Within-season forecasting of crop yields is of great economic, geo-strategic and humanitarian interest. Satellite Earth Observation now constitutes a valuable and innovative way to provide spatio-temporal information to assist such yield forecasts. This study explores different configurations of remote sensing time series to estimate of winter wheat yield using either spatially finer but temporally sparser time series (5-daily at 100 m spatial resolution) or spatially coarser but denser (300 m and 1 km at daily frequency) time series. Furthermore, we hypothesised that better yield estimations could be made using thermal time, which is closer to the crop physiological development. Time series of NDVI from the PROBA-V instrument, which has delivered images for retrieving vegetation and land surface characteristics at a spatial resolution of 100 m, 300 m and 1 km since 2013, were extracted for 39 fields for field level analysis and 56 fields for regional level analysis across Northern France during the growing season 2014-2015. An asymmetric double sigmoid model was fitted on the NDVI series of the central pixel of the field. The fitted model was subsequently integrated either over thermal time or over calendar time, using different baseline NDVI thresholds to mark the start and end of the cropping season. These integrated values were used as a predictor for yield using a simple linear regression and yield observations at field level. The dependency of this relationship on the spatial pixel purity was analysed for the 100 m, 300 m and 1 km spatial resolution. At field level, depending on the spatial resolution and the NDVI threshold, the adjusted R2 ranged from 0.20 to 0.74; jackknifed – leave-one-field-out cross validation – Root Mean Squared Error (RMSE) ranged from 0.6 to 1.07 t/ha and Mean Absolute Error (MAE) ranged between 0.46 and 0.90 t/ha for thermal time analysis. The best results for yield estimation (adjusted R2 = 0.74, RMSE = 0.6 t/ha and MAE = 0.46 t/ha) were obtained from the integration over thermal time of 100 m pixel resolution using a baseline NDVI threshold of 0.2 and without any selection based on pixel purity. The field scale yield estimation was aggregated to the regional scale using 56 fields. At the regional level, there was a difference of 0.0012 t/ha between thermal and calendar time for average yield estimations. The standard error of mean results showed that the error was larger for a higher spatial resolution with no pixel purity and smaller when purity increased. These results suggest that, for winter wheat, a finer spatial resolution rather than a higher revisit frequency and an increasing pixel purity enable more accurate yield estimations when integrated over thermal time at the field scale and at the regional scale only if higher pixel purity levels are considered.

DURGUN Ozum;  GOBIN A.;  DUVEILLER BOGDAN Grégory Henry E;  TYCHON Bernard; 

2019-12-09

ELSEVIER SCIENCE BV

JRC117210

0303-2434 (online),   

https://www.sciencedirect.com/science/article/pii/S0303243419306439?via%3Dihub,    https://publications.jrc.ec.europa.eu/repository/handle/JRC117210,   

10.1016/j.jag.2019.101988 (online),