In our study area, a savannah ecosystem, provisioning services (food production) and regulating services (e.g. erosion prevention) are in conflict. The main motivation of this study is to assess the temporal changes of land cover and the feasibility of two change detection techniques. We aim at responding to the following research questions: Are results of a bi-temporal change detection supported by results of a change detection method using full temporal resolution? What are the limitations of both methods in an ecological context?
Material and Methods
We analysed the Normalized Difference Vegetation Index derived from MODIS images from 2002-07-04 to 2017-03-30. As a change detection technique using full temporal resolution Breaks for Additive Seasonal and Trend (BFAST) was used. BFAST decomposes time series into trend, seasonal and remainder components and then iteratively detects break points. A Change Vector Analysis (CVA), calculating pixel differences between bands, was used as a bi-temporal change detection method.
BFAST detected trend changes in approximately half of the analysed pixels. Positive magnitudes for biggest trend breaks tended to coincide with an increased precipitation amount in the main rainy season. In contrast, BFAST identified 10 pixels only with seasonal breaks. Using an alternative test and significance level hardly affected the number of detected breaks in the seasonal component. Constantly increasing amplitudes or short-term seasonal variations were ignored by BFAST. CVA results are to be evaluated.
We found a disagreement between the results of the BFAST analysis and the visual inspection of the time series. In particular, changes of the amplitude in the seasonal component are hardly ever detected as changes. Further analysis needs to determine which ecosystem behaviour remains undetected by BFAST and hence whether it is suitable for change analyses in a savannah ecosystem.