The suitability of existing open data weather data for agro-meteo advisory

With relevant, reliable, timely and accessible weather information, smallholder farmers can make important farming decisions, especially for the semi-arid to arid environments such as those found in Ethiopia. Therefore, weather information, in the form of forecasts and Near Real Time (NRT) observations or estimates, is essential for any farm advisory system targeting smallholder farmers.

Is open data best suited for agro-meteo advisory services to smallholders?

Open data in the weather domain could address the information needs of agro-meteo farm advisory systems. However, is open data ‘fit-for-purpose’; does it match the needs of being reliable, relevant, timely and accessible? Some answers come from the CommonSense project targeting smallholder farmers in Ethiopia.

With relevant, reliable, timely and accessible weather information, smallholder farmers can make important farming decisions, especially for the semi-arid to arid environments such as those found in Ethiopia. Therefore, weather information, in the form of forecasts and Near Real Time (NRT) observations or estimates, is essential for any farm advisory system targeting smallholder farmers. 

The CommonSense project is exactly working into this, by bringing together Communities and Sensors in Ethiopia. The project is funded under the Geodata for Agriculture and Water (G4AW) facility of the Dutch Ministry of Foreign Affairs and executed by the Netherlands Space Office (NSO). 

Although having a broader scope, CommonSense has an important weather service component, which runs in collaboration with the National Meteorology Agency of Ethiopia (NMA) and the Sesame Business Network. Partners in this component include the Dutch weather services company Weather Impact, the Ethiopian IT company Apposit and Wageningen Environmental Research, also based in the Netherlands. One of the aims of the project is to assess what weather information should be part of a national agro-meteorological advisory system. To achieve this, CommonSense works with the Federal Ministry of Agriculture and Natural Resources (MoANR) and other Ethiopian government organisations, like EIAR and RARIs. 

Open weather data sources

Weather data that are used in agricultural advisory systems can be categorised in meteorological observations from stations (NRT and historic archives), simulated weather variables from numerical weather models either for historical periods (so-called ‘reanalysis’) or weather forecasts for the coming days or months, and finally NRT and archive products derived from remote sensing.

Currently, open weather data comes from several sources. Open meteorological observations include the Global Summary of the Day of the US National Oceanic and Atmospheric Administration (NOAA), while open products based on remote sensing include rainfall estimates (such as CHIRPS-USGS CHG) and Incoming Solar Radiation (MSG-LandSAF). Short-term weather forecasts, based on open data, include the Global Forecast System (GFS) from NOAA’s National Centres for Environmental Prediction (NCEP). 

Of course, many other weather data are available, although subject to restrictions. These typically include datasets generated by most national meteorological agencies, but also data from numerical weather prediction models. Examples of the latter are weather forecasts generated by the European Centre for Medium-Range Weather Forecasts (ECMWF). In contrast, ECMWF reanalysis datasets such as ERA-INTERIM and the upcoming ERA5 are available as open data as part of the European Copernicus programme. With reference to forecasts, for instance GFS short term forecasts can be freely obtained, while ECMWF data undergo restrictions and are usually subject to license fees. 

It is therefore important to assess pros and cons, costs and benefits of the datasets in respect to the requirements of the advisory systems envisaged, and ultimately of the smallholder farmers operating in specific environmental conditions. 

Indicators

While it is relatively simple to evaluate the costs associated to the datasets (in terms of license fees or data processing), the benefit side needs much more elaboration since it requires knowledge of end-user requirements and the value added when making use of a specific dataset. 

There are, however, indicators for relevance, reliability, timeliness and accessibility of the data that can be used as proxies for such added value. Consequently, most of the effort would then be in the assessment of compliance of available datasets with the requirements of the system. 

If we take once more short-term weather forecasts as an example, what would be the requirements of the proposed advisory system? What time range and interval, spatial resolution, accuracy (skill) of the prediction would serve the purpose? Is the knowledge of the uncertainty associated to the prediction also important? These are all questions which should be guiding decisions in the choice of the data and that CommonSense has been confronted with.

GFS for instance provides 10-days range predictions, at a 0.25° (~28km at the Equator) spatial resolution. ECMWF provides a short range (1-3 days) as well as a medium range (4-7 days) forecast; it generates both a deterministic high-resolution prediction (HRES) with 0.1° resolution (~11km), and an ensemble one (ENS), with a spatial resolution of 0.20° which corresponds to 22 kilometres. 

Quantifying quality of weather data 

According to the World Bank Ethiopia Socioeconomic Survey 2015-2016, the average field size for the country is 0.13 hectares. Although still far from matching the size of the average Ethiopian fields, the ECMWF forecasts (both HRES and ENS) do provide a better spatial resolution than GFS and are thus considered better fit-for-purpose from this perspective. 

In the context of CommonSense we are also conducting a statistical validation of the accuracies of different weather variables (namely rainfall occurrence and amount) between GFS and ECMWF predictions. This is done by comparing the two against observations from all National Meteorological Agency’s synoptic weather stations. The work is still on-going, but results so far indicate that ECMWF outperforms GFS forecasts not so much in terms of occurrence, but especially in terms of amounts, which is again critical information to support actual farming practices. 

Good interpretation of uncertain forecasts such as rainfall, can only be done if uncertainty can be quantified. A forecast expressed in probability terms, such as ENS, seems therefore more useful than a single prediction (like for HRES or GFS) for making decisions related to farming. For example, planting operations in Ethiopia are highly dependent on the onset of rainfall. Therefore, the forecast can be tailored to provide high certainty that rainfall will fall on the predicted date. Similarly, sesame plants at maturity stage are highly susceptible to damage due to wind and rainfall. Therefore, the forecast can be tailored such that even low probabilities of such conditions in the forecast are reported to farmers. 

A first ‘test bed’ for the forecasts, based on ECWMF predictions, has been a pilot SMS service run over the 2017 growing seasons in the regions of Tigray and Amhara and reached 1,520 users including farmers, extension workers and research staff. Users’ perception of the forecast is positive, which enforces the CommonSense point of view that high quality weather forecasts are a basic requirement for smallholder farmers.

From this perspective the still scarce sources might suggest that open weather data are less fit-for-purpose, which might currently result in opting for paid, non-open services. Nevertheless, it can be expected that more open weather data will become available in the future, hopefully better suited for the purpose of agricultural services for smallholders.