| Maize is a major staple crop worldwide,widely grown in many countries for its remarkable adaptability to various environments.In this regard,it is of great practical significance for farmers to have timely,accurate,and nondestructive pre-harvest estimation of maize yield,which helps them make reasonable decisions,such as crop insurance,storage requirements,fertilizer,water,and other input factors.Traditional yield measurement techniques involve sampling surveys,which are time-consuming and labor-intensive as they require a large area of ground destruction in maize fields.Fortunately,advancements in unmanned aerial vehicle(UAV)platforms and sensor technology offer novel insights into yield estimation.However,existing yield estimation models only consider traditional feature indicators such as vegetation index,canopy coverage,meteorological factors,and soil data without factoring in the unique characteristics of maize under lodging stress.Additionally,the existing feature indicators are instantaneous values,while spectral reflectance can be easily affected by external environmental factors,leading to discrepancies in the instantaneous indicators extracted at different times.Therefore,the objective of this study was to use a low-altitude UAV platform equipped with RGB and multispectral sensors to collect high spatial resolution images of maize canopy under different varieties and nitrogen treatment experiments for two years.Subsequently,we extract the spectral information(vegetation index)and structural information(lodging index and stay-green index)of the maize canopy to accurately estimate maize yield at each growth stage.The main contents and conclusions of this study are as follows:(1)In this study,the yield estimation model was constructed based on vegetation indices.The input parameters were divided into two categories of total vegetation indices and the best combination of vegetation indices after feature screening.The accuracy changes of both models were consistent in the years 2020 and 2021 at different growth stages.The study showed that in the yield estimation model involving all vegetation indices,the highest accuracy of the model was obtained at the denting(R5)stage in 2020(R~2=0.709,RMSE=1.207t/ha,r RMSE=14.6%),and at the doughing(R4)stage in 2021(R~2=0.861,RMSE=0.636t/ha,r RMSE=8.6%),this result can also be obtained in the yield estimation model of the best vegetation index combination.It should be noted in the research that the 28 vegetation indices selected in the 2021 dataset are significantly correlated with the actual yield,while the correlation between them in the 2020 dataset is low.Moreover,after feature screening,the number of feature combinations in each growth stage in 2021 was much less than that in each growth stage in 2020,which may be due to the effect of lodging stress on the field maize population in 2020.Furthermore,based on the data from two years,the effect of the yield estimation model between years was verified.Taking the R5 stage dataset as the benchmark,the model was constructed with the dataset of 2020,and the model was verified with the dataset of 2021.The results showed that the effect of estimated yield was far lower than that of the data set of a single year.Results indicated that the yield estimation effect between years was much lower than that of the single year dataset for the model.(2)Lodging disaster is a common occurrence in the tasseling(VT)stage,which results in the decrease of maize actual yield in the final harvest.Therefore,an index representing lodging information in the field(lodging index)was introduced into the yield estimation mode in this study,which improved the performance of the yield estimation model in the single growth stage,especially near the VT stage when lodging occured,the accuracy of estimation was greatly improved,R~2 increased by 62.26%,RMSE decreased by 12.86%.Furthermore,the study found that the R5 growth stage remained the optimal period for yield estimation when incorporating the lodging index in the model.The inclusion of the lodging stress information of maize as a feature in the model allowed for a more comprehensive description of the yield change differences and resulted in an overall improvement in the stability of the yield estimation model.(3)Traditional yield estimation models are partly based on spectral information(vegetation indices)of maize in different growth stages,but the accuracy and stability of such yield estimation models are poor.Meanwhile,the longer the green duration of maize leaves,the longer the photosynthetic duration,and the higher the yield at harvest.Therefore,in this study,the relative index(stay-green index)extracted from time series remote sensing images was introduced to construct the yield estimation model,which alleviated the systematic errors of the inconsistent photosynthetic duration of maize caused by different maize varieties and different nitrogen treatments.Thus,improving the accuracy and stability of the yield estimation model.The unitary regression model of five forms(linear function,exponential function,logarithmic function,quadratic polynomial function,and power function)was constructed by using the stay-green index,SAVI,and EVI as input parameters,meanwhile the dataset was randomly divided 20 times(training set:testing set=8:2),the accuracy distribution interval was obtained,and the model corresponding to the three parameters was comprehensively compared.Results indicated that the yield estimation model corresponding to the stay-green feature had the best accuracy and stability.In addition,under the comparison of different algorithms,the yield estimation model corresponding to the green holding index is better than the model without the green holding index.The follow-up study will take this as a reference to further explore the close relationship between the stay-green feature and crop yield.Figure[23]Table[9]Reference[131]... |
PDF Full Text Request