Yield forecasting for winter wheat using time series NDVI from HJ satellite

Abstract: A NDVI time series curve, proposed from high time resolution remote sensing images, contains rich information for crop yield forecasting. MODIS-NDVI and AVHRR-NDVI are normally used. However, the spatial resolution of MODIS and AVHRR are low. When they used for crop yield forecasting in China, the prediction accuracy will be reduced by a mixed pixel problem, as the farmland is small. China launched a HJ satellite constellation in 2009. The satellite constellation can provide an image with a time resolution of 2 days and a spatial resolution of 30 m. It would be helpful to make crop yield forecasting in field scale in China, based on the NDVI time series curve designed from higher spatial resolution HJ images. Taking Yucheng city as a research area, this study research the feasibility of proposing a NDVI time series curve using HJ satellite images and then making a winter wheat yield forecast using parameters extracted from the above curve. For this purpose, 11 images were acquired with a nearly 10 day interval from February 19th to June 5th, and the winter wheat yields of 12 sample sites were also measured during a field campaign in 2012. Firstly, quintic polynomial least square fitting was used to propose a NDVI time series curve using NDVI values extracted from the above images. The curve covered reviving, jointing, flowering, filling and mature stages of winter wheat. Secondly, parameters of a NDVI time series curve were calculated. They were maximum NDVI and accumulated NDVI during winter wheat growth stage, NDVI value at winter wheat reviving stage, NDVI change rate during winter wheat vegetation growth stage, and NDVI change rate during the winter wheat reproduction growth stage. Thirdly, using collected sample data, the yield prediction models were created, based on the above parameters respectively. Meanwhile, a single-phase image NDVI was also used to propose a yield prediction model, and it was compared to the above models, in order to show if a NDVI time series curve can provide more information than a single-phase image for yield prediction. At last, the multivariate yield prediction model was proposed based on the parameters of a NDVI time series curve using a stepwise multiple linear regression method, and was validated by a leave-one-cross-validation method. The results showed that the winter wheat NDVI series curve can be proposed by a HJ satellite image; A NDVI time series curve can provide more information for winter wheat yield prediction than a single phase image; The parameters extracted from a NDVI time series curve can be used to make a reliable winter wheat yield prediction model, and the designed multivariate yield prediction model has the determination coefficient (R2) value of 0.87 and relative error (RE) value of 5.02% at model calibration, and R2 value of 0.78 and RE value of 6.87% at model validation.