Intra-annual land cover mapping and dynamics analysis with dense satellite image time series: a spatiotemporal cube based spatiotemporal contextual method

Land covers provide essential information for understanding and detecting ecosystem, resources, and environmental dynamics. However, they are generally mapped at coarser temporal scales to study the inter-annual changes, while scant attention has been paid to map intra-annual land cover dynamics at finer temporal scales. Moreover, existing studies are still limited in intra-annual land cover mapping with dense satellite image time series (SITS). Accordingly, this study proposed a novel approach to accurately classify dense SITS for mapping intra-annual land cover dynamics. First, dense SITS is segmented at multiple spatiotemporal scales to generate optimal spatiotemporal cubes (ST-cubes), which are chosen as classification units. Second, the ST-cubes based on spectral, textural, spatial, and temporal features are integratively defined and employed in SITS classification. Third, the spatiotemporal context is modeled by a spatiotemporally extended conditional random field model that measures both spatiotemporal features and semantic correlation between geographic objects. Finally, the proposed method is applied to map the intra-annual land cover dynamics. Comparative experiments of SITS classification are carried out between our method and three existing competitors in a suburban area in Beijing, China, with a dense Sentinel-2 SITS. Moreover, based on the classification results, we analyzed the quantitative intra-annual dynamics of land cover. The result shows that our approach achieves significant improvements in classification accuracy over existing methods, indicating the effectiveness and superiority of the proposed method in mapping intra-annual land cover dynamics with dense SITS.

Automated summary

This study proposed a novel approach to accurately classify dense satellite image time series for mapping intra-annual land cover dynamics. The method involves segmentation of dense SITS to generate optimal spatiotemporal cubes and utilizes spectral, textural, spatial, and temporal features for classification. By modeling spatiotemporal context with a conditional random field model, the approach demonstrates significant improvements in classification accuracy over existing methods.