Research On Typical Crop Classification Based On Multiple Features

Crop remote sensing classification is of great significance for crop species identification,planting area acquisition,yield prediction,and guidance of agricultural production management.Affected by various factors,it has always been an important and difficult topic in agricultural remote sensing monitoring.The existing research has insufficient research on the classification applicability of different features and the comprehensive application of multiple features.In order to solve the problem that the results of crop remote sensing classification are not ideal due to the lack of classification applicability research for different features and comprehensive application research for multiple features,this paper takes Youyi County,Shuangyashan City,Heilongjiang Province as the research area to conduct remote sensing classification and recognition of major crops such as soybeans,corn,rice,and so on.The following work has been completed:Firstly,the spectral characteristics of different crops under ground measured hyperspectral data were studied,and the classification applicability of characteristic bands,the ability to carry original band information,and the changes of different classification methods before and after band selection were discussed.The classification applicability and importance of spectral and texture features under Landsat 8 image data were analyzed;Secondly,combine the phenological information and temporal characteristics of crops to determine the optimal feature subset of different temporal data and the optimal acquisition time for crop remote sensing classification,and combine multiple temporal features to solve the problem of category mismatch or missing points caused by the inability of single temporal data to reflect differences in vegetation growth;Finally,the differences between the presence and absence of red edge band satellites in crop remote sensing classification were studied,and the impact of red edge band on the separability of crop categories,the degree of impact on crop classification results,the importance of different red edge bands,and the role of red edge band in crop remote sensing classification were analyzed.The main conclusions are as follows:(1)The combination of multi temporal features can improve the effectiveness of crop recognition and classification.After combining multi temporal features,the overall classification accuracy reaches99.8%,and the Kappa coefficient reaches 0.9972.This can effectively complement the feature information that has the best classification effect for various types of crops at each time,avoiding the problem that single temporal data cannot reflect differences in vegetation growth;Different crops have different recognition effects at different times.In early August,the overall classification effect was the best,with an overall classification accuracy of 94.9% and a Kappa coefficient of 0.9346.At this time,soybean is at the pod setting stage,maize is at the flowering stage,and rice is at the heading stage.(2)Among the spectral characteristics,the red light band,near infrared band,and vegetation index are of high importance;Among texture features,the mean value and the feature values calculated from the near infrared band are of high importance.(3)The red edge band has significantly improved the classification effect of crop remote sensing,especially for cash crops and rice.Among the three types of satellite data,Sentinel 2 satellite with red edge band has the best effect on crop classification and recognition,with an overall classification accuracy of 96.5% and a Kappa coefficient of 0.9550;After adding red edge bands,the J-M distance between rice and cash crops increased from 1.55116278 to 1.9998-2017,significantly enhancing the degree of separability;Among the three red edge bands,B5 is more important,with the overall classification accuracy of 94.1%,higher than B6 and B7,with B5 participating.(4)Feature band filtering can effectively improve the accuracy and operational efficiency of classification methods,while retaining original information.After filtering the feature bands,the classification accuracy of the Markov distance is improved from 14.42% to 96.33%,the time consumption of the BP neural network is reduced by 94%,and the convolution layer number of the 1DCNN is reduced from 8 to3,reducing the time consumption by 51%;The overall classification accuracy and Kappa coefficient of 1DCNN are the highest,with 99.46% and 0.9942 respectively.However,the training process is relatively complex and has poor interpretability.The classification effect of Markov distance on economic crops is better,with no errors or omissions.The BP neural network is somewhere between the two,with a total classification accuracy of92.52% and a Kappa coefficient of 0.9216.