Estimating the canopy chlorophyll content of winter wheat under nitrogen deficiency and powdery mildew stress using machine learning

As an important indicator of the photosynthetic capacity of crops, the canopy chlorophyll content (CCC) is nondestructively estimated by reflectance using various spectrometers. Crop growth is often severely affected by Nitrogen (N) deficiency and diseases, and the compatibility of the data collected for different stressors needs further clarification to develop unified estimation models. In this field experimental study, hyperspectral data of the wheat canopy were collected, along with canopy chlorophyll content, to assess nitrogen deficiency and powdery mildew stress. Comparative analysis of hyperspectral remote sensing data input features (original reflectance (OR), spectral index (SI) and wavelet features (WF)) was conducted. A combination of feature selection and machine learning was used to determine the best estimation mode for the accurate inversion of CCC under these two stressors. The results showed that the canopy spectra under nitrogen deficiency and powdery mildew stress had the same change trend. Under nitrogen deficiency, the sensitive wavelengths to CCC mainly reflected canopy structure characteristics, followed by pigments. Under powdery mildew stress, the sensitive wavelengths mainly reflected pigment characteristics, followed by canopy structure. Eight input features (two reflectance wavelengths, two spectral indices and four wavelet features) were selected using competitive adaptive reweighted sampling (CARS) and variance inflation factor (VIF) methods. Machine learning (ML) produced better estimates for both stressors. For CCC estimation under nitrogen stress, random forest regression (RFR) was more suitable (R2 = 0.828; RMSE = 0.363 g/m2) and showed a higher accuracy for both the calibration and validation sets. For CCC estimation under powdery mildew stress, support vector machine regression (SVR) was more suitable (R2 = 0.787; RMSE = 0.126 g/m2), especially when OR and WF data were used as input features. For the unified estimation of CCC under both stressors, WF is most effective as an input feature and good accuracy is achieved for both SVR (R2 = 0.846; RMSE = 0.296 g/m2) and RFR (R2 = 0.844; RMSE = 0.297 g/m2), and their differences are very small. The results demonstrated that the CWT-CARS-VIF-ML mode was appropriate for CCC estimation under two different stressors, which provides an ideal reference and technical guidance for the evaluation of photosynthetic potential and improved crop management.