A Markov regime-switching regression approach to modelling NDVI from surface temperature and soil moisture

Building accurate relationships between vegetation amount and climatic variables is helpful in understanding and informing sustainable environmental management. The common approach in this regard is to develop a generic, linear relationship or season-dependent relationships. Such approaches, however, fail to hold if data characteristics deviate from expected patterns. This study applied a regime-switching regression model, namely the Markov-switching (MS) approach, to predict time-series Normalized Difference Vegetation Index (NDVI). This was done using surface temperature, soil moisture and the interaction of surface temperature and soil moisture as regressors at monthly temporal resolution. Modelling was executed at the biome spatial (broad vegetation categories) scale. The results showed that the MS approach captured the non-linear dynamics in the data for each of the eight biomes considered in the study. The accuracy of the MS approach compared to non-switching modelling approach was evident in model comparison criteria including significance of parameter estimates, coefficient of determination (R²), Akaike Information Criterion (AIC), Bayesian Information Criterion (BIC) and log-likelihood as well as post-modelling diagnostics such as residual plots, autocorrelation function (ACF) and partial autocorrelation function (PACF) of residuals, and squared residuals. Overall, the study clearly demonstrates the superiority of MS modelling that captures non-linear relationships that may not be modelled using conventional non-switching modelling. Further studies are encouraged to test the approach at larger spatial scales.