A Saliency Guided Semi-Supervised Building Change Detection Method for High Resolution Remote Sensing Images

Bin Hou, Yunhong Wang, Qingjie Liu2016 · DOI

TL;DR

This paper introduces a novel semi-supervised approach for detecting building changes in high-resolution remote sensing images. By combining pixel-level feature extraction with object-level analysis and a saliency-guided training strategy, the method achieves improved accuracy and robustness compared to existing techniques, particularly for complex urban environments.

What Problem Does This Paper Address?

Accurate and timely change detection (CD) in remote sensing (RS) images is crucial for urban planning, environmental monitoring, and resource management. While traditional CD methods have focused on lower-resolution imagery, the advent of high-resolution (HR) RS data presents new challenges due to increased intraclass variability and decreased inter-class variability. This makes it difficult to accurately identify changes at the building level. Existing object-based CD methods often struggle with defining object boundaries and extracting discriminative features that capture both spectral and spatial information. Furthermore, supervised CD methods require extensive labeled data, which is costly and time-consuming to obtain. This paper aims to bridge this gap by developing an effective and efficient semi-supervised CD method specifically for HR building change detection.

Methods

The proposed method integrates pixel-based and object-based strategies in a semi-supervised framework.

  • Pixel-based Feature Extraction: Extended Morphological Attribute Profiles (EMAPs) are used to extract multi-resolution spatial and spectral features from the HR images. EMAPs are designed to capture structural information at different scales.
  • Object Segmentation and Feature Representation: Superpixel segmentation (using SLIC) is employed to create image objects. For each object, a hierarchical fuzzy histogram is constructed based on fuzzy clustering of the EMAP features. This object-level representation aims to overcome the limitations of pixel-based methods.
  • Semi-Supervised Classification: A novel strategy is introduced to generate pseudo-training sets. This involves using a combination of saliency detection (based on spectral residual) and a Morphological Building Index (MBI) applied to the difference image. Saliency highlights potentially changed regions, while MBI specifically targets buildings. Objects that strongly overlap with salient regions are considered potential "changed" samples, and those with minimal overlap are considered "unchanged" samples. These pseudo-training sets are then used to train a Random Forest (RF) classifier for object-based change detection.

Datasets

The authors evaluated their method on two sets of datasets:

  1. QuickBird Data: A pair of images acquired over Beijing in September 2002 and November 2003, with a spatial resolution of 1m, cropped to 472x472 pixels. This dataset was previously used in reference [16].
  2. Google Earth Data: Five pairs of bi-temporal images over the Beijing urban area. These images have three bands in the visible spectrum and a size of 500x500 pixels. The spatial resolutions vary: 1m for four pairs (taken on 30 September 2012 and 4 March 2013; 28 June 2009 and 19 September 2015) and 4m for one pair (taken on 4 March 2013 and 12 November 2014).

Key Insights and Ideas

  • EMAPs for HR Imagery: EMAPs prove effective in capturing rich spatial and spectral information from HR remote sensing images, overcoming some limitations of traditional feature extraction methods.
  • Hierarchical Fuzzy Histograms for Objects: Representing objects with hierarchical fuzzy histograms allows for a more robust capture of spectral and spatial information compared to simple object features or traditional bag-of-words approaches.
  • Saliency and MBI for Pseudo-Training: Combining saliency detection and MBI is a clever way to automatically generate reliable training samples for building change detection, significantly reducing the need for manual labeling. This strategy effectively filters out false positives and negatives, guiding the semi-supervised classification.
  • Object-Based Advantage: The object-based approach, particularly when combined with superpixel segmentation, is shown to be more robust to issues like misregistration and noise compared to pixel-based methods.

Key Results

The proposed method demonstrated promising performance across various datasets and was compared against seven widely-used CD algorithms.

  • Quantitative Performance: The method achieved competitive results, often outperforming others in terms of Overall Alarm Rate (OAR) and Kappa coefficient, indicating a good balance between detecting changes and avoiding false alarms. For instance, on the first dataset, the proposed method achieved a kappa of 0.5167, which was the highest among all tested methods. On the second dataset (Image 1), it achieved a kappa of 0.7458, also the highest.
  • Qualitative Performance: Visual results showed that the proposed method generated fewer scattered changed areas and avoided fragmented detections compared to some traditional methods. It successfully detected most changed building areas while minimizing false alarms.
  • Parameter Sensitivity: The study analyzed the sensitivity of key parameters (saliency threshold, MBI threshold, and overlapping ratio) and found that while saliency and MBI play crucial roles, overly strict thresholds can negatively impact performance.

Limitations and Future Work

The authors acknowledge several limitations and suggest avenues for future research:

  • High Missed Alarm Rate (MAR): The method still exhibits a relatively high MAR, meaning some changed areas are missed.
  • Future Work: Incorporating a priori information, such as shadow information (to identify buildings and filter shadow-induced changes) and vegetation indices (to differentiate backgrounds), could help improve reliability and representative sampling.
  • False Detections from Spectral Differences: The approach can still produce a few false detections due to spectral differences within the same class.
  • Future Work: Exploring color space transformations and normalization as preprocessing steps could mitigate this.
  • Feature Extraction and Difference Measures: Further research is needed to extract more discriminative features and develop more robust difference measures for object-based change features.
  • Automatic Parameter Selection: Developing methods for the automatic selection of parameters is a key area for future investigation.

Summary updated

BibTeX
@misc{hou2016saliency,
  author = {Bin Hou and Yunhong Wang and Qingjie Liu},
  title = {{A Saliency Guided Semi-Supervised Building Change Detection Method for High Resolution Remote Sensing Images}},
  year = {2016},
  doi = {10.3390/s16091377},
}
Figure