Noise-resistant Optimization Method Of Convolutional Neural Network For Remote Sensing Image Classification

Remote sensing image classification is the prerequisite and basic work of remote sensing image analysis.The accuracy of remote sensing image classification directly affects the effect of subsequent analysis.At present,with the large-scale