Multi-label Classification Of High-resolution Remote Sensing Image In Opencast Coal Mines

Opencast coal mining has caused huge changes in the terrain of the mining area and caused problems such as damage and pollution of the land in the mining area.The classification of land types in opencast coal mines can provide strong support for the investigation of land use in opencast coal mines.The high spatial resolution remote sensing image contains rich feature information,the types of data are rich and the acquisition method is convenient.So,it has become one of the important information sources in the mining area.Traditional pixel-based and object-oriented remote sensing image classification methods are difficult to effectively use the rich information of high spatial resolution remote sensing images to obtain semantic information at the scene level of the image.Therefore,this paper adopts the "scenario-oriented" remote sensing image classification idea.In view of the problem that the single-label classification of remote sensing images in open-pit coal mines cannot fully reflect the feature information contained in the area,which is not conducive to scene understanding,this paper uses different multi-label classification methods of remote sensing images and different label relationship learning strategies to achieve multi-label classification of high-resolution remote sensing images in open-pit coal mines.The main research contents of this article are as follows:(1)Based on the interpretation and analysis of the high-resolution remote sensing images of open-pit coal mines,this paper proposes a multi-label data set of highresolution remote sensing images of open-pit coal mines,which is used for multi-label classification of remote sensing images of open-pit coal mines.This paper improves the transfer convolutional neural network used for single label image classification,making it suitable for multi-label classification of remote sensing images.(2)This paper discusses the dependency relationship between the category labels of various objects in the dataset.On this basis,a multi-label classification model of remote sensing images based on the encoding-decoding structure of convolutional neural network,attention mechanism and recurrent neural network is adopted.Convolutional neural network is used as the feature extractor of the image.Use the attention mechanism to capture the corresponding feature information of a specific label.Finally,the feature map will be input into the recurrent neural network.Using the powerful decoding ability of the recurrent neural network,the label relationship is modeled and the prediction results of multiple labels are generated.(3)In order to better study and model the relationship between labels in datasets from a global perspective,a multi-label classification method for remote sensing images is proposed,which integrates attention mechanism and graph convolution neural network.On the basis of convolution neural network,the attention module of convolution block is added to strengthen the correspondence between category label and region in image.Finally,the graph convolution network model is used to map the labels to a group of interdependent object classifiers to classify the features.The classification effect is improved by adding convolution block attention module.Finally,multi-label classification of remote sensing images in mining areas is realized.