Research Of Crop Classification With Remote Sensing Images Based On Convolutional Neural Network

Crop classification using remote sensing images is a very important technical method for crop area statistics and yield estimation.The traditional methods for image classification such as those based on single pixel or object-oriented,etc.,usually make use of limited spectral information or handcrafted features,but these spectral information or handcrafted features may not be the optimal for the classification task.So far there is not a method can extract features automatically and accomplish classification efficiently,in the smallholder pattern of our country.It is urgently needed to find a robust and efficient method for the crop classification in the broken blocks in China.The prevalent deep learning method can fit a complex function consisted of a large quantity of parameters with gradient descent and extract hierarchical features automatically.It has made outstanding achievements in computer vision,natural language processing and speech recognition.So far there doesn't exist any systematic application for the task of crop classification with deep learning method.Therefore,we attempts to solve the problem of accurate crop classification using convolutional neural network-a special kind of deep neural network with the expectation of finding a kind of model to learn the spatial-temporal information in remote sensing images automatically.In this paper,2D convolutional neural network is firstly used for crop classification.By using the GF-1 multi-temporal multi-spectral images,we analyzed the crop classification ability of convolutional neural network.We also compared the accuracy of convolutional neural network with some other classification methods,such as SVM?KNN?PC A?object-oriented method,and finally we find that convolutional neural network using multi-temporal multi-spectral images gained the highest accuracy.In order to prove that convolutional neural network can automatically extract the potential useful features for the crop classification task,the multi-temporal vegetation index data is taken as input data to the convolutional neural network.The corresponding accuracy is lower than that of original image data.In addition,when the original image data and the vegetation index data are all fed to the convolutional neural network,the classification accuracy has no notable improvement,this validate the fact again that the automatically extracted features with convolutional neural network are highly abstract and general.Considering the defect of 2D convolutional neural network when processing sequential images,a 3D convolutional neural network which can extract spatial-temporal features is proposed,and the experiment results show that 3D convolutional neural network has certain advantage in crop classification with multi-temporal remote sensing images.Typically,deep learning method always requires a large number of training samples.Considering the high cost of obtaining crop samples in remote sensing images,this paper proposes a method combined with active learning which can greatly improve the classification accuracy with a small amount of manual work.