Retrieving SPAD Values of Summer Maize Using UAV Hyperspectral Data Based on Multiple Machine Learning Algorithm

Abstract

Using unmanned aerial vehicle (UAV) hyperspectral images to accurately estimate the chlorophyll content of summer maize is of great significance for crop growth monitoring, fertilizer management, and the development of precision agriculture. Hyperspectral imaging data, analytical spectral devices (ASD) data, and SPAD values of summer maize in different key growth periods were obtained under the conditions of a micro-spray strip drip irrigation water supply. The hyperspectral data were preprocessed by spectral transformation methods. Then, several algorithms including Findpeaks (FD), competitive adaptive reweighted sampling (CARS), successive projections algorithm (SPA), and CARS_SPA were used to extract the sensitive characteristic bands related to SPAD values from the hyperspectral image data obtained by UAV. Subsequently, four machine learning regression models including partial least squares regression (PLSR), random forest (RF), extreme gradient boosting (XGBoost), and deep neural network (DNN) were used to establish SPAD value estimation models. The results showed that the correlation coefficient between the ASD and UAV hyperspectral data was greater than 0.96 indicating that UAV hyperspectral image data could be used to estimate maize growth information. The characteristic bands selected by different algorithms were slightly different. The CARS_SPA algorithm could effectively extract sensitive hyperspectral characteristics. This algorithm not only greatly reduced the number of hyperspectral characteristics but also improved the multiple collinearity problem. The low frequency information of SSR in spectral transformation could significantly improve the spectral estimation ability for SPAD values of summer maize. In the accuracy verification of PLSR, RF, XGBoost, and the DNN inversion model based on SSR and CARS_SPA, the determination coefficients (R2) were 0.81, 0.42, 0.65, and 0.82, respectively. The inversion accuracy based on the DNN model was better than the other models. Compared with high-frequency information, low-frequency information (DNN model based on SSR and CARS_SPA) had a strong estimating ability for SPAD values in summer maize canopy. This study provides a reference and technical support for the rapid non-destructive testing of summer maize.

Keywords

• UAV
• hyperspectral data
• successive projection algorithm
• XGBoost
• deep neural network (DNN)