Automated mapping of vegetation trends with polynomials using NDVI imagery over the Sahel
Summary, in English
Over the last few decades, increasing rates of change in the structure and function of ecosystems have been brought about by human modification of land cover, of which a major component is vegetation. Metrics derived from linear regression models applied to high temporal resolution satellite data are commonly used to estimate rates of vegetation change. This approach implicitly assumes that vegetation changes gradually and linearly, which may not always be the case. In order to account for non-linear change in annual observations of vegetation from satellites, we test and apply a polynomial fitting-based scheme to annual GIMMS (Global Inventory Modeling and Mapping Studies)–NDVI (Normalized Difference Vegetation Index) observations for North Africa (including the Sahel) for the period 1982–2006. The scheme divides vegetation change into cubic, quadratic, linear, and “concealed” trend behaviors, the latter indicating that while no net change in vegetation amount has occurred over the period, the curve exhibits at least one minimum or/and maximum indicating that the vegetation has undergone change during the elapsed time period. Our results show that just over half the study area (51.9%) exhibit trends that are statistically significant, with a dominance of positive linear trends (22.2%) that are distributed in an east-west band across the Sahel, thus confirming previous studies. Non-linear trends occur much less frequently and are more widely scattered. Nevertheless, they tend to cluster within or on the outskirts of zones of linear trend, underscoring their importance for detecting anomalous change features. We also show that the ratio of linear vs. non-linear trends tends to be associated with different land cover types/land cover change estimates, many of which reflect biome-level controls on vegetation dynamics. However, more local drivers related to direct human impact, such as urbanization, cannot be ruled out. Our change detection approach retains the more complex signatures embedded in long-term time series by preserving details about change rates, therefore allowing for a more subtle interpretation of change trajectories on a case-by-case basis. The fitting method is entirely automated and does not require the judicious selection of thresholds. However, while polynomials can give a better fit, they like linear models are based on assumptions, and may sometimes lead to oversimplification or miss short-term variations. Our method can help to contribute more accurate information to one of the major goals of the burgeoning field of land change science, namely to observe and monitor land changes underway throughout the world.