Improving Unmanned Aerial Vehicle (UAV) remote sensing of rice plant potassium accumulation by fusing spectral and textural information

Abstract

Plant potassium accumulation (PKA) plays an important role in evaluating the production capacity of crops. Remote sensing from unmanned aerial vehicles (UAVs) is increasingly used in precision agriculture, but relatively little is known about its use for determining the potassium (K) nutritional status of rice (Oryza sativa L.). This study compares the performance of different sets of spectral and textural indices derived from UAV data for estimating rice K nutritional status. A UAV equipped with three cameras (RGB, color-infrared (CIR) and multispectral cameras) was used to acquire imagery of rice canopies at different key growth stages. Immediately following the overflights, rice canopies were sampled for rice PKA. Regression models were then built to predict rice PKA with spectral and textural indices as predictor variables. Finally, stepwise multiple linear regression (SMLR) was used to determine if fusing spectral and textural indices significantly improved UAV-based PKA estimates. The renormalized difference vegetation index (RDVI) calculated from multispectral imagery proved to be the best predictor of rice PKA (R2 = 0.72, RMSE = 5.42 g m−2), though spectral vegetation indices calculated from RGB and CIR imagery also showed moderately strong prediction capability (R2 ≤ 0.56, RMSE ≥ 7.33 g m−2). Among the textural indices, the renormalized difference texture index [RDTI (MEA800, MEA680)] calculated by mean texture (MEA) of the multispectral imagery was the best predictor of rice PKA (R2 = 0.74, RMSE = 5.57 g m−2) while textural indices calculated from the RGB and CIR imagery showed only weak relationships with rice PKA (R2 ≤ 0.40, RMSE ≥ 8.11 g m−2). Fusing textural and spectral vegetation indices improved our ability to remotely sense PKA, with the SMLR model combination of RDTI (MEA800, MEA680) and spectral vegetation index DATT performing best (R2 increased by 11.11% to 0.80 and RMSE decreased by 7.5% to 5.15 g m−2). Our findings suggest that spectral and textural indices derived from UAV data allow for accurate mapping of PKA and that fusion of the two further improves the accuracy.