Spatialization of rice crop yield using Sentinel-1 SAR and Oryza Crop Growth Simulation Model

Rice is a staple food across the majority of the world’s population that is expected to exceed 9 billion by 2050 and will require approximately 60% more food. In season and accurate information on the spatiotemporal distribution of rice cultivation, phenology across the region and spatial distribution of yield is of significant importance. This information is used by various stakeholders such as government, policymakers, insurance companies, and agri-input companies. Methods involving manual surveys for developing spatial crop yield are constrained by short harvest window and availability of the skilled human resource. Estimation of regional crop yield with precision and accuracy requires the use of high-resolution remote sensing data. The key contribution of this study is the spatial estimation of rice yield by assimilation of parameters derived from Synthetic Aperture RADAR (SAR) data from Sentinel-1 satellite into a process-based Oryza crop growth simulation model. The study has been carried out in four districts of coastal Andhra Pradesh, India viz., Guntur, Krishna, East Godavari and West Godavari during monsoon season locally called Kharif (mid-Jun. to midDec.) 2018. In the study area, rice is transplanted during mid-Jun to Aug. end and harvested from Oct. to mid-Dec. months. The methodology for in-season regional rice area estimation using random forest classifier has been described in our previous work. This study provides insights into the estimation of rice crop phenology and Leaf Area Index (LAI) using early time series of Sentinel-1 SAR observations. The rice phenology parameter such as Start of the Season (SoS) is estimated using Sentinel-1 SAR time series available during Jun.-Sept. 2018. The pixel-wise SoS estimation method comprises finding the local minima from the time series and image compositing. Total of six different SoS estimates is considered to cover early and late transplanted areas. The equation presented in literature has been used to estimate LAI from VH backscatter. Further, to facilitate the compute-intensive crop growth simulation task and cover maximum variation, the estimated LAI was categorized into five classes. Other datasets required for crop growth simulation such as weather was obtained from NOAA. A lookup table based approach was used wherein yield simulations were generated considering five SoS classes, five LAI classes, and four weather combinations. The total of 120 yield simulations were finally mapped to each pixel’s SoS, LAI, and weather categories. The plot-wise crop yield data for fifty-two (52) plots was collected for independent validation of yield estimates. The comparison of simulated and actual yield showed Normalized Root Mean Squared Value (NRMSE) of 9.21%. The overall agreement between actual and simulated yield is 83-89%. The results showed that spatialization of crop growth simulation for yield estimation using remote sensing observations provides fairly accurate yield estimates. Also, it is observed that the look-up table based approach reduced the computational complexity and crop growth model simulation time.