Assessment Of Plant Growth Traits Based On Proximal Reflectance Spectral Imaging

Assessment of plant growth traits is of great significance for monitoring plant growth,assessing yield potential,and exploring the process of plant-environment interactions and energy cycle at the field scale.With the rapid development of sensing technologies in the past 30 years,proximal reflectance spectral imaging has become the primary method for efficient perception of plant information to assess plant growth traits.However,assessment of plant growth traits is still limited by lacks of the full understanding of model mechanism,model accuracy and generalization,and the difficulty in upscaling conversion from leaf to canopy.Therefore,this study aims to develop efficient and accurate approaches for plant growth assessment,with the main focuses on the interaction mechanism between light and leaves,assessment of leaf biochemical traits,upscaling conversion from plant leaf to canopy,quantification of canopy traits,and yield-related performance traits.Based on the radiative transfer theory,this study proposed a new method to eliminate the difference between leaf bidirectional reflectance factor(BRF)and direction-hemispherical reflectance factor(DHRF)spectra.By using proximal reflectance spectral images within the regions of visible,near infrared,and shortwave infrared(VIS-NIR-SWIR),we developed models for assessing leaf biochemical,canopy structural,and yield-related performance traits of rice(Oryza sativa L.)and oilseed rape(Brassica napus L.).Model validations were also performed with different plant species.The proposed approaches provide new ways for accurate monitoring of crop growth and technical support for field precision management and crop breeding.The main contents and conclusions are summarized as follows:(1)Due to the lack of the full understanding of commonly used BRF spectra and the difficulty in the practical application of mechanistic models based on DHRF spectra,this study explored the difference between leaf BRF and DHRF spectra across different species and their effects on PROSPECT model inversion.A framework for the retrievals of leaf biochemical traits was constructed based on the radiative transfer model with leaf BRF spectra.The results showed that the difference between DHRF and BRF spectra varied with wavelengths and plant species,which was especially significant in the SWIR region.Spectral derivatives were able to eliminate the wavelength-independent difference between BRF and DHRF spectra,while implementation of Manhattan distance compensated the limitation of spectral derivatives to further reduce the wavelength-dependent difference.Thus,this study developed a novel framework by coupling the PROSPECT model with spectral derivatives and similarity metrics,namely PROSDM,to assess leaf biochemical traits.For assessment of the contents of leaf chlorophyll,carotenoid,water,and dry matter based on leaf BRF spectra,the root mean square errors(RMSEs)of PROSDM were 20.33%-44.19%lower than those of PROSPECT.The proposed approach provides a basic model framework with strong mechanism for multi-scale interpretation of reflectance image data.(2)In view of the complex relationships between leaf biochemical traits and lacks of accuracy and generalization of estimation models,the above radiative transfer theory was combined with transfer learning to characterize the diversities of leaf biochemical traits and reflectance spectra,and a new hybrid method with model update was proposed to estimate leaf nitrogen concentration(LNC).The results showed that LNC was closely associated with the contents of pigment and water,but their relationships varied with plant species and growth stages.The assessment of leaf biochemical traits was affected by differences in plant species and spectral regions,and the combination of VIS-NIR-SWIR reflectance was recommended to establish the transferable estimation model for LNC across different datasets.Compared with traditional partial least squares regression model,transfer component analysis can effectively extract similar spectral features from different datasets,and can be coupled with support machine regression to develop a transfer learning model,namely TCA-SVR,which improved transferable assessment accuracies of LNC by reducing the average RMSE by 36.76%.Furthermore,the model generalizability of TCA-SVR for LNC assessment was further improved by transferring only 5%of the samples from the target dataset to the source dataset.(3)In order to achieve the upscaling conversion from leaf to canopy,this study explored the effects of varied leaf and canopy structures on canopy reflectance spectra and interpretation of proximal reflectance spectral images.The results showed that different structures of leaf adaxial and abaxial surfaces of plant species(such as rice and oilseed rape)led to that leaf abaxial surface exerted the higher reflectance in the VIS and SWIR regions than leaf adaxial surface with the close reflectance in the NIR region.The difference in leaf adaxial and abaxial reflectance at different spectral regions caused the variations in canopy reflectance,and thus affected the estimations of leaf biochemical traits based on proximal reflectance spectral imaging.Furthermore,leaf structural differences further affected the response of canopy reflectance to the variations in canopy structural traits.The erectophile distribution produced the smallest canopy reflectance,and the variations in leaf angle distributions affected the assessment of leaf area index.The results provide a theoretical foundation for interpretation of proximal reflectance spectral images upscaling from leaf to canopy.(4)In order to eliminate the influence of leaf and canopy structural variations on canopy proximal reflectance spectral imaging,this study developed methods to improve the assessment of canopy biochemical and structural traits based on proximal reflectance spectral imaging with minimized insensitivity to leaf and canopy structural variations.The results showed that assessment of leaf biochemical traits at canopy scale was affected by leaf structural differences.This study proposed spectral indices upscaling from leaf to canopy to accurately estimate leaf biochemical traits with the relative RMSE of below 14%,which were insensitive to the differences of leaf adaxial and abaxial sides.The PROSAIL model and gap probability model were combined to simulate the effects of canopy structural variations on proximal reflectance spectral imaging,and the PROSAIL-GP model was constructed to estimate fractional vegetation cover in oilseed rape,rice,wheat,and cotton.Compared with traditional random forest regression model,the RMSEs for PROSAIL-GP model were reduced by 16.67%-30.77%.These findings provide novel methods for improving estimation accuracies of leaf biochemical and canopy structural traits from multi-scale remote sensing.(5)In order to accurately estimate the most complex and key yield-related performance traits(biomass and yield)of crops,based on the results mentioned above,this study revealed the relationships between leaf biochemical,canopy structural and yield-related performance traits.We constructed the estimation models for crop yield-related performance traits by combining temporal spectral and structural information,which improved the estimation accuracies of biomass and yield of rice and oilseed rape.The results showed that leaf chlorophyll content and leaf area index were closely related to crop yield-related performance traits,while their relationships significantly varied between vegetative and reproductive growth periods.The retrieved leaf biochemical and canopy structural traits from unmanned aerial vehicle(UAV)-based temporal spectral and structural information accurately assessed biomass of rice and oilseed rape with RMSEs of 0.22 kg/m~2 and 0.03 kg/m~2,respectively.The same good performances were also obtained for the assessment of yield of rice and oilseed rape with RMSEs of 0.39 t/ha and 0.30 t/ha,respectively.These methods on interpretation of UAV images and spectra provide technical support for field precision management and smart breeding.