Forecasting and Prediction
The remote sensing tools and products that have been presented in the previous sections of this chapter characterize current drought conditions that are appropriate for drought monitoring purposes, but are somewhat limited for early warning applications that typically require information about future conditions to give decision makers lead time to implement drought mitigation actions (Tadesse et al. 2016). Satellite remote sensing has proved to be an effective means for real-time drought monitoring, as demonstrated by the numerous tools summarized in this chapter that have emerged over the past decade. However, the development of complementary remote sensing-based tools for drought forecasting and prediction has been very limited until recently, with efforts built upon the previous work of several monitoring tools (presented earlier) to provide projections of future drought conditions.
The vegetation outlook (VegOut) is a vegetation condition prediction tool that was developed experimentally for the CONUS by Tadesse et al. (2010) and is currently being implemented for Ethiopia and the Great Horn of Africa region in work underway by Tadesse et al. (2014). VegOut builds upon the VegDRI modeling methodology presented earlier by applying the same regression tree analysis technique to satellite-based VI observations, climate-based drought index data, biophysical information (e.g., land use/land cover, soils, and elevation), and several oceanic indicators. The empirical VegOut models are based on the historical analysis of the relationship between the VI-based vegetation conditions and the preceding climate conditions represented in the climate-based SPI input and the teleconnection signal from several oceanic indicators (e.g., Pacific decadal oscillation [PDO] and Atlantic multidecadal oscillation [AMO] indices) while considering the biophysical characteristics of a location such as land cover and elevation. The rationale underlying the VegOut is that time-lagged relationships exist between vegetation response and prior climatic and oceanic conditions. The historical analysis of these variables for a 20+ year period using a regression-tree-based data mining method is used to reveal these historical interactions and develop models that predict future vegetation conditions at multiple time steps (e.g., upcoming 1-3 months). The specific vegetation condition measure being estimated by VegOut is the standardized NDVI (SDNDVI), which is a standardized NDVI value calculated from the historical time-series NDVI data using a z-score approach that represents seasonal greenness of a given time period during the growing season compared to the historical average greenness conditions for the same period. Work by Tadesse et al. (2014) for the central United States and Tadesse et al. (2010) for East Africa showed that VegOut had a reasonable predictive accuracy of forecasting future vegetation conditions for outlook periods spanning 1-3 months. This work found that correlations between predicted and observed conditions were generally greater than 0.70 for the shorter outlook periods and 0.60 for the longer 3 month period over the central United States and Ethiopia. An operational VegOut-Africa tool producing dekadal (10-day) updates of 1-, 2-, and 3-month outlook maps is currently being developed for the Greater Horn of Africa region.
Another drought forecasting approach using remotely sensed data inputs has been developed by Otkin et al. (2015) using the rapid change index (RCI) data derived from the ET-based ESI presented earlier in this chapter. The RCI is designed to detect unusually rapid decreases in the ESI that are indicative of either rapid onset drought events or changes in drought severity. Otkin et al. (2015) expanded the RCI concept to include a forecasting component by using linear regression between the RCI and USDM drought severity classes to compute drought intensification probabilities based on the current RCI value. The initial RCI forecasting work showed that the probability of drought development and/or intensification over subseasonal time scales is higher than normal when the RCI is negative. Figure 10.5 demonstrates the utility of these forecasts as a drought early warning tool for a drought event that occurred across the central United States during 2002. On June 3, high drought intensification probabilities were present across most of South Dakota where the drought severity subsequently intensified by up to three USDM categories during the next 4 weeks. By July 1, a long band of high probabilities had developed from southwestern South Dakota into eastern Kansas. As was the case earlier in the summer, these high probabilities occurred several weeks before a period of rapid drought intensification, with some locations experiencing a two-category increase in USDM drought severity by the end of July. This example also illustrates the strong correspondence between regions experiencing rapid decreases in the ESI (as depicted by negative RCI values) and rapidly deteriorating crop conditions. This example demonstrates that statistical regression methods that combine drought early warning signals in remote sensing datasets such as the ESI with information from other variables can produce useful probabilistic forecasts of drought development.
Evolution of the USDM drought depiction, USDA topsoil moisture and crop condition anomalies, rapid change index, and the probability of at least a one-category increase in USDM drought severity over a 4-week period or at least a two-category increase over an 8-week period from June 3 to July 29, 2002.
As remote sensing of the land surface has led to an improved characterization of the initial land surface states through advanced data assimilation techniques that were discussed earlier, a need to improve short-term forecasts of drought intensification or recovery has emerged. While the drought community has largely focused on seasonal forecasts of drought, which are mainly tied to the prediction of the large-scale circulation patterns (e.g., El Nino southern oscillation [ENSO]), forecasts in the 1- to 3-week range can provide actionable information, especially in the agricultural sector, where decisions are still made within such a timeframe. Through the combination of advanced land data assimilation methods and ensemble-based numerical weather prediction models (e.g., from operational centers such as National Climate Prediction Center [NCEP], European Centre for Medium-Range Weather Forecasts [ECMWF], Canadian Meteorological Centre [CMC], and United Kingdom Meteorological Office [UKMET]), a probabilistic forecast of drought improvement or recovery is possible. Operational numerical weather prediction (NWP) systems currently produce hundreds of 10- to 15-day forecasts each day, which can be used to force a land surface model ensemble toward improving drought forecasts. Such a system would ingest all available land surface remote sensing inputs through data assimilation in a land surface modeling framework that would then be forced with all available bias-corrected NWP ensemble-based forecasts. This land surface model ensemble could then be used to produce probabilistic forecasts of variables such as changes in soil moisture, which are closely tied to drought intensification or recovery.