UAV-Based Estimation of Grain Yield for Plant Breeding: Applied Strategies for Optimizing the Use of Sensors, Vegetation Indices, Growth Stages, and Machine Learning Algorithms

Abstract

Non-destructive in-season grain yield (GY) prediction would strongly facilitate the selection process in plant breeding but remains challenging for phenologically and morphologically diverse germplasm, notably under high-yielding conditions. In recent years, the application of drones (UAV) for spectral sensing has been established, but data acquisition and data processing have to be further improved with respect to efficiency and reliability. Therefore, this study evaluates the selection of measurement dates, sensors, and spectral parameters, as well as machine learning algorithms. Multispectral and RGB data were collected during all major growth stages in winter wheat trials and tested for GY prediction using six machine-learning algorithms. Trials were conducted in 2020 and 2021 in two locations in the southeast and eastern areas of Germany. In most cases, the milk ripeness stage was the most reliable growth stage for GY prediction from individual measurement dates, but the maximum prediction accuracies differed substantially between drought-affected trials in 2020 (R2 = 0.81 and R2 = 0.68 in both locations, respectively), and the wetter, pathogen-affected conditions in 2021 (R2 = 0.30 and R2 = 0.29). The combination of data from multiple dates improved the prediction (maximum R2 = 0.85, 0.81, 0.61, and 0.44 in the four-year*location combinations, respectively). Among the spectral parameters under investigation, the best RGB-based indices achieved similar predictions as the best multispectral indices, while the differences between algorithms were comparably small. However, support vector machine, together with random forest and gradient boosting machine, performed better than partial least squares, ridge, and multiple linear regression. The results indicate useful GY predictions in sparser canopies, whereas further improvements are required in dense canopies with counteracting effects of pathogens. Efforts for multiple measurements were more rewarding than enhanced spectral information (multispectral versus RGB).

The Transferability of Spectral Grain Yield Prediction in Wheat Breeding across Years and Trial Locations

“…Considering the limited studies on GY modelling for predicting genotypic differences in GY in different locations and years, this analysis aims to evaluate the dataset effect for GY prediction in winter wheat breeding yards based on data from three years and two locations. It extends the evaluation of within-trials models from two locations and two years [19]. This study recommended a red edge index amongst different RGB and multispectral indices and measurements around anthesis and early milk ripeness.…”

“…The methodology is based on that described in detail in Prey et al (2022) [19]. Based on that study [19], which identified the normalized difference red edge index (NDRE1) as overall best vegetation index for GY estimation among several tested multispectral and RGB bands and vegetation indices, this study focuses on this spectral index. The index was calculated as…”

“…The data was analyzed in R [22] using an automated processing pipeline based on the caret package [23]. The partial least squares regression (PLSR) was chosen based on its overall good performance and its calculation efficiency in the previous analysis within trials [19]. The NDRE1 input data was centered to mean 0 and scaled to standard deviation 1.…”

“…In the case of across-trials models, all plots of the test trials were used as a test set. The WYWL models were reported in [19] for the trials in 2020 and 2021 and evaluated in the same way for the MR_22 and HZ_22 trials, based on an independent test set using 20% of the data based on random data splitting. For modelling across trials, dates were matched between trials based on the least temporal differences as calculated from the day of the year (DOY), assuming that the measurement dates, i.e., for the training and the test data should be the most similar.…”

“…Since models need to achieve sufficient prediction accuracies within particular trials, testing is only conducted within trials. Moreover, based on the results in the analysis from Prey et al (2022) [19], where GY predictions differed substantially between locations and years, we expect that the combination of data from multiple trials for model training will not directly improve GY predictions, since models would be fitted to the inter-trial variation as opposed to the intra-trial variation. We expect that the use of models trained on trials most similar to the trial used for application will be more expedient.…”