Forecast warnings are based on C3S multi-system⁽³⁾ single-level total precipitation forecast (https://climate.copernicus.eu/seasonal-forecasts) sourced at the monthly time step. Data includes forecasts created in real-time (since 2017) and retrospective forecasts (hindcasts) initialized at equivalent intervals during the period 1993–2016. Forecasts are global and at 1° x 1° spatial resolution. We use precipitation forecasts aggregated to a monthly temporal resolution with a 6-month forecasting horizon. Real-time forecasts are released once per month on the 10^th at 12 UTC.

Forecast warnings build on the methodology described in the scientific paper Increasing the prospective capacity of global crop and rangeland monitoring with phenology tailored seasonal precipitation forecasts. As compared to the scientific paper which is focusing on adapting seasonal forecasts to ASAP base 500m grid cell phenology, the operational forecast warnings described here are adapted to the ASAP unit average phenology for consistency with standard ASAP warnings and easier interpretation.

Forecast warnings are issued for each ASAP unit and target the specific unit growing seasons⁽⁴⁾, one or two per solar year. Growing seasons, defined by inactive to active, active to inactive transitions and breakpoints, are shown in the top-left panel of the dashboard, under the “Time series” tab.

At the time of analysis and for a given unit under consideration we focus the forecasts on the active growing season (denoted by “ongoing season”) and on the next growing season (denoted by “upcoming season”). It is noted that at the time of analysis an ongoing season may be present or not. The upcoming season can be instead always identified and covered by the forecasts provided that it starts in the next six months.

It is important to note that forecast warnings are linked to ASAP growing seasons although they cover a slightly different time period, denoted as “precipitation relevant period”. The precipitation relevant period starts earlier than the unit season to cover the period when crops growth has not started yet but precipitation may contribute to increase initial soil moisture and early development of the crop. The precipitation relevant period also ends earlier than the unit season to avoid considering precipitation deficit in the final part of the growing season, i.e. crops senescence. Pragmatically, the precipitation relevant period is defined as unit season start – 1 month till unit season end – max(1 month, 25% of the unit season length). This way, we stop either one month before the end of season to ensure no overlap with the start of the next relevant period or, for a season lasting more than four months at the start of the senescence period (as 25 % of the unit total season length is expected to represent the senescence period).

Finally, as seasonal forecasts have monthly temporal resolution, the precipitation relevant period is rounded to the monthly time step.

The Tab provides four main pieces of information: temporal information, area covered by most likely category of precipitation (drier, normal, wetter), category average probability, and category skill score.

Top bar chart: the chart shows the timing of the unit crop or rangeland season (green), the precipitation relevant period for this season (light blue) and the seasonal forecast temporal extent available (olive green). Note that the forecast extent may not fully cover the precipitation relevant period. The red vertical line marks the time of analysis, as selected in the left panel of the Warning Explorer. Dates of the mentioned periods, the percent of precipitation relevant period remaining and the one covered by the forecast are displayed hovering the mouse on the chart.

Bottom pie chart: the most likely category of precipitation (i.e. drier, normal, wetter) is identified in each target (crop or rangeland) grid cell. If the probability of the identified tercile is less than 0.4 or two terciles have equal probability, the grid cell is assigned to an additional category "other". The pie chart shows the most probable categories of the unit (drier, normal, wetter and other) and their respective crop (or rangeland) area percentage. For example, a pie slice with the dry category occupying half of the pie indicates that the this category was identified as most probable in 50% of the target area. In addition, pie slices are color coded with the average probability of the category. The outer ring shows the skill score (Ranked Probability Skill Score, RPSS) associated with the most likely categories. RPSS larger than zero indicates that forecast performs better (i.e. smaller errors) than climatology derived from CHIRPS data. RPSS smaller than or equal to zero indicates no skill as our forecasts are less or equally accurate than simple climatology. Skill is only available in the latitude band covered by CHIRPS (i.e. between 50° N and 50° S). The skill score of the additional category "other" is not shown.

