InfoSequia: Towards an operational satellite-based Drought Early Warning and Forecasting System for quantifying risks of crop and water supply by using machine learning and remote sensing

Droughts have directly affected at least 1.5 billion people in the last century, generating economic losses up to $124 billion. They are a recurrent, creeping meteorological hazard that may endanger the water and food security of large regions. The frequency and severity of droughts are expected to increase with climate change, especially in Africa, Central America, and also in Europe where annual losses may multiply by 7 and represent up to 2 times the size of the European economy in the medium-long term.

Drought Early Warning Systems (DEWS) are key pillars of a risk-based, proactive management strategy. The increasing number of sources of EO data has remarkedly improved the monitoring capabilities of DEWS. Despite there are good examples of global and regional drought monitoring systems, these tools still lack of seasonal forecasting capabilities able to provide enough accurate and specific predictions of drought impacts at the subregion level (e.g. basin, district). These deficiencies constitute a challenge for the scientific community and provide an opportunity to improve the current services.

To address this gap in the DEWS landscape, the InfoSequia DEWS is developed to integrate the strengths of spatial, satellite-derived data with machine learning techniques for seasonal forecasting. InfoSequia consists of two modules:

• InfoSequia-MONITOR provides more than 50 drought predictors including meteorological (SPI, SPEI), vegetative (VCI / TCI / VHI), and hydrological (water level in reservoirs, groundwater storage status) drought indices, as well as atmospheric oscillation indices, all of them retrieved from satellite (e.g. MODIS, Sentinel-2, Sentinel-3, GRACE), hybrid (eg CHIRPS), or reanalysis and modeling (ERA5-Land) products. All indices are obtained from dekad values ​​ which are timescale aggregated at 1, 3, 6 and 12 months. The spatial resolution of the indices ranges from 5km (SPI, SPIE) to 250m (VH indices).
• Acknowledging the limitations of physically-based modelling on the seasonal time scale, the InfoSequia-4CAST module rests on the Fast and Frugal Tree (FFT) algorithm, a machine learning technique in which binary decision trees are trained and generated at the subregional level with the historical and spatially-aggregated predictors of drought. Final outputs are delivered in the form of monthly warnings of risk of failure up to 6 month lead times.

All InfoSequia algorithms run on a cloud platform, with cloud geoprocessing functionalities.

With support of the European Space Agency (ESA), InfoSequia is being developed and piloted to provide operational seasonal forecasts of: a) crop yield failures at the district level in Mozambique, and b) water supply failures in the Segura river basin in SE Spain.

Seasonal outlooks of drought impact support improvement of the water and food security of a region by allowing the early exploitation of groundwater reserves or unconventional water resources (desalination, reuse), the optimal water allocation of limited resources among users, or the implementation of ex-ante cash transfers or food vouchers. This research introduces the general workflow which underpins InfoSequia, how limitations due to technical barriers and data gaps are addressed, and the key performance indicators generated for both pilot cases.