UAV-based hyperspectral and ensemble machine learning for predicting yield in winter wheat

Authors: LI, ZONGPENG - Chinese Academy Of Agricultural Sciences; CHEN, ZHEN - Chinese Academy Of Agricultural Sciences; CHENG, QIAN - Chinese Academy Of Agricultural Sciences; DUAN, FUYI - Chinese Academy Of Agricultural Sciences; Sui, Ruixiu; HUANG, XIUQIAO - Chinese Academy Of Agricultural Sciences; XU, HONGGANG - Chinese Academy Of Agricultural Sciences

Submitted to: Agronomy
Publication Type: Peer Reviewed Journal
Publication Acceptance Date: 1/10/2022
Publication Date: 1/14/2022
Citation: Li, Z., Chen, Z., Cheng, Q., Duan, F., Sui, R., Huang, X., Xu, H. 2022. UAV-based hyperspectral and ensemble machine learning for predicting yield in winter wheat. Agronomy. 12(1):202. https://doi.org/10.3390/agronomy12010202.
DOI: https://doi.org/10.3390/agronomy12010202

Interpretive Summary: Crop yield prediction in a large scale can provide a scientific foundation for establishing crop production plans and ensuring food security. Timely and accurate estimates of wheat yield is crucial for rapid decision-making in wheat crop management. In a collaborative research, ARS researcher in Stoneville, MS and the researchers in Chinese Academy of Agricultural Sciences used an unmanned aerial vehicle (UAV) to acquire hyperspectral imagery of wheat crops in China to predict the crop yield. They extracted hyperspectral indices from the images to develop different models for the yield prediction. It was found that the decision-level fusion (DLF) model gave the highest accuracy in the yield prediction. This study demonstrated the effectiveness of using hyperspectral images to build models for estimating crop yield. The results provided useful information for applications of UAV-based remote sensing in agricultural production.

Technical Abstract: Winter wheat is a widely grown and indispensable food crop worldwide, and timely and accurate estimates of wheat yield play a vital role in accurate management and rapid decision-making in agriculture. In this study, a low-altitude unmanned aerial vehicle (UAV) was used to obtain hyperspectral image data of the winter wheat crop canopy at two growth stages, flowering and grain filling, to predict the crop yield. Field trials with 30 winter wheat cultivars and three irrigation treatments were used. A large number of hyperspectral indices were extracted from the high spatial resolution hyperspectral imagery, and three feature selection methods, Recursive Feature Elimination (RFE), Boruta and Pearson Correlation Coefficient (PCC), were used to filter the hyperspectral indices to reduce the data dimensionality. A decision-level fusion (DLF) model in ensemble learning was then developed by combining four underlying learners: support vector machine (SVM), Gaussian process (GP), linear ridge regression (LRR), and random forests (RF). The results showed that the SVM model had the best performance in predicting the yield under the preferred characteristics of the two growth stages with a R2 in the range of 0.62-0.73, while the GP, LRR, and RF model had an R2 in the range of 0.57-0.72, 0.62-0.66, and 0.58-0.68, respectively. In the modeling, the preferred characteristics obtained by the RFE method at flowering stage performed better while the preferred characteristics obtained by the Boruta method at the grain filling stage had a better performance. The results demonstrated the validity of the proposed DLF model, which has a maximum R2 of 0.66 for the RFE and Boruta methods and a maximum R2 of 0.78 for the Boruta method.