UAV time-series imagery with novel machine learning to estimate heading dates of rice accessions for breeding

Abstract

Timely monitoring of rice heading dates is essential for estimating growth status and grain yield in rice breeding. This study aims to develop a machine learning method to classify growth stages and estimate the days from sowing to initial heading (IHD) and days from sowing to full heading (FHD) of rice accessions from unmanned aerial vehicle (UAV)-based imagery. A two-year field experiment was conducted on different rice accessions, including doubled-haploid (DH) lines, early maturing F1 hybrids, and cultivars at two different locations (i.e., EXP 1 and EXP 2). A neural network, namely Res2Net50, was first used to classify the rice growth stages based on UAV RGB images. The partial least squares regression (PLSR) model was then developed to predict IHD and FHD by using canopy reflectance extracted from multispectral images. The results showed that the Res2Net50 model successfully identified five growth stages with overall accuracies of 86% and 83% in EXP 1 and EXP 2, respectively. PLSR models developed with image data after the Res2Net50 classification achieved good performances on predicting IHD and FHD with the coefficient of determination (R2) of 0.60–0.78, root mean square error (RMSE) ranging from 3.9 days to 5.1 days, and relative root mean square error (rRMSE) varied between 4% to 6%. Model validation with ten cultivars further demonstrated the capability of proposed method for IHD and FHD predictions and achieved the rRMSE of 3%-4%, which was comparable with that obtained by field inspection. Furthermore, the feasibility of the proposed method in monitoring of the rice heading dynamics was validated by EXP 2 using the model transfer. The proposed approach enables quantification of heading dates of diverse rice accessions in breeding and could be used as a high-throughput screening technique for accelerating rice breeding.