Time-Resolved UAV Multispectral Characterization of Cultivar-Dependent Yield Variation in Rice Under Planting-Density Treatments

Abstract

UAV multispectral imagery can support rice yield estimation, but spectral–yield relationships may change with cultivar and planting density. We evaluated three japonica rice cultivars under five planting-density treatments in a split-plot experiment and collected UAV multispectral imagery on 13 dates. We examined five vegetation indices: NDVI, RVI, NDRE, OSAVI, and CLRE. Cultivar significantly affected grain yield and yield components. Planting density had no significant main effect, while the cultivar × density interaction was significant for grain yield. Spectral–yield correlations were strongest during 19–25 August, which broadly matched reproductive development based on field observations. NDRE and CLRE showed the most consistent associations. A four-index PLSR model using 19 August data achieved an outer-LOOCV R2 of 0.773 and an RMSE of 530.28 kg ha−1. Tuned RF gave comparable performance (R2 = 0.783, RMSE = 518.60 kg ha−1). Adjustment for cultivar, planting density, and block substantially reduced the pooled mid-August correlations. However, the four-index spectral set still improved prediction beyond cultivar, planting density, and their interaction. Grouped validation changed little when density levels or cultivar × density combinations were withheld, but performance declined sharply when an entire cultivar was excluded. These results show that mid-August UAV observations were informative in this experiment, but wider use will require validation across more cultivars, years, and sites.