires complex multi-scale and multi-directional operations that are computationally intensive. Furthermore, most current methods provide only binary outputs (change vs. no change), failing to identify the nature of the change, such as construction or demolition. Finally, high-resolution images are prone to noise from shadows and complex textures, leading to false positives in change detection.
Methodology
The proposed framework addresses these challenges through a two-stage process: building detection followed by grid-based change analysis.
Stage 1: Building Detection via MFBI
Instead of traditional MBI, the authors introduce the Multi-scale Filtering Building Index (MFBI). This index utilizes median filtering across four scales (window sizes of 3, 6, 12, and 24 pixels) to extract spatial information. By averaging differential images from these filters, the method captures building features while bypassing the heavy directional calculations required by MBI. This is supplemented by Otsu thresholding and spectral indices (NDVI/NDWI) to remove vegetation and water interference.
Stage 2: Grid-Based Change Detection
The images are partitioned into a grid (e.g., 14x14 or 20x20 segments). For each cell, the ratio of building area at Time 2 ($A_2$) to Time 1 ($A_1$) is calculated. Using an optimized threshold $T$ (typically 2.5), grids are classified into three semantic patterns:
- Significantly Increase (SI): If $A_2/A_1 > T$
- Significantly Decrease (SD): If $A_2/A_1 < 1/T$
- Approximately Unchanged (AU): If the ratio falls between $1/T$ and $T$.
Results and Discussion
The method was evaluated on QuickBird (QB-Wuhan) and Gaofen-2 (GF2-Ezhou) datasets. One of the primary insights is that using a ratio of building areas is much more fault-tolerant than using simple differences, as small extraction errors at one time-step are less likely to trigger a false change signal.
Efficiency and Accuracy
The MFBI approach proved to be nearly 8 to 10 times faster than MBI. For instance, on the QB-Wuhan dataset, processing time dropped from 69 seconds to just 8 seconds. More importantly, the grid-based ratio analysis significantly improved detection accuracy.
As shown in the confusion matrices, the Overall Accuracy (OA) for the QB-Wuhan1 dataset increased from 76.39% to 91.67%, and for the GF2-Ezhou dataset, it rose from 87.00% to 95.50%. The grid-based approach effectively filtered out noise from roads and other non-building structures that often share similar spectral properties.
Conclusion
By replacing exhaustive morphological operations with multi-scale median filtering and adopting a ratio-based grid partition strategy, this method provides a faster, more accurate, and semantically aware solution for building change detection. While the method is slightly more prone to mistaking wide roads for buildings than MBI, its speed and ability to categorize change types make it a powerful tool for large-scale urban analysis. Future work will focus on refining spatial arrangement analysis within grids to detect building replacements of similar sizes.
Summary updated