Modeling variability and trends in NDVI using a hierarchical Bayesian spatio-temporal approach in semi-arid Kenyan rangeland landscapes

Background/Question/Methods

: Global rangelands support hundreds of millions of pastoralist peoples and are hotspots of mammalian biodiversity. A confluence of increasing variability of climate, changes in fire regime, and fragmentation of movements of livestock and wildlife, all draw the future function and diversity of rangeland vegetation into question. In particular, changes in productivity of grasses, and woody encroachment have strong implications for pastoralist livelihoods and wildlife habitat alike. A number of remotely-sensed approaches using time series analysis of historical NDVI data have been utilized to model rangeland productivity dynamics due to stochastically variable rainfall and temperature, and to distinguish climatic variability from trends due to anthropogenic influence. However these approaches often assume a linear relationship between rainfall and productivity, and rarely consider spatial variability due to gradients of intensity of land use or contextual variables such as soils and topography that likely play an important role in observed trends. We test the ability of a hierarchical Bayesian spatio-temporal linear regression model to predict time series NDVI (AVHRR, MODIS) patterns across two extensive semi-arid rangeland landscapes in Kenya from remotely-sensed estimations of rainfall, temperature, soils, and topography data coupled with detailed understandings of livestock mobility and other seasonal land use patterns.

Results/Conclusions

: We present results from a comparison of several dynamic spatio-temporal models with different levels of hierarchical grouping according to topography, soil, and land use variables. We assess these models in comparison to non-hierarchical models, as well as other common time series models used to assess productivity trends without considering spatial autocorrelation. Through flexible parameterization of lagged rainfall, preliminary results indicate an improved ability to predict productivity through consideration of inter-annual, cumulative variability of daily rainfall, and non-linear regression analysis. Preliminary results also indicate an improved global model by partitioning variance according to soil, topographic, and land use groupings. We expect final results to indicate an enhanced ability to predict productivity patterns for specific contexts across rangeland landscapes from stochastic abiotic variables and other spatial variables such as edaphic contexts and land use gradients. We discuss the implications of these results for the design of models to forecast rangeland conditions under expected changes in temperature and rainfall patterns, as well as under ongoing limitations in pastoralists’ livelihood strategies posed by political and economic policies. In conclusion we discuss the importance of statistical models that are parameterized for appropriate observational scales in semi-arid rangeland contexts with high heterogeneity and high spatio-temporal variability.