| Chlorophyll is one of the most important pigments in plants.Rapid,accurate,nondestructive and real-time quantitative acquisition of chlorophyll content in each part of plants is of great significance for understanding plant growth status and guiding scientific fertilization.In recent years,the three-dimensional(3D)distribution model construction of chlorophyll for plants based on hyperspectral LiDAR has made some progress,however,it is only aimed at the adaxial of the leaf,which is limited to a single species.This thesis uses hyperspectral LiDAR to collect hyperspectral-spatial point cloud data of different types of leaves and plants,uses empirical statistical method and leaf optical model respectively to complete the mapping from hyperspectral to chlorophyll content,and finally combines spatial coordinates to construct 3D chlorophyll distribution model for plants,the main works are as follows:(1)Collected hyperspectral point cloud data of different plant samples and designed a 3D chlorophyll modeling method based on the classification of the adaxial and abaxial sides of leaves.Firstly,the chlorophyll content prediction model of leaf adaxial and abaxial was constructed by partial least squares regression(PLSR).Secondly,an adaptive threshold selection strategy based on the feature difference index combined with ten-fold cross-validation was proposed to achieve the classification of adaxial and abaxial leaves.Finally,the chlorophyll content was calculated according to the category label selection mode.The experimental results show that the 3D distribution of chlorophyll on plants obtained using the proposed method is more consistent with the real situation,the coefficient of determination is improved from less than 0 to 0.69,and the root mean square error is reduced from 14.45 to 4.97.(2)The leaf optical model PROSPECT-5 was used to estimate chlorophyll content at the leaf scale.Firstly,the performance of look-up table algorithms at different table sizes and iterative optimization algorithms are compared.Secondly,the PLSR(Partial Least Square Regression),ANN(Artificial Neural Network)and RFR(Random Forest Regression)algorithms are combined with the PROSPECT-5 model to avoid the model training process falling into an overfitting state by adding additive Gaussian noise to the training data,and to improve the model generalization ability and robustness.Using the prediction results of the public dataset ANGERS’03 as a criterion,the ANN algorithm was the best compared to the other algorithms,achieving a coefficient of determination of 0.90 and a root mean square error of 4.51 for the prediction model trained on the noisy dataset,an increase of 0.13 and a decrease of 1.82 compared to the noisy dataset.(3)Using the PROSPECT-5 model to invert the chlorophyll content at the 3D scale and combining the spatial coordinates,the 3D chlorophyll distribution model was constructed.To address the distance-incidence angle effect of hyperspectral LiDAR at the 3D scale,two adaptive vegetation indices based on sensitivity analysis were established,and the form of the value function in the look-up table algorithm was improved using the two indices.Hyperspectral-spatial point cloud data of several water spinach samples were collected using hyperspectral LiDAR,and the performance of the look-up table algorithm before and after the improvement was compared.The experimental results show that the improved look-up table algorithm is better than the unimproved in predicting the leaf edge parts and parts with larger incidence angles of the vegetable samples,and the 3D chlorophyll distribution model is more consistent with the real growth conditions of the plants.Figure [32] Table [11] Reference [67]... |
PDF Full Text Request