Detection Of Growth Stages In Rapeseed Using Remote Sensing Spectral Data

The phenology information of rapeseed is important for assisting field manage-ment,catching appropriate viewing time and evaluating yield.Remote sensing technol-ogy can efficiently and non-destructively obtain large-scale vegetation information which has been Awidely applied in precision agriculture.In this study,spectral reflec-tance data was used to identify the growth period of rapeseed,and the leaf area index was retrieved by combining the phenology information of rapeseed.The main research results are as follows:(1)According to rapeseed spectral information of rapeseed,combined with the actual growth period of ground field observations.Some famous machine learning al-gorithms was used including k-nearest neighbor(k-NN),support vector machine(S VM),random forest(RF),Gradient Boosted Decision Trees,(GBDT),artificial neu-ral network(ANN)and Stacking combined classification methods to identify key growth stages of rapeseed.The results showed that the Stacking combination algorithm combined with different machine learning can significantly improve the classification accuracy compared with a single classifier.The accuracy of Stacking algorithm is high than the best single classifier(k-NN)1%.The developed model was applied to data collected from different sensors(ground-and aerial-collected data),yielding the accu-racy of 90.07%with ground-collected spectra and the accuracy of 81.48%with aerial-collected data.(2)A key growth state identification of rapeseed based on machine learning method and vegetation index was proposed.By analyzing the misclassification of the k-nearest neighbor algorithm in leaf stage and pod-development stage and the canopy spectral information of rapeseed,it is concluded that the reflectance in leaf stage and pod-development stage will be significantly different in the green band.Select VARI-green=0.35 for the threshold value distinguishes the fruit-and-leaf period which is more difficult to decimate in the near-nearest neighbor algorithm,and can significantly im-prove the identification accuracy of the rape growth period.The recognition accuracy of ASD data on the ground platform is 92.4%.The recognition accuracy of UAV plat-forms with low altitude UAV is 79.62%.The recognition accuracy rate of data without discrimination is 86.92%.The results show that it can be applied to different locations.The data of different platforms has better generalization ability.(3)According to the spectral information of rapeseed,the leaf area index was es-timated with the whole quantity of samples was trained on a single machine learning model.The best of machine learning method support vector machine has a test error of 0.401,which is higher than title test error of using enhanced vegetation index 2(0.379).Therefore,the vegetation index method can perform inversion of leaf area index with a simple and effective model.(4)Analyzing the retrieval results of LAI using vegetation index at different phe-nological states,it is shown that the flowering state of rapeseed is more confusing and brings errors to LAI retrieval.Therefore,an empirical model of LAI vegetation indices combined with phenology information is used.EVI2 had the best results at the leaf stage of rapeseed(RMSE=0.247).vegetation index CIgreen had the best results(RMSE=0.108)during pod-development stage.However,at the flowering stage of rapeseed,the error of the individual flowering period got root mean square error 0.792.The error was higher than that of using all data of rapeseed(0.753).(5)Estimating leaf area index using machine learning,vegetation index at all pe-riods and vegetation index at different periods,which indicated that the error of choos-ing the optimal retrieval model was 0.346,so estimating LAI for rapeseed combined with the phonological performance has better results.