Research On Object-oriented Classification Based On GF-1 Image

In recent years,with the continuous development of remote sensing technology,more and more satellites with high temporal resolution and high spatial resolution have appeared one after another,which provides a lot of convenience for our lives.How to make good use of these huge amounts of data and information-rich remote sensing images have always been a hot issue for domestic and foreign research.Among them,the classification of remote sensing images is the basis for applying a large number of remote sensing images to various fields.Earlier remote sensing images had lower resolution,less feature information,and the relationship between features was not obvious.Most studies used pixel-based methods to analyze,that only take advantage of the spectral information of the image,which is prone to misclassification and leakage.Later,an object-oriented analysis method for images with high resolution,rich ground features,and clear relationships between them appeared.At the same time,the object’s spectrum,topology,and spatial structure were considered,which provides a good theoretical basis for the classification of ground features using remote sensing images.At present,the traditional classification model has been difficult to meet the requirements in practical applications,and the rapid development of deep learning has made it not difficult to solve the problem of extracting large-scale image features.In this study,Xihu District,Hangzhou City,Zhejiang Province was used as the research object,and GF-1 remote sensing data was used as the information source.After performing routine pre-processing on the image data,the optimal index method is used to determine the optimal band combination.Then a two-dimensional space of the Moran’s I index and the geographic detector q was constructed to determine the optimal segmentation scale.The maximum area method was used to determine the shape factor and compactness weights.The spectral,geometric,and texture features of the segmented objects were used as input factors for the classification model.Pixel-based support vector machine classification model,object-oriented support vector machine,nearest neighbor method and BP neural network classification model were established.Finally,based on the TensorFlow learning framework,a one-dimensional convolutional neural network(1D-CNN)classification model of deep learning was established,and all models are analyzed by using the confusion matrix method.The main conclusions are as follows:(1)Using the optimal index method,the optimal band combination can be confirmed which is 4、2、1.The selected domestic GF-1 remote sensing image data was preprocessed including radiation calibration,atmospheric correction,orthorectification,mosaicking and cropping.The optimal band combination was determined to be 4、2、1 by using the optimal index method,which provides a certain reference for future research on GF-1 satellite imagery.(2)The two-dimensional space of the geographic detector q statistics and the moran’s Ⅰ index was constructed to confirm the optimal segmentation scale which was 90,and the maximum area method was used to determine the homogeneity factor of image segmentation which was 0.7 and 0.3 respectively.The two-dimensional space was constructed with the normalized geographic detector q statistics as the abscissa and the normalized moran’s Ⅰ index as the ordinate.The Euclidean distance between the segmentation optimization function(SOF)value of each segmentation scale and the best point(1,0)of its segmentation quality was calculated,and the optimal segmentation scale for image segmentation was determined to be 90.The maximum area method was used to determine the homogeneity factor of image segmentation.The result was a shape factor weight of 0.7 and a compactness weight of 0.3.The image was segmented with these parameters,and the segmented image was used to extract 62-dimensional features including spectral features,geometric features,and texture features as the input factors to the classification model.(3)Among the traditional models,the BP neural network is the best one with the overall accuracy 89.52%and Kappa coefficient 0.8737.Pixel-based support vector machine classification model,object-oriented support vector machine classification model,object-oriented nearest neighbor classification model,and object-oriented BP neural network classification model were adopted.The overall accuracy of the models are 68.75%,86.90%,84.86%,89.52%,Kappa coefficients are 0.6244,0.8421,0.8110,0.8737 respectively.It can be seen that the classification effect of the object-oriented classification model is better than the pixel-based classification model.Analyzing the user accuracy and producer accuracy of each model,we can get that the extraction of buildings and roads is difficult for the above models.(4)An object-oriented 1D-CNN-based classification model of deep learning based on the TensorFlow learning framework was established,the overall accuracy of which is 93.10%,and the Kappa coefficient is 0.9167,which is the best on of all.The 1D-CNN model has two layers of convolution,two layers of pooling,three layers of fully connected,and Dropout was used in the first two fully connected layers to prevent overfitting.The overall classification accuracy of this model is 93.10%,and the Kappa coefficient is 0.9146.In addition,the user precision and producer precision of the five selected land types are higher than 85%,except the land of transportation which is hard to classify.The classification effect is ideal,which means that the 1D-CNN model has a great advantage compared with the traditional model.