NDVI time series stochastic models for the forecast of vegetation dynamics over desertification hotspots

Land degradation in semi-arid natural environments is usually associated with climate vulnerability and anthropic pressure, leading to devastating social, economic and environmental impacts. In this sense, remotely sensed vegetation parameters, such as the Normalized Difference Vegetation Index (NDVI), are widely used in the monitoring and forecasting of vegetation patterns in regions at risk of desertification. Therefore, the objective of this study was to model NDVI time series at six desertification hotspots in the Brazilian semi-arid region and to verify the applicability of such models in forecasting vegetation dynamics. We used NDVI data obtained from the MOD13A2 product of the Moderate Resolution Imaging Spectroradiometer sensor, comprising 16-day composites time series of mean NDVI and NDVI variance for each hotspot during the 2000-2018 period. We also used rainfall measured by weather stations as an explanatory variable in some of the tested models. Firstly, we compared Holt-Winters with Box-Jenkins and Box-Jenkins-Tiao (BJT) models. In all hotspots the Box-Jenkins and BJT models performed slightly better than Holt-Winters models. Overall, model performance did not improve with the inclusion of rainfall as an exogenous explanatory variable. Mean NDVI series were modelled with a correlation of up to 0.94 and a minimum mean absolute percentage error of 5.1%. NDVI variance models performed slightly worse, with a correlation of up to 0.82 and a minimum mean absolute percentage error of 22.0%. After the selection of the best models, we combined mean NDVI and NDVI variance models in order to forecast mean-variance plots that represent vegetation state dynamics. The combined models performed better in representing dry and degraded vegetation states if compared to robust and heterogeneous vegetation during wet periods. The forecasts for one seasonal period ahead were satisfactory, indicating that such models could be used as tools for the monitoring of short-term vegetation states.