Semi-supervised Multi-scale Deep Neural Networks For Remote Sensing Image Classification

Remote sensing technology provides a convenient technology for people to explore the world.This technology can select different detection bands for different targets and tasks,quickly monitor the target area from space,and obtain real-time information to expand human vision range.Therefore,it is widely used in many fields such as medicine,agronomy,geology and military.However,there are still some problems in the field of remote sensing image classification.The number of labeled data that can be used for training classifier is small and the classes are unbalanced,which put high requirements on the performance of the classifier.Moreover,unique imaging mechanism of remote sensing images makes it different from natural images.Therefore,if deep neural networks are applied to remote sensing image classification without combining with data characteristics,information of remote sensing image would be wasted.To solve these problems,this thesis carries out in-depth research on the classification of remote sensing images.The main contributions are listed as follows:1.A Pol-SAR image classification method based on multi-scale fully convolutional network is proposed.This method combines multi-scale wavelet and deep semantic segmentation network,and proposes a new classification network named non-subsampled contourlet fully convolutional network.In addition,a lightweight non-subsampled contourlet fully convolution network is also designed for the small-area Pol-SAR image classification.Experimental results on three data sets prove that the proposed method can extract the multi-scale deep feature of Pol-SAR image,thereby improving the classification accuracy.Moreover,these networks save time and hardware consumption and improve efficiency as they do not need to be trained and tested on pixel cubes like traditional CNN does.2.A hyperspectral image spatial-spectral feature fusion classification method based on dualchannel network is proposed.Based on previous chapter,this method uses non-subsampled contourlet to design a non-subsampled contourlet convolutional neural network.Combined with recurrent neural network,these two networks form a multi-scale dual-channel network to extract multi-scale deep spatial and spectral features of hyperspectral image.Then,multikernel learning classifiers are designed according to the one-to-all strategy to merge extract-ed spatial and spectral features.Finally,the results of these classifiers are integrated to obtain the classification result.The experiment results on three data sets show that the proposed method can fuse spatial-spectral feature,improve the classification accuracy of hyperspectral images,and obtain more homogeneous classification map.3.A hyperspectral image classification method based on MDCPE co-training algorithm is proposed.In this method,a novel co-training algorithm MDCPE is proposed to train multiscale dual-channel network in the previous chapter.This algorithm can extract spatial and spectral features both from labeled data and unlabeled data,which makes up for the shortcomings of insufficient labeled data.At the same time,the proposed algorithm uses k means clustering to make the updated samples more balanced,which effectively improves the low accuracy caused by class imbalance.Experiments are carried out on three sets of data sets and various comparative methods.It is concluded that the proposed method can obtain high classification accuracy on a small number of samples with unbalanced class.