Research On The Application Of Convolution Neural Network In High-resolution Satellite Remote Sensing Image Land Cover Classification

Land cover is an important geographic information resource for natural resource monitoring,territorial spatial planning,geo-national census,and macro-control.It is an important and indispensable information for analysis.In recent years,high-resolution satellite image data acquisition capability has been increasing,which provides a solid data basis for land cover classification.However,in practice,a large number of land cover classifications still rely on manual decoding and processing,which is difficult to meet the requirements of current large-scale,massive image land cover classification processing and extraction.In order to meet the requirements for efficient,accurate,and automated land cover classification applications,this paper investigates the use of the Convolutional neural networks for high-resolution satellite imagery land cover classification,which is proposed as a key technique based on the specific method of cover classification improves the accuracy and efficiency of land cover classification.The main research content of the paper is as follows.(1)With the characteristics of high-resolution satellite images and the requirements of land cover classification taken into account,a Land Cover Classification Convolutional Neural Network(LCC-CNN)including an encoding and decoding structure is designed.The LCC-CNN enhance the ability of feature extraction and fusion by increasing Multiscale perception block and multiscale fusion block.By fusing the low-level feature information and improving the loss function,the effect of edge learning and classification is enhanced,and the classification ability of LCC-CNN is enhanced.(2)Based on the characteristics of remote sensing image data,the ground cover classification training sample augmentation method was investigated.An improved SMOTE algorithm is proposed to perform data augmentation and filtering by computing the learning effect of training samples by convolutional neural networks.The interference of invalid samples on the training process is reduced.The training sample set constructed using the improved SMOTE algorithm provides more accurate land coverage at the same data augmentation scale.(3)The spectral,textural and slope features of the training data were expanded to take into account the characteristics of multispectral image data,and the LCC-SPL was optimized.LCC-CNN training process,adding CRF post-processing method to further improve the classification accuracy.The experimental data show that LCC-CNN classification accuracy is better than the comparison method,and the improvement of IOU and Kappa coefficient is more distinct.Using the pre-trained network with a small amount of training sample data,the LCC-CNN has the ability to extract and identify sugarcane growing areas.(4)In this paper,multi temporal data is used to improve the accuracy of land cover classification.Based on the strong generalization ability of LCC-CNN in land cover classification,a parallel classification network of front and back time phases is constructed with LCC-CNN as the backbone network,and a multi time phase modified land cover classification method is proposed.The method is able to determine the correction area based on the similarity between the truth value of the pre-temporal phase and the classification feature of the pre-temporal phase land cover classification result.