Winter wheat SPAD estimation from UAV hyperspectral data using cluster-regression methods

Abstract

Soil plant analysis development (SPAD) values indicate the relative chlorophyll content in leaves. Chlorophyll plays a vital role in wheat growth and fertilization management as a photosynthetic agent. Unmanned aerial vehicle (UAV) hyperspectral data, combined with measured SPAD values in fields, are widely used to study wheat chlorophyll concentrations over time. In this study, considering the spectral differences in fields with different soil fertility, we propose a new modeling method named as cluster-regression to estimate the winter wheat SPAD at a national soil observation and experimental station in northern Anhui, China. The wheat spectrum was divided into several clusters according to the spectral angular distance, and regression models were built between wheat spectrum and the measured SPAD for each cluster. We used K-means as the clustering method and two types of ensemble learning algorithms, namely, random forests and Extreme Gradient Boosting (XGBoost) as the regression methods. The root-mean-square error (RMSE), median absolute percentage error (MAPE), and R2 were used as model accuracy indicators. The results showed that the XGBoost model slightly outperformed the random forest in wheat SPAD estimation. Coupled with the first step of K-means clustering, the overall performance of the regression models can be easily improved, and cluster-XGBoost showed the best performance, with an RMSE of 1.444, MAPE of 2.36%, and R2 of 0.925. Moreover, addition of soil organic matter and soil total nitrogen positively impacted the accuracy of SPAD estimation models. This study provides an effective UAV hyperspectral technique for wheat SPAD concentration estimation and demonstrates the impact of soil organic matter and total nitrogen on wheat SPAD estimation.