Spectral Unmixing of Time Series Data to Provide Initial Object Seeds for Change Detection on Google Earth Engine

Nowadays satellite and aerial imagery provide High Resolution (HR) and Very High Resolution (VHR) time series orthophotos. Image classification and object detection is an important and challenging task in urban areas. Object matching in time series images is more challenging and hard to acquire. Accurate object detection is important in change detection methods and plays a significant role in damage assessment algorithms. This issue brought us to develop a method to create initial seeds to improve the change detection in time series images. A part of next to last image is clustered into specified number of clusters. Time series images will construct a multi-spectral image including all images, except the last one. Cluster geometry is utilized to find endmembers in the multispectral image. Space reconstruction is applied to the multispectral image to find initial seeds. Our method could be a part of many conventional image processing algorithms to improve their capability in time-series image analysis. It is implemented on Google Earth Engine, a cloud computing platform for remote sensing. The proposed method is evaluated by comparing the result of change detection in seeded and non-seeded algorithms.