Can field-scale yield be reconstructed from commune-scale yield using Sentinel-1?

The long-term yield estimation, which leverages and learns dependency relationships between dense-in-time and dense-in-spatial Earth Observation (EO) products, is a popular topic in agriculture, but all based on commune-scale statistical information. However, there is a need for yield information at a field-scale resolution, which is a key descriptor for agricultural practices and which is difficult to achieve with current yield estimation models. In this study, the three-year nationwide field-scale Sentinel-1 and Landsat-based vegetation index measurements will be explored to see if commune-scale yield information can be projected to the field-scale using the time series measurements. To discuss this problem, firstly three-years cotton crop mask covering Turkey is produced based on only Sentinel-1 and Sentinel-2 images. Then, all field-scale time series measurements are clustered by dynamic time warping approach without considering any prior information about the fields. Finally, in order to understand if there is any relationship with these cluster centers and the cotton field yield condition, the commune scale statistical information is projected to the field scale according to the cotton mask. The Sentinel-1 VV backscattering shows a higher sensitivity to the yield variation compared to the vegetation index one, showing more sensitivity to geographical distribution but not as much as climate variables do. Relationship between yield and climate variables remains constant at different spatial scale, instead, imaging data has the potential to leverage field-scale yield from the statistical commune scale yield information.