Estimation of sowing dates at the village level: a probabilistic approach based on vegetation onset detections.

The length of the crop season is a major determinant of yields in Sahel. Higher variability being observed on when the season starts rather than when it finishes, accurate estimates of sowing dates are vital for yield forecasting and consequently food-security. Based on green-up onset detections derived from Moderate Resolution Imaging Spectroradiometer (MODIS) time series at 250m, this paper proposes a new methodology for their estimation. It builds upon a stochastic model that translates vegetation onsets detected around villages in sowing probabilities. Results for Niger show that it outperforms the standard method adopted in the region based on rainfall thresholds. Vegetation onsets, defined as the transition from a bare to a vegetated surface, are detected by the method proposed by Pekel et al. (2010). The selection of this method is threefold motivated: (i) it exploits the multi-spectral information and consequently avoids usual confusions between bare soils and vegetation, (ii) it synthetizes multi-spectral information in one value, and (iii) it reduces the noise due to the observation conditions. The proportion of vegetated pixels around each village is then computed in a buffer of 5km and converted into a probability of sowing. Adopting a parsimonious position, only two parameters are estimated in the conversion process. The number of villages having sowed by decade and by department is assumed to follow a Poisson-Binomial distribution. The parameters are then obtained by maximizing the likelihood of observing the figures officially stated by the Nigerien Ministry of Agriculture for the years 2008 and 2009. Cross validation between years shows that despite the apparent absolute difference on the estimated parameters, the method generates significantly higher R-squares and lower root mean square errors (RMSE) when compared with estimations based on rainfall thresholds (between 8% to 18% higher R-squares and 17% to 29% lower RMSE). This improvement can be explained, among others, by the higher spatial resolution of the input data (250m for the vegetation onsets against 10km for rainfall estimates), which is more consistent with the higher weather spatial variability in the area.

DE ARAUJO MARINHO FILHO Eduardo;  VANCUTSEM Christelle;  FASBENDER Dominique;  KAYITAKIRE Francois;  PINI Giancarlo;  PEKEL Jean-François;  MARINHO Eduardo; 

2012-12-20

IEEE-INST ELECTRICAL ELECTRONICS ENGINEERS INC

JRC72437

1939-1404,   

https://publications.jrc.ec.europa.eu/repository/handle/JRC72437,   

10.1109/Agro-Geoinformatics.2012.6311695,