PROSPECT combined with deep transfer learning for citrus chlorophyll estimation from multiscale reflectance

Large-scale citrus orchard cultivation is an inevitable trend in the development of precision agriculture, but single-scale hyperspectral data can hardly satisfy the monitoring of chlorophyll parameters in large-scale orchards. To effectively exploit the advantages of remote sensing data at multiple scales (leaf, canopy, and unmanned aerial vehicle (UAV)), this study employed Murcott citrus trees as the research subject .This study conducted a comparative analysis of response characteristics and correlations between citrus leaf hyperspectral (measured and simulated) and chlorophyll content, applied the CARS algorithm to PROSPECT-simulated spectra for selecting hyperspectral feature bands associated with citrus chlorophyll content, and established a DNN model to predict citrus leaf chlorophyll content. This study used deep transfer learning to explore the performance of measured data for estimating citrus chlorophyll content,and the multiscale portability of the model was explored via citrus canopy hyperspectral data and UAV hyperspectral data. The results show that: (1) The 11 feature bands(526, 545, 553, 572, 706, 708, 722, 760, 764, 860 and 869 nm) selected in 400-1000 nm could resist the influence of leaf chlorophyll content (LCC) changes and had good specificity for citrus LCC. (2) The LCC prediction model was constructed by combining the PROSPECT model with the DNN model. Transfer learning was implemented to refine the pre-trained model through partial of ground observed spectral data. Comparative validation demonstrated that the accuracy of the migration learning model improves significantly at different fine-tuning ratios relative to the pre-trained model.(3) The PROSPECT model combined with DNN transfer model studied in this paper has good versatility, accuracy and portability in predicting citrus LCC at multiple scales. With the gradual reduction in spatial resolution, the accuracy gradually decreased, and the R 2 distributions were 0.890, 0.752 and 0.685 for the leaf, canopy and UAV scales, with RMSEs of 3.747, 6.704 and 8.988, respectively. • Selected wavelengths are resistant to chlorophyll content variations and have good specificity for citrus chlorophyll content. • PROSPECT model combined with DNN model to achieve multi-scale chlorophyll content prediction in citrus.