Multi-feature fusion and random multi-graph synthetic building change method
WANG Chang, ZHANG Yongsheng, JI Song, ZHANG Lei
2021, 50(2):
235-247.
doi:10.11947/j.AGCS.2021.20200097
Asbtract
(
)
HTML
(
)
PDF (6924KB)
(
)
References |
Related Articles |
Metrics
Change detection of building with remotely sensed image is a challenge work, as it has many issues, e.g. the approach for calculating difference image (DI) which could highlights the changes is not ideal, poor strategies for the training sample collection and low classification accuracy as well. This study analyzed the process from three aspects, namely DI construction, sample selection reliability, and classification method selection, and proposed a remote sensing image building change detection method based on multi-feature fusion and random multi-graphs. First, the spectral and textural features (gray level co-occurrence matrix), morphological building index features, and shape features (after optimum scale segmentation) of multi-temporal, multi-source remote sensing images were extracted. The spectral and textural DI features obtained using change vector analysis, morphological building index DI, and shape feature DI obtained by the subtraction method, were fused to construct the final DI which effectively highlighted the building change information. Second, we obtained the DI saliency map using the frequency-domain significance method. The coarse change detection map was derived by selecting pre-classified thresholds for the DI saliency map (changed pixels “buildings”, unchanged pixels, undetermined pixels) using the fuzzy c-means clustering algorithm to obtain high-quality building and non-building training samples. Finally, the neighborhood features of the non-building and the building were extracted from the remote sensing and feature images, and these were used as the training sample for random multiple training. Subsequently, this trained random multiple classification model was used to perform change detection on the coarse change detection map, resulting in the final change detection map. To verify the efficiency of the proposed method, homogeneous and heterogeneous images were selected for experimental analysis. The results showed that the proposed method could detect more building change information than other methods, and the Com, Cor, and FM values were significantly higher than those of other methods.