In-season mapping of rice yield potential at jointing stage using Sentinel-2 images integrated with high-precision UAS data.

Developing regionally applicable models for timely mapping of rice yield potential still faces challenges like the poor accuracy of yield prediction at early stages, and the scale mismatch between ground samples and satellite image pixels. Here, we explored the potential of using high-resolution multispectral images and texture features derived from unmanned aerial system (UAS) to fill these gaps. Rice field experiments were carried out in eastern China during 2017, 2018, and 2020 to acquire UAS (eBee-SQ) and satellite (Sentinel-2) images at jointing stage as well as the rice final yield. High spatial-resolution (HSR, 0.1 m) eBee-SQ images were used to extract both textures and vegetation indices (VIs), the resampled medium spatial-resolution (MSR, 20 m) eBee-SQ and Sentinel-2 images were only used to extract VIs. Firstly, using experimental data from 2017 and 2018, five kinds of yield prediction models at jointing stage were developed with: [ⅰ] VIs derived from MSR Sentinel-2 images and linear regression (LR) (least-RMSE = 1.68 t ha−1), [ⅰi] VIs derived from MSR eBee-SQ images and LR (least-RMSE = 1.482 t ha−1), [ⅱi] VIs derived from HSR eBee-SQ images and LR (least-RMSE = 1.371 t ha−1), [iv] VI-texture mixed indices derived from HSR eBee-SQ images and LR (least-RMSE = 1.082 t ha−1), and [v] random forest (RF) regression combining VIs and textures derived from HSR eBee-SQ images (least-RMSE = 0.997 t ha−1). These results showed that yield prediction models developed by both MSR and HSR eBee-SQ data performed better than by MSR Sentinel-2 data. Secondly, using optimal models derived from HSR eBee-SQ data and in-season HSR eBee-SQ data, different farm-scale HSR yield potential maps were generated. These yield maps were processed to Sentinel-2 resolution (20 m) to relate with Sentinel-2 VI for building integrated yield prediction models (IPMs) based on linear and quadratic regression. IPMs were then used to map town-scale rice yield potential by importing Sentinel-2 VI from jointing stage, and were compared with simple yield prediction model (SPM) that directly used Sentinel-2 VI. The independent validation results using yield data collected in 2020 suggested that all of the IPMs were more accurate (RMSE = 0.64 – 0.74 t ha−1) than the SPM (RMSE = 0.96 t ha−1). Besides, IPMs involving eBee-SQ derived textures and VIs showed higher accuracy in mapping yield potential than IPM involving eBee-SQ derived VI only. Overall, this study provides a novel approach to accurately mapping the regional rice yield potential through the integration of satellite and UAS data. • Combining VI and texture improved yield prediction accuracy at rice jointing stage. • Textures from red-edge band were more conducive to yield prediction in RF model. • UAS-predicted yield map was upscaled and correlated with satellite VI to build IPM. • IPM improved accuracy for mapping town-level yield potential compared with SPM. • UAS served as an intermediate source to fill gaps between ground and satellite data. [ABSTRACT FROM AUTHOR]