Design And Implementation Of Crop Classifier Based On Hyperspectral Image

China is a large agricultural country,and agriculture is the basis for survival and development of the country.The crops classification of a large scale land is great significance for crop monitoring,crop yield estimation,general survey and post-disaster compensation statistics.In recent years,with the development of hyperspectral imaging technology,sensor technology and the satellite equipped with imaging spectrometer has been launched,making it easier to obtain hyperspectral data,and the information contained in the hyperspectral data is also more abundant,which can lay a good data foundation for the design of large-scale crop classifier..Therefore,in order to meet the needs of agricultural development,we took hyperspectral data as the data source,focused on the spectral features of crops,and designed a kind of classifier based on hyperspectral images.The classifier can effectively distinguish different crops and their related vegetation types.Its main work is as follows:(1)In order to make effective use of the crops spectral features,we proposed a crops classification framework based on the bidirectional recurrent neural network.In the framework,we directly related to the spectral value of each pixel as a spectrum sequence,and focused on sequence characteristics.The framework can extract sequence forward and backward united spectral features for classification.The experimental results showed that the bidirectional recurrent neural network has stronger feature extraction capacity,and has better classification performance than other classification algorithms.(2)As spectral curves of different crops are similar,it is difficult to distinguish them in the classification task.Firstly,we analyzed the spectral curves of the crops and eliminated the redundant bands that are obviously coincident in the curves.Then,a vegetation index set that can highlight crop characteristics was added to the original data set,and a feature selection algorithm based on redundancy and relevance was improved.The algorithm can flexibly adjust the influence of the relevance between the band to be selected and the label on the feature selection.In the classification experiment,we obtained higher classification accuracy by eliminating the redundant bands、adding vegetation index set and selecting higher relevant feature set for crops classification by the proposed algorithm.