Identification Of Growth Stages In Rice Using Remote Sensing Data

Rice growth stage information is an important indicator of farmland management,and the growth status and environment of crops can be judged by the information of the time and process of crop growth.In addition,it can also assist in large-scale crop extraction and environmental analysis.Remote sensing technology has become a powerful means of monitoring crop growth stage with its advantages of large-scale,fast,non-destructive and continuous space.In this study,the canopy spectral characteristics of rice at different growth stages were analyzed first,then,used the remote sensing data obtained from different platforms to identify rice growth stages.The main results of the study are:(1)Based on the spectral reflectance of rice obtained from ground and unmanned aerial vehicle(UAV)platforms,the K nearest neighbor(KNN),decision trees,support vector machine(SVM),random forest(RF),gradient lifting decision tree(GBDT)and Stacking method were used to identify the growth stage of rice.Based on the reflectance obtained from the ground platform,the Stacking model using random forest as the secondary learner has the highest overall accuracy(89.44%).And based on the reflectance obtained from the UAV platform,the overall effect of the GBDT is the best,with the accuracy of 98.75%.(2)The rice growth stage identification model based on the high temporal resolution data obtained from the ground platform was extended to the data obtained from the UAV platform to evaluate the applicability of the model in the cross-platform application.The overall accuracy of the cross-platform application of the decision trees model is the highest,which is 76.25%,and there are many confused samples at the jointing-booting stage and heading-flowering stage of rice.(3)The main reason for the poor applicability of the model in cross-platform applications is that the reflectance obtained by the two platforms is different,the secondary reason is that the four-band spectral characteristics of rice were similar at the two stages.First,spectral assimilation was carried out through a simple linear model to make up the reflectance difference between the two types of data.After data assimilation of the UAV platform,the recognition accuracy of the SVM model is the highest(78.96%).On the basis of the SVM model,the overall accuracy of the final rice growth stage identification is 86.67% and 83.54%,respectively,by combining the visible atmospherically resistant index(VARI)threshold(0.27)and the entropy threshold(0.4)of the 5th band of MCA image to distinguish the jointing-booting stage and heading-flowering stage.The combination of spectral assimilation and vegetation index threshold or texture feature threshold can effectively improve the applicability of the model in cross-platform application.The combination of vegetation index-VARI has the best effect.