High Resolution Remote Sensing Image Classification Based On Regional Consistency

In recent years,as the continuous development of space technology,information technology and sensor technology,the ground details description ability of remote sensing image has been improved,high-resolution remote sensing images are widely used in practical applications in various fields.The semantic segmentation of high-resolution images is the basis of the practical application of high-resolution images,and the semantic segmentation quality has a direct impact on the practical application,so the research on the semantic segmentation method of high-resolution images has important practical application value.Remote sensing image annotation method based on the deep convolution network has shown a more superior performance than traditional methods.But due to its context information retrieval method based on fixed receptive field size does not explicitly use the constraint relationship between pixels,lead to failure to overcome the challenge of high resolution image and its spectral character bring,and result in inconsistent semantic annotation results in the same object internal.Based on the assumption that pixels in the same area have lager probability to belong to the same category,this paper attempts to introduce the consistency of semantic annotation in region internal to improve the existing depth convolution neural network's ability to describe the context information and correct the spatial relationship of the prediction results of convolution neural network.Specifically,this paper introduces regional consistency constraint information in the following two ways:(1)Based on existing fully convolutional network model for feature extraction and classification of images,bring in a loss function to expression consistency of pixels features in a region by utilizing last convolution neural network features,combine this loss function with softmax loss function to co-train and get network model parameters.Experiments were performed on the ISPRS(International Society for Photogrammetry and Remote Sensing)Vaihingen 2D semantic classification data set to validate the method proposed in this paper.The experimental results show that compared with the existing FCN model,the method proposed in this paper has achieved better classification results on most category,the overall classification accuracy achieved 85.18%.This paper introduce category consistency of regional internal pixels into fully convolutional network which can effective capture the context information among the pixel features in a region and improve the spatial relationship consistency of prediction results of traditional FCNs,and get better classification results with better regional consistency,thus improve image annotation effect.(2)In order to further utilize the constraint relation between adjacent pixels,this paper introduces the conditional random field(CRF)to deal with the output of the convolution neural network.In the post-processing of CRF for the FCN prediction results,a fully connected CRF based on pixel point and regional consistency is considered..Two pairwise potentials are introduced and defined in terms of the average color vectors and average positions of hyperpixel region.Experiments were carried out on the Vaihingen 2D semantic annotation dataset of ISPRS.The results show that the CRF proposed in this paper can smooth the rough prediction boundary after combining the region information of meanshift pre-segmentation,the segmented contour boundary is more natural and smooth,the noise at the boundary and the hole in the area are reduced,overall accuracy increased to 85.70%.Moreover,the experiment found that the meanshift algorithm for pre-segmentation caused the decrease of the final classification performance when the spatial bandwidth and color bandwidth increased,and the regional scale bandwidth parameters were not very sensitive to the results.