Risk of drought for agriculture in Morocco is increasing due to dual pressure of decreasing and fluctuating precipitations and increasing domestic and industrial needs. This risk has to be considered and managed to insure food security. Early Warning Systems, and particularly agro-meteorological crop yield models, constitute early decision-making tools to warn of production drop due to drought. At continental to global scales, operational yield forecasting systems exist, which do provide timely estimates on the yields of the major crops in Morocco, but in a rudimentary way and only at national level. However, no specific system exists for Morocco and so far the official production estimates for the major crops are based on costly field surveys during May to September period and final results are published during the next crop season in July to October. The objective of this study is to prospect the feasibility of accurate early prediction models for wheat (Triticum aestivum L.) grain yields in Morocco, as it is by far the most consumed and cultivated crop. The challenge consists on elaborating prediction models to be used in an operational mode by decision-makers, based on reliable approaches and easily available agro-climatic indices or weather data.
The used approach to predict yields consists on using 4 different methodologies, depending on data availability: (1) Ordinary Least Squares (OLS) regression models, using only seasonal rainfall as predictor, (2) OLS regression models, using Seasonal rainfall, temperature and Normalized Difference Vegetation Index (NDVI) as predictors, (3) Artificial Neural Network (ANN) analysis, using Seasonal rainfall and NDVI as predictors and (4) AgroMetShell (AMS) water balance model, that uses dekadal rainfall and potential evapotranspiration as inputs. Seasonal rainfall, temperature and NDVI were used, as they display strong associated inter-annual variation with wheat yields in Morocco.
The first methodology is based on empirical OLS regression models to predict grain yields at province level, using only rainfall information. Grain yield was predicted at national level with 11.1% (119 kg.ha-1) error, when taking into account all the information at provincial level. The predicted error depends on province and can range from 67 kg.ha-1 to 595 kg.ha-1. Seemingly Unrelated Regressions (SUR) were used, as an original methodology in agro-meteorology, to improve these predictions taking into account spatial information for a set of neighbor provinces. SUR models improved yield predictions, as R2 between observed and predicted grain yields increased from R2=90.7%*** to R2=91.3%***. The models used are promising and could be applied for predicting wheat grain yields in Morocco if only rainfall information is available.
The second methodology keeps the same OLS modelling methodology to predict wheat yields but using seasonal NDVI and temperature information in addition to rainfall, still at both provincial and national levels. The predictions used dekadal (10-daily) NDVI/AVHRR, dekadal rainfall sums and average monthly air temperatures. The global land cover map GLC2000 was used to select only the NDVI pixels that are related to agricultural land. This second methodology is simpler and more accurate, comparatively to the first methodology, mainly thanks to NDVI information. At province and country levels most of the yield variation was accounted for by NDVI. Provincial wheat yields were assessed with errors varying from 80 to 762 kg.ha-1, depending on the province. At national level, wheat yield was predicted at the third dekad of April with 6.8% (73 kg.ha-1) error, using NDVI and rainfall. However, earlier forecasts are possible, starting from the second dekad of March with 84 kg.ha-1 error. The proposed models can be used in an operational context to forecast wheat yields in Morocco when NDVI is available in addition to rainfall.
The third methodology attempted to use ANN to predict yields using NDVI and rainfall information. ANN were used as they seem to have a great potential as they can theoretically deal with linear or even non linear relations for various levels of complexity, without any a priori assumption regarding the processes involved. This third methodology was compared to the second methodology, but only at country level. Multiple linear regression models performed better that ANN analysis for predicting wheat yields in Morocco. National wheat grain yields could be forecasted with 73 and 94 kg.ha-1 errors in validation at the second dekad of April, respectively using OLS models and ANN analysis. The lower performance on ANN analysis was probably due the linearity between yields and the predictors (NDVI and rainfall) and to the shortness of the used times series, in respect to the high year-to-year variation of NDVI, rainfall and yields.
The fourth methodology is based on AMS software, which derives agro-climatic indices, based on rainfall and evapotranspiration information. Three indices (Water Surplus Deficit, Water Requirement Satisfaction Index and Soil Water Storage), derived from AMS were correlated to wheat yields in Morocco. A fourth and new index was added to the evaluation, calculated as the integration of the Water Surplus Deficit index over dekads from the start of season in November. Two sample provinces, Meknès and Safi, located respectively in a sub-humid and a semi-arid agro-ecological zone were considered. Amongst the 4 indices, the Soil Water Storage (SWS) and the Water Requirement Satisfaction Index (WRSI) were the most correlated to wheat grain yield at the dekad level, respectively in Meknès (maximum R2=82% at 1st dekad of March) and Safi (maximum R2=67% at 1st dekad of April). AgroMetShell performance was also compared to the NDVI provided by NOAA-AVHRR. The four AgroMetShell indices were better correlated to yield than NDVI in Meknès, contrarily to Safi where NDVI was the best index. Regression models for predicting wheat yield at the province level were determined based on AMS and NDVI indices. Models combining AMS and NDVI explained 79% (330 kg.ha-1 error) and 92% (110 kg.ha-1 error) of wheat yield variability respectively in Meknès and Safi, even if most of grain yield variability was accounted for by NDVI in Safi. AMS seems to be an interesting tool in Morocco for wheat monitoring when no NDVI information is available in a sub-humid province such as Meknès.
At national level, potential improvements could stem from the inclusion of non-weather predictors (diseases, pests, soils, and irrigation) and water balance calculation in the proposed models. In addition, the predictions could be improved using better NDVI quality, derived from SPOT-vegetation for example, instead of NOAA/AVHRR. Our preliminary evaluation (not shown) of the SPOT-vegetation sensor, displayed better correlations between wheat yields and NDVI/SPOT in Morocco for the 1988-2004 time period. However, SPOT time series is actually too short for modelling use and, pooling these two datasets is restricted, as poor correlations were found between NDVIs derived from SPOT-vegetation and NOAA/AVHRR sensors. At provincial level, R2 widely ranged from 24 to 98.5% using only rainfall predictors and, from 64 to 98% using NDVI, rainfall and temperature predictors. The use of vegetation information could be more efficient at the provincial level, if higher spatial resolution NDVI images and land covers maps are used. In addition, provincial models should be based on more representative weather stations, as only one synoptic station per province was available in our study.
If adopted, the proposed models will certainly help policy-makers to warn populations for drought and plan well in advance annual imports, ultimately helping for food security of the country. These models are early, fast and low costly if compared to the actually used surveys-based methodology, needing only real time dekadal rainfall, temperature data and, NDVI images for nine dekads (February to April). The proposed approach is relatively easy to understand and not constraining, as it relies on well robust methodologies and could be adapted according to data availability. The proposed models could be applied by decision-makers to accurately predict wheat yields in an operational context, at both provincial and national levels in Morocco.