An Object-Based Change Detection Method Considering Temporal-Spatial Similarity in Long Time Series

A new object-based change detection method has been proposed to address the limitations of existing research based on pixel-based change detection, as well as the neglect of concurrent changes, pixel spatiotemporal information, and the alteration of boundary information due to changes. Additionally considering the concurrence of changes, the measurement of similarity in time series of adjacent pixels needs to take into account the phase shifting and scaling of the time series. In this study, based on the considerations mentioned above, we first constructed time series using available Landsat 5, 7, and 8 data collected in the study area and used a pixel-based change detection method to obtain change information. Then, considering the heterogeneity within objects, we first combined the change information to extract change seeds. After that, the inside_similarity metric was introduced, which is computed by dynamic time wraping (DTW) algorithm, and it was used to impose sequential constraints on the expansion of seeds. Considering that changes can alter both the interior and boundaries of objects, we applied a conditional judgment to all pixels outside the seeds. Through quantitative assessment in three experimental areas, the method proposed in this article improved producer's accuracy (PA) by 6.9%, 2.7%, and 5.5% and user's accuracy (UA) by 6.1%, 3.1%, and 6.6% with F1-score improved by 6.49%, 2.91%, and 7.45% compared to purely pixel-based change detection methods. Combined with qualitative assessment, the object-based change detection method is proved to increase the accuracy of change detection.