| Leaf area index(LAI)is an important parameter for crop growth evaluation and yield estimation.Traditional measurement of LAI generally requires destructive sampling,which is time-consuming and labor-intensive,and is easily affected by environment factors.With the continuous development of remote sensing technology,remote sensors mounted on unmanned aerial vehicle(UAV)are widely used to obtain field crop information to estimate LAI,which greatly improves crop phenotype acquisition efficiency.To date,there are few reports on phenotypic studies using different UAV-based sensors in soybean.In this study,20 soybean lines including landraces and elite cultivars were planted in the summer 2021.Three different sensors mounted on UAV,i.e.hyperspectral,multi-spectral and Li DAR,were used to obtain the remote images of soybean at flowering,podding,bulging and mature stages.LAI data at the four development stages were also obtained by ground measurement.Using eight algorithms of unary linear regression(ULR),multiple linear regression(MLR),random forest(RF),support vector machine(SVM),multi-layer perceptron(MLP),BP neural network algorithm(BP)and Recurrent neural network(RNN),reversion analysis of LAI at different stages was carried out in soybean.The main results were as follows:1.There were significant correlations between LAI and the sensitive bands of original spectrum,the first derivative spectrum extracted from hyperspectral remote sensing data,optimization soil adjusted vegetation index(OSAVI)extracted from multispectral remote sensing data,and the height scale correlation parameters from Li DAR remote sensing data,indicating these remote parameters could be used for the estimation of LAI in soybean.2.Comparison of models built by the data from different types of remote sensing sensors showed that the retrieval abilities of the models based on hyperspectral remote sensing data were comparable to those of the models based on multispectral remote sensing data,and the retrieval abilities of the models based on Li DAR data were relatively poor at the maturity stage.3.The inversion capability of the model constructed by fusing of hyperspectral,multispectral and Li DAR remote sensing data was not significantly improved compared with the single hyperspectral and multispectral spectral data,but the remote sensing data fused with hyperspectral and multispectral data showed a substantial improvement in the inversion capability of soybean LAI.4.Comparison of inversion models constructed by different algorithms showed that the prediction ability and stability of the inversion model constructed based on multivariate were better than those of the inversion model constructed based on univariate.The inversion models constructed by RF,XGBoost and RNN are the best among all the reversion models.Therefore,these three algorithms could be used as important reference algorithms for LAI inversion in soybean.In addition,among all the models,the RF-LAI inversion model based on the fusion of hyperspectral and multispectral remote sensing data performed best at flowering stage,with highest R~2 value and smallest RMSE of 0.8154 and 0.2420,respectively,in the modeling set,and 0.7844 and 0.2701,respectively,in the validation set.5.There were significant differences in the inversion ability between the models constructed from the data obtained from different development stages.Among all development stages,the flowering stage and mature stage were the best,followed by podding stage and bulging stage.To sum up,in this study,the inversion capabilities of LAI based on the data from hyperspectral,multispectral and Li DAR,the univariate and multivariate inversion models constructed by different algorithms,and the optimal inversion models combining hyperspectral and multispectral remote sensing data were analyzed and compared.The results could provide theoretical guide for the selection of remote sensing sensors and algorithm for the inversion of LAI,and might also provide a new idea for high throughput phenotype retrieval in soybean. |
PDF Full Text Request