Towards a Hydrological Drought Index for anticipatory action against hunger

MARCH 29, 2023

Lucas Kruitwagen, Francisco Dorr, Chris Arderne, Alex Money, Amine Baha, Fiona Huang, Nicolas Longépé

Results from OxEO's collaboration with the ESA Phi-Lab and the World Food Programme as part of the EO & AI for SDGs Innovation Initiative.

A record 349 million people across 82 countries face acute food-insecurity, with climate-exacerbated drought a leading contributor [1]. Insufficient water prevents or reduces crop growth, leading to food shortages with dire impacts on human health and welfare. As part of the EO & AI for SDGs Innovation Initiative, Oxford Earth Observation (OxEO) collaborated with WFP and the European Space Agency (ESA) Φ-lab to improve the prediction of drought events in Mozambique and Zimbabwe. We developed a Hydrological Drought Index (HDI) and tested its increased predictive capability over a baseline Standard Precipitation Index (SPI). A SPI is based on monthly forecasted precipitation anomalies. Our HDI extends the SPI by adding historic measurements of surface water availability and soil moisture. We find that the HDI can improve the forecasting of vegetation growth in agricultural areas by an average of 20% over a conventional SPI in the short term.

The HDI is developed by combining Earth observation data with the meteorological forecast. Our intention is to predict hydrological drought, which refers directly to shortfalls in water availability for end uses (whereas meteorological drought refers to precipitation anomalies only)[2]. A key motivation for exploring markers of hydrological drought was to increase the longitudinal utility of the prediction. Forecasts based on meteorological drought capture near-term weather events, but a hydrological forecast additionally captures observation data on surface water extents and soil moisture, providing a more complete and durable understanding of likely conditions on the ground. These insights are particularly valuable given the anticipatory nature of humanitarian and development work. By extending the lead time of drought predictions and providing a more comprehensive understanding of local hydrological conditions, agencies can work more effectively for the benefit of target communities.

As a proxy for crop growth and yield[3], we use a normalised-difference vegetation index (NDVI) which can be measured continuously since the mid-1980s using the Landsat-5, -7, -8, and Sentinel-2 satellite constellations. We used priority level-2 (i.e. municipality) administrative areas in Mozambique and Zimbabwe for the study. For each administrative area, we obtained representative land cover maps from DynamicWorld[4], vectorising a set of agricultural areas in each administrative area. These areas are shown in Figure 1 and are coloured by both predictive performance and improvement of the HDI over the baseline SPI.

Figure 1: select agricultural areas colored by predictive performance for study administrative areas, both absolute (a) and relative to the baseline SPI (b).

The NDVI of each agricultural area was used as the target of our prediction problem. NDVI isn’t a perfect proxy for crop growth (which in turn is a proxy for food insecurity only in limited contexts), but provides a high-resolution signal, both spatially and temporally, to be used in our modeling framework. To examine the relevance of NDVI to other signals of drought and food insecurity, we compare our NDVI signal to i) the Water Requirement Satisfaction Index (WRSI)[5], ii) FEWS-NET food insecurity activations, available on a four-month cadence[6]; and iii) our baseline model, the SPI. Figure 2 shows the correlations between these four proxy measurements of food insecurity.

Figure 2: Cross-correlation between WRSI, SPI, 3-month lagged SPI, mean NDVI anomaly, and FEWS-NET food insecurity activations for 21 administrative areas

We measure soil moisture using Sentinel-1 imagery, adapting the algorithm of Bhogapurapu et al[7]. We add to this the baseline precipitation forecast available from the European Centre for Medium-Range Weather Forecasts[8]. NDVI, soil moisture, and forecast data are each rendered as a time series for individual agricultural areas in priority administrative areas in Mozambique and Zimbabwe. We obtain surface water extents for 155 small- and medium-sized reservoirs in Mozambique and Zimbabwe using a convolutional neural network trained on Landsat-5, -7, -8, and Sentinel-2 imagery, applied to water bodies accessed through Open Street Map[9]. Surface water extents measurements for each agricultural area are allocated according to the anomaly in surface water extents of upstream reservoirs, weighted by inverse distance. With this ensemble of features, we predict NDVI for each agricultural area up to 7 months in advance. Figure 3 shows cross-validation predictive performance across all agricultural areas.

Figure 3: Predictive performance of additive feature combinations showing distributions for all agricultural areas across forecast horizons.

It is only at the one-month horizon that a prediction model trained with the HDI feature combination clearly outperforms a conventional SPI prediction model across all agricultural areas. Figure 1 shows certain high-performance hot spots, including Chiredzi, Beibridge, and Mwenezi in Zimbabwe, and Chibuto and Macossa in Mozambique. The improvement varies across agricultural areas, with improvements over 50% across all forecast horizons for certain agricultural areas. More study is now required to obtain longer, more accurate timeseries of soil moisture, NDVI, and water extents, and further analysis is needed of the complex relationships between these variables. This predictive task can then be approximated by heuristical rules to obtain a useful, simple, and predictive hydrological drought index.

The source code for the machine learning model training and inference[10], computation orchestration pipeline[11], and API service[12] have all been made available open-source. Our pipeline is built and hosted on AWS Web Services, is orchestrated using Prefect Cloud, and uses the STAC endpoints provided by SentinelHub and generously funded by an ESA’s Network of Resources project sponsorship. It was a fruitful collaboration amongst OxEO, WFP, and the ESA Φ-lab. As the next step, we plan to further test the interpretability, reliability, and scalibility of the HDI in various countries. Analytical tools such as this are critical for building parametric humanitarian anticipatory action protocols that could lead to providing better support to the people we serve.

[1]: Anthem, P. (2022) WFP and FAO sound the alarm as global food crisis tightens its grip on hunger hotspots, The World Food Programme, https://www.wfp.org/stories/wfp-and-fao-sound-alarm-global-food-crisis-tightens-its-grip-hunger-hotspots.

[2]: E.g. Gumus, Veysel, and Halil Murat Algin. "Meteorological and hydrological drought analysis of the Seyhan− Ceyhan River Basins, Turkey." Meteorological Applications 24.1 (2017): 62-73.

[3]: Wall, L., Larocque, D. and Léger, P.M., 2008. The early explanatory power of NDVI in crop yield modelling. International Journal of Remote Sensing, 29(8), pp.2211-2225

[4]: Brown, Christopher F., et al. "Dynamic World, Near real-time global 10 m land use land cover mapping." Scientific Data 9.1 (2022): 1-17.

[5]: WRSI calculated using the CHIRPS (https://www.chc.ucsb.edu/data/chirps) and EDDI (https://www.chc.ucsb.edu/data/eddi) datasets as principal sources

[6]: FEWS NET. 2021. Food Security Classification System. Washington, DC: FEWS NET.

[7]: Bhogapurapu, N., Dey, S., Homayouni, S., Bhattacharya, A., Rao, Y.S. (2022) Field-scale soil moisture estimation using sentinel-1 GRD SAR data, Advances in Space Research.

[8]: Johnson, Stephanie J., et al. "SEAS5: the new ECMWF seasonal forecast system." Geoscientific Model Development 12.3 (2019): 1087-1117.

[9]: OpenStreetMap: https://www.openstreetmap.org/

[10]: oxeo-water: https://github.com/oxfordeo/oxeo-water

[11]: oxeo-flows: https://github.com/oxfordeo/oxeo-flows

[12]: oxeo-api: https://github.com/oxfordeo/oxeo-api