Effects of phenological uniformity on prediction accuracy of wheat yield and biomass

Abstract

Context

Accurate prediction of wheat yield and biomass is essential for breeding new cultivars and optimizing field management. The accuracy of yield and biomass predictions can be affected by the phenological phase of data collection. However, phenological transitions are gradual, and wheat fields rarely consist of a single phenological stage. The influence of phenological uniformity (PU) on prediction accuracy has been largely overlooked, particularly in multi-cultivar study areas.

Objective

This study aimed to quantitatively evaluate PU, to classify wheat phenological stages using hyperspectral data collected by an unmanned aerial vehicle (UAV), to identify the optimal growth stage for yield and biomass prediction, and to assess the impact of PU on prediction accuracy.

Methods

A two-year field experiment was conducted using 210 wheat cultivars with diverse phenological stages, and time-series hyperspectral images were collected using an UAV. The study first defined and quantitatively evaluated PU. Subsequently, the classification accuracies of five models were compared to identify the most effective approach for phenological stage classification. Hyperspectral data collected at four key growth stages were then used to determine the optimal stage for yield and biomass prediction. Finally, datasets with varying PU were constructed to predict yield and biomass, and the influence of PU on prediction accuracy was assessed.

Results

PU of wheat exhibited a fluctuating trend throughout the growth stages, with most values ranging between 0.5 and 0.8. Hyperspectral data enabled effective discrimination of key phenological stages, among which the end-to-end mixhop superpixel-based graph convolutional networks (EMS-GCN) model achieved the highest classification accuracy, with an overall accuracy of 86.2 %. The PLSR model achieved the most accurate predictions of both yield (R² = 0.692, RMSE = 1.091 t/ha, CV = 0.152) and biomass (R² = 0.827, RMSE = 1.873 t/ha, CV = 0.113) at the flowering stage. The results of yield and biomass prediction based on datasets with varying PU values indicated a positive correlation between PU and prediction accuracy.

Conclusions

Accurate classification of key wheat phenological stages can be achieved by combining deep learning with hyperspectral data. The flowering stage is the optimal period for yield and biomass prediction. PU positively correlates with the prediction accuracy of yield and biomass.

Implications

This study emphasizes the important role of PU in wheat yield and biomass prediction, and accurate monitoring of PU can provide theoretical guidance for data collection. This is of great significance to the development of precision agriculture and guiding field management.