Farmland Extraction From Multi-band Remote Sensing Image Based On Semantic Segmentation Network

With the development of remote sensing technology,it is more and more easy to acquire multi-band high-resolution remote sensing image,so how to automatically and quickly process the rapidly increasing remote sensing data has become an urgent problem to be solved.At present,deep learning technology has shown increasingly excellent results in image processing,therefore,it is becoming an increasingly important trend to explore how to use deep learning technology in remote sensing images.However,in the study of extracting typical resources and environment elements from multi-band remote sensing images,the following problems will be faced:(1)For multi-band remote sensing image,there is no dataset with hundred thousands or millions labeled images to pretrain the deep learning network.(2)High accuracy is often required for processing the multi-band remote sensing tasks,on the other hand,this task needs to process a large number of remote sensing images,which requires the algorithm used to have a fast running speed.(3)For multi-band remote sensing images,labled images only accounts for a small part,while there are a large number of unlabeled data.In order to solve these above problems,we studied the extraction of farmland from multi-band remote sensing image based on semantic segmentation network.The main work and innovation of this master paper are as follows:(1)For the problem of no pretraining dataset,the pretraining steps are abandoned and a semantic segmentation network with few parameters and good segmentation performance----Deep Feature Aggregation Network(DFANet),is utilize to extract farmland.The experimental results on the farmland dataset show that compared with remote sensing image with RGB bands,the semantic segmentation network of DFANet using multi-band remote sensing image can improve the precision of farmland extraction.Meanwhile,compared with random forest,Convnet and Integrated Convnet,the DFANet using multi-band images has more excellent results of farmland extraction.(2)In order to further improve the precision of farmland extraction through DFANet based on the farmland dataset,the feature extraction part,the semantic module and the decoder of the semantic segmentation network are optimized,then the Improving Deep Feature Aggregation Network(IDFANet)is proposed.Due to the redundant subnetworks in IDFANet,in order to improve the performance of farmland extraction,pruning is carried out on this IDFANet,therefore,Pruning of Improved Deep Feature Aggregation Network-1(PIDFANet-1)and Pruning of Improved Deep Feature Aggregation Network-2(PIDFANet-2)are proposed.Experiments on the dataset of farmland show that IDFANet can further improve the precision of DFANet for the farmland extraction.At the same time,compared with IDFANet,PIDFANet-1 after pruning can improve the speed of the network by 37% without reducing the precision of farmland extraction.The PIDFANet-2 after pruning has a running speed similar to PIDFANet-1,which is 0.97 times as much as that of PIDFANet-1.Moreover,the Miou and Kappa coefficient of farmland extraction by PIDFANet-2 are improved by 0.19% and 0.21% respectively.(3)For the multi-band remote sensing images,there are a small number of labeled samples and a large number of unlabeled data.In order to construct two different frameworks,that is GAN-1 and GAN-2,the semantic segmentation network is used to replace the discriminator and generator in Generative Adversarial Networks(GAN).Therefore,by using semi-supervised learning method,the unlabeled data is applied to network training to reduce the use of labeled data.In GAN-1 framework,the semantic segmentation network is used as discriminator,and this semantic segmentation network is trained by few labeled samples,a large number of unlabeled samples and many generated samples.In GAN-2 framework,the semantic segmentation network is used as generator,the unlabeled samples are utilized to constrain the network to ensure that their predictive maps of them are consistent with that of labeled samples,which can prevent overfitting and improve the generalization ability of the network.Experiments on the dataset of farmland show that when GAN-1 and GAN-2framworks are applied to the training of PIDFANet-1 and PIDFANet-2,a good performanece of farmland extraction can be maintained while the labeled samples are reduced.Compared with Seg Net,ENet and ICNet with 100% labeled samples,PIDFANet-2 trained in GAN-1 and PIDFANet-2 trained in GAN-2 with40% labeled samples have relatively few parameters and faster running speed,but still have relatively high accuracy when using such a small amount of labeled data.