Effects of Different Sample Set Partition Strategies on Crop Remote Sensing Classification Accuracy
The extraction accuracy of crop distribution has a profound impact on the subsequent inversion of farmland parameters and estimation of crop yield per unit area.In the process of crop classification and recognition,the accuracy and number of training samples are crucial to the final classification results.Aiming at the problem of small number of samples and uneven distribution,the crop classification sample data set was constructed by two ways of field identification and visual interpretation,and five sample data set construction schemes were designed:①all the measured sample points(70%training,30%verification)were used in the scheme;② all visual interpretation sample points were used(70%training,30%verification);③the same proportion of training samples and verification samples were selected from the measured sample points and visual sample points respectively,and then the training sample set and verification sample set were constructed combinedly(70%training,30%verification);④ the visual interpretation sample points were used as training samples,and the measured sample points were used as verification samples;⑤the same number of samples were selected from the visual sample points and the measured sample points to construct a sample set(70%training,30%verification).The accuracy of crop remote sensing classification was studied under different schemes.The results showed that except④,the overall accuracy of ①②③⑤ four sample data set partition schemes was more than 95%,and the classification results were good.Using visual interpretation to supplement sample points could effectively solve the problem of fewer sample points and uneven distribution.As the best classification scheme for crop recognition and extraction in the study area,scheme③ had an overall accuracy of 97.6%and a Kappa coefficient of 0.970,and the accuracy of corn,rice and soybean was all more than 97%,indicating that the combination of training samples and validation samples selected from visual interpretation samples and measured samples to construct training and validation sample sets can not only improve the accuracy of classification results,but also improve the authenticity and accuracy of classification results.
Remote sensingCrop classificationVisual interpretationSample set constraction
国家科技期刊平台
NSTL
万方数据
