Crop Yield Forecasting

type of analysis, it is also possible to envisage a forecasting error, considering that the similar
seasons, if any, are a sample of all similar seasons. Similarity analyses can be performed
through the following: visual graphics, simple statistical techniques, principal components
and cluster analyses. One or more agro-climatic factors, such as rainfall, temperature or
vegetation indicators (e.g. the NDVI or the DMP) can be used for the analysis. Similarity
analyses are used to forecast cereal yields as early as the month of February, using rainfall
or the NDVI. Because the climatic conditions of the months of March and April are decisive
for final cereal yields, signs of bad or good cereal harvests may be observed as early as the
end of February.
The linear regression analysis uses the ordinary least-squares method to identify the
relationships between crop yields and agronomic, environmental or economic variables,
which are used as predictors. Predictors can be either agronomic factors (variety,
temperature, fertilizers, irrigation, etc.), climatic factors (rainfall, temperature, humidity, etc.),
environmental indicators (NDVI, water balance, etc.), or economic indicators (price, cost,
accessibility, etc.). Indicators are used in crop forecasting because they can complete the
agronomic and climatic factors that are directly used. For example, the NDVI is a measure
of the vegetation’s vigour and is the result of measurements of rainfall, temperature, soil
water balance, the agronomic techniques used, etc. However, it is difficult to quantify the
contribution of each factor to the NDVI. The predictors can be either quantitative variables
(rainfall, temperature, NDVI, etc.) or qualitative variables (presence or absence of drought,
etc.), used separately or in combination. In Morocco, the predictors to be considered for
cereal yield forecasting are: the rainfall, the temperature and the vegetation index (NDVI and
derived indices). They can be computed for the entire cropping cycle or for parts thereof, and
can be used in a simple linear regression model with yield, or in combination, in multiple
regression equations.
The results of the aggregated crop simulation at regional level are the indicators that are used
for crop yield forecasting by means of the CST. The results from the CGMS-MAROC at Level
2 are sent to the CST database and, at the same time, further indicators are added, which
are derived directly from the weather data and from satellite processing chains at VITO. The
CST incorporates a tool to perform linear regression and similarity analyses for cereal yield
forecasting (see Figure 3.6 below). The estimate of the forecast error is a fundamental quality
of a crop forecasting system. For example, the JRC considers that the forecast error is low
when it is lower than 3 percent and no higher than 6 percent (Genovese & Bettio 2004).
The European CGMS’ forecasting system provides figures for the yields of major crops in
Europe with an error between 3 and 5 percent (8.6 percent for wheat). The accuracy of
the CGMS-MAROC’s cereal yield forecasting (see Figure 3.7 below) is continuously refined,
because it depends upon four essential parameters that are subject to constant improvement:
• The number of years during which data is recorded, which helps to build better
databases that account for more diverse situations; this, in turn, can train the
regression models;
• The number of meteorological stations within the DMN’s network, but also within
the Ministry of Agriculture and Marine Fisheries and the INRA; this will enable better
assessment of climatic conditions;
Crop Yield Forecasting: Methodological and Institutional Aspects 97