Remote Sensing Image Classification Based On Fully Convolutional Networks

The classification of remote sensing image is one of the key problems in remote sensing image interpretation,which has attracted the attention of domestic and foreign scholars.The classification of remote sensing image can be classified into two categories:unsupervised classification and supervised classification.The supervised classification method selects the underlying characteristics like color,texture and shape from the image as prior knowledge,and normally achieve higher classification accuracy.However,the improvement of image spatial resolution makes the similarity of characteristics of the same surface feature decrease and different surface feature increases.In this case,whether using a single feature or multiple features,there is always a big difference between the underlying characteristics and the upper understanding of the image,which makes it difficult to meet the actual demand of the remote sensing image classification.The development of deep learning method provides new ideas for the classification of high-resolution remote sensing images,which can learn from the input data a large amount of bottom-to-top features by the self-learning of multi-layer neurons,which can avoid the limitation of traditional classification methods.This paper summarizes the relevant research theory and applications at home and abroad.Taking the part of Postdam City,German as the research area and true-orthophoto and DSM of the ISPRS WG II/4 dataset as experiments data,the fully convolutional networks(FCN)of deep learning field is used to study the effect and advantages of FCN in remote sensing imagery classification,and the impact on classification results of using DSM data and conditional random fields(CRF)model to introduce spatial information constraints.This paper includes the following contents and achievements:(1)On the basis of existing training samples,a series of geometric transformations and radiation transformations are adopted to obtain the final training set data,making the training set expanded by 8 times.Therefore,to some extent,the over-fitting phenomenon of model training is reduced.(2)The FCN model is used to realize the classification of remote sensing image,and the classification results of FCN model are compared with the support vector machine,the k-nearest neighbor method and the maximum likelihood method.The experiment result shows that FCN is superior to the other three traditional classification methods both in visual effect and index of quantitative assessment,which is a kind of remote sensing image classification method with great potentiality for application.(3)The DSM data is added to the training set data to participate in the training process.The experiment result shows that the classification results after adding DSM data improve obviously both in visual effect and index of quantitative assessment,and the ROC curve is also obviously better than before,which verified the validity of elevation information for improving classification results.(4)The CRF model is used to process the classification results,and multi-set of classification experiments are conducted to determine the optimal combination of parameters of the gaussian kernel.The experiment result proves that the CRF model has edge smoothing and denoising effect on classification results to a certain extent.