Remote Sensing Image Classification And Segmentation Based On Convolutional Neural Network

Remote sensing image is a surface image taken by remote sensing satellite or aerial camera.Because of its bird's-eye view of the earth,it can directly observe the earth's information.It can be used for map drawing,crop area statistics,urban planning,weather monitoring and so on.It has great social and economic value.This paper presents methods based on convolutional neural network for classification and segmentation of remote sensing image: The classification is to classify remote sensing image with a label;segmentation is to classify a remote sensing image in pixel-level.The main research contents include:A multi-stage network model suitable for classification of remote sensing images is proposed.According to some kind of remote sensing images are similar,construct a two phase transformation model to do coarse to fine classification process.In the first stage,the model outputs the classification probability,and according to the mapping relation in the conversion table,the image is sent to the corresponding second stage model to do fine classification,and the final classification probability is obtained through the probability fusion layer.The experimental results show that the proposed model improves the classification accuracy and reduces the computational complexity and parameter size compared with the ensemble method.A pixel weight adjustment loss function and generative adversarial network for remote sensing image segmentation is proposed.In the segmentation network,the cross entropy based on pixel is usually used as the loss function,ignoring the phenomenon that the number of pixels in the remote sensing image is unbalanced.In this paper,we improve the loss function and increase the weight of a small number of pixel categories in remote sensing images.The experimental results show that the proposed method accelerates the training speed and segmentation performance of the model.Then,the segmentation is regarded as the generator,and the discriminator is used to discriminate the segmentation results.The generator and the discriminator alternately improve their performance through adversarial training.Experiments show that the performance of the segmentation model under the framework of generate adversarial network can improved.In addition,we crawled more than 10000 remote sensing images from Google Map to build a dataset.The experimental results show that the above classification and segmentation method can also achieve good results in the database.