Leaf Chlorophyll Content Retrieval from Hyperspectral Measurement Using Convolutional Neural Network and Continuous Wavelet Transform Techniques

Dynamic monitoring chlorophyll variations is imperative for overseeing crop growth dynamics and biophysical exchanges. Previous studies have established methods for leaf chlorophyll content assessment using hyperspectral reflectance. Compared with traditional remotely sensing methods, convolutional neural network (CNN) excels in representation learning ability, while continuous wavelet transform (CWT) has the capacity to unveil nuanced spectral changes. Nevertheless, integrative studies employing both CWT characteristics and CNNs within the context of leaf chlorophyll content (LCC) assessment were rarely reported. Aiming at this problem, the present study harnessed abundant leaf data, namely ANGERS, NEON, and winter wheat, to investigate the capability of synthesizing AlexNet and CWT features for LCC assessment. Results revealed that AlexNet model informed by CWT features, at 21 decomposition scale, best captured LCC variations, achieving an R² of 0.98, an RMSE of 3.52 μg/cm², an RRMSE of 11.00 %, and an RPD of 6.41. Further validation confirmed an increment in R² by 0.19, a reduction in RMSE by 2.83 μg/cm², and a 7.70 % reduction in RRMSE for LCC estimates when compared with results using raw spectral reflectance. These findings corroborated the substantial potential CWT holds in conjunction with AlexNet for robust LCC estimation, advocating for increased integration of these methodologies in remote sensing assessments of crop physiological parameters.