Spatio-temporal spectral unmixing of time-series images

Mixed pixels exist widely in remotely sensed images. To obtain more reliable land cover information than traditional hard classification, spectral unmixing methods have been developed to estimate the composition of the mixed pixels, in terms of the proportions of land cover classes. The existing spectral unmixing methods usually require pure spectra (i.e., endmembers) of each land cover class. However, in areas dominated by mixed pixels (e.g., highly heterogeneous areas), it can be a great challenge to extract a large number of pure endmembers, especially for long time-series data. Meanwhile, intra-class spectral variation remains a long-standing issue in spectral unmixing. In this paper, we propose a spatio-temporal spectral unmixing (STSU) approach to address these issues. The proposed method extends spectral unmixing from the traditional spatial domain to the spatio-temporal domain. It exploits fully the multi-scale spatio-temporal information, by using temporally neighboring fine spatial resolution images to detect land cover changes and, further, extracts the proportion information of unchanged mixed pixels required for training. The STSU method is free of the need for endmember extraction, using directly the extracted mixed training samples to construct a learning model, and it accounts for intra-class spectral variation. Therefore, it is a fully automatic method suitable for dynamic monitoring of land cover changes. The effectiveness of the STSU method was validated through experiments on Moderate Resolution Imaging Spectroradiometer (MODIS) data in five different areas. The proposed STSU method provides a new solution for spectral unmixing of time-series data based on the goal of continuous monitoring at the global scale.

Automated summary

The study introduces a spatio-temporal spectral unmixing (STSU) approach, which extends spectral unmixing into the spatio-temporal domain to obtain more reliable land cover information. This method does not require pure endmember extraction and directly uses extracted mixed training samples to construct a learning model, making it suitable for dynamic monitoring of land cover changes.