Application of satellite image time series and texture information in land cover characterization and burned area detection

Land cover is critical information to various land management and scientific applications, including biogeochemical and climate modeling. In addition, fire is an essential factor in shaping of vegetation structures, as well as for the functioning of savanna ecosystems. Remote sensing has long been an important and effective means of mapping and monitoring land cover and burned area over large areas in a consistent and robust way. Owing to the free and open Landsat archive and the increasing availability of high spatial resolution imagery, seasonal features from the temporal domain and the use of texture features from the spatial domain create new opportunities for land cover characterization and burned area detection. This thesis examined the application of satellite image time series and texture information in land cover characterization and burned area detection. First, the utility of seasonal features derived from Landsat time series (LTS) in improving accuracies of land cover classification and attribute prediction in a savanna area in southern Burkina Faso was studied. Then, the temporal profiles from LTS were explored for mapping burned areas over a 16 year period, and MODIS burned area product was used for comparison. Finally, the application of texture features derived from high spatial resolution data in land cover classification and attribute predictions was investigated in a savanna area of Burkina Faso and an urban fringe area in Beijing. According to the results, firstly, seasonal features from LTS based on all available imagery during one year as input led to a significant increase in land cover classification accuracy in comparison to the dry and wet season single date imagery. The harmonic model used for time series modeling provided a robust method for extracting seasonal features, and the influence of burned pixels on seasonal features could be considered simultaneously. Secondly, the annual burned area mapping based on a harmonic model and breakpoint identification with LTS was capable of detecting small and patchy burn scars with higher accuracy than MODIS burned area product. The approach demonstrated the potential of LTS for improving burned area detection in savannas, and was robust against data gaps caused by clouds and Landsat 7 missing lines. Thirdly, predictive models of tree crown cover (CC) using RapidEye and LTS imagery achieved similar accuracy, indicating the importance of texture and seasonal features from RapidEye and LTS imagery, respectively. Predictions of aboveground carbon and tree species richness, which were strongly correlated with CC, were promising using RapidEye and LTS imagery. Finally, the optimized window size texture classification improved classification accuracy in comparison to the classifications with single window size texture features and multiple window size texture features in an urban fringe area in Beijing, indicating the importance of multiscale texture information.