Estimation of forage yields of pastures in the climatic spaces of Burkina Faso from satellite data

Estimation of forage yields of pastures in the climatic spaces of Burkina Faso from satellite data

Background : Forage resource assessment is a key element in the governance of livestock food crises in Burkina Faso. Objective-Methods : This study aimed to assess, for the first time, the possibility of estimating forage yields of pastures in the climatic spaces of Burkina Faso through the use of univariate and multivariate linear statistical models constructed from forage plant biomass data collected in the field in 2017, 2018 and 2019, phenological satellite variables (normalized difference vegetation index [NDVI] and fraction of absorbed photosynthetically active radiation [FAPAR]) and agroclimatic variables (precipitation, soil moisture, evapotranspiration, surface temperature). Results : An exhaustive search for the best linear statistical models with one to four variables was carried out and the best models according to the Bayesian information criterion (BIC) were identified. The performance of the univariate to quadrivariate models obtained was quite low with, for all climatic areas except the Sahelian area, RRMSE press varying from 55% to 61% (R² press from 0.07 to 0.36), and for the Sahelian climatic area RRMSE press varying from 42% to 49% (R² press from 0.59 to 0.69). The decrease in correlation of the majority of variables with forage plant biomass according to the north-south gradient results in a decrease in model performance according to this gradient. The agroclimatic variables were found to be useless, and those from FAPAR are generally more efficient than those from NDVI. A very low added value of multivariate models compared to univariate models was observed, except for the Sahelian area. The models developed on more homogeneous climatic areas were found to be more efficient. Conclusions : A series of recommendations were identified to improve the coupling between forage plant biomass data collected in the field and variables extracted from satellite images, and thus improve model performance.