(1) Most likely category and its probability

Categories are computed as terciles of historical forecasts. That is, terciles are computed scrutinizing the ensemble set of the forecasts made in the past (hindcasts) and represent the three 33rd percentiles (pct) of the lowest, the highest and in between (normal) precipitation accumulated throughout the period of interest, for each grid cell in the hindcast dataset. Having 24 years of hindcasts (1993–2016) and taking the ECMWF forecast system as an example, each grid cell always contains 600 values of precipitation forecasted in the past (24 years × 25 ensemble members). The lowest 200 values (precipitation < 33rd pct) are included in the lower tercile and represent below normal precipitation (“drier than usual”), the highest 200 values (precipitation > 66th pct) are included in the upper tercile and represent above normal precipitation (“wetter”), the values in between represent the normal precipitation tercile. Thus, each category (tercile) is characterized by a 33% probability of occurrence. Tercile probabilities for a single prediction system are computed using all its forecast members and by counting how many of the ensemble members fall in each tercile. That is, for each tercile, we sum the number of members falling in that tercile. Then we divide such number of occurrences by the total number of members, resulting in a probability for each tercile. Once the three tercile probabilities of each prediction system are computed, we compute the average probability by tercile. In this way, we assign equal weight to each prediction system, independently of the number of ensemble members produced by prediction system.

(2) Skills

Skill is expressed by the ranked probability skill score (RPSS). It measures cumulative squared error between categorical forecast probabilities and the observed categorical probabilities relative to a reference (or standard baseline) forecast. As we are using the terciles computed from observation climatology (CHIRPS precipitation over the same period) as a reference, RPSS informs if the forecasts are better than simple climatology. RPSS larger than zero indicates better performances (i.e. smaller errors) for forecast than for climatology. RPSS smaller than or equal to zero indicates no skill as our forecasts are less or equally accurate than simple climatology. Skills are computed for the multi-system and for each forecasting system for internal check (not shown on the website) in the latitude band covered by CHIRPS data (i.e. between 50° N and 50° S).

(3) C3S multi-system

We use contributions from the following 9 prediction systems: ECMWF, UK's Met Office, Météo-France, the German Weather Service (DWD), the Euro-Mediterranean Center on Climate Change (CMCC), the US National Weather Service's NCEP, the Japan Meteorological Agency (JMA), Environment and Climate Change Canada (ECCC) and the Australian Bureau of Meteorology (BOM).

(4) Spatial unit growing seasons

Unit seasons definition fully is described in ASAP warning classification scheme v 8.0. In short: multi-annual average satellite-derived phenology (start and end of season) is first computed at the 500 m ASAP base resolution. In the time period between start and end of season, the grid cell is considered “active”. A unit season is deemed to start when 15 % of the target area (cropland/rangeland) is active and to end when less than 15 % is active. In addition to such unit definition, forecast warnings may further divide a single ASAP season in two sub-season using “breakpoints”. Breakpoints mark the transition between one season to another when it is likely that two seasons actually coexist but are not separated by an inactive period.

Typically, small admin units show a clear dominance of phenological pattern: monomodal or bimodal (i.e. one or two growing season per year). In such cases, ASAP will nicely depict the unit level seasons. However, two exceptions exist: i) presence of two seasons and the end of the first overlapping with the start of the second (in different areas); ii) presence of both monomodal and bimodal areas in the unit.

The presence of both monomodal and bimodal areas may have different origins. In Sub Saharan African countries, the Intertropical Convergence Zone (ITCZ) oscillates from South to North and back to South during the solar year. This results in two distinct growing seasons in the South and a single and longer one in the North. In the transition zone, both seasons may coexist in a single ASAP unit that will likely be characterized by a long season mixing the two types of modalities. This can happen also where different crops (with different sowing dates) coexist on different fields in the same ASAP unit. This is the case in Europe where winter and spring/summer crops coexist in the same region at a certain period of the year. In these cases the unit is active (i.e. has at least 15 % of active pixels) during a long crop season actually resulting from two crop seasons. In these cases, we opted to split the unit single active season with breakpoints marking the end of season for a large part of the cropland pixels.

Breakpoints are determined using a set of fixed rules based on the unit-level season definition (time of inactive to active and active to inactive transitions) and the temporal profile of the percentage of crop/rangeland pixels ending the season. Under certain conditions (e.g. percentage greater than 25 % of total area in a two-month period, resulting new seasons of three months minimum), a breakpoint is identified and used as a splitter of the original ASAP season unit in two seasons.

Breakpoints, together with transition inactive to active and active to inactive, define main agricultural seasons at the unit level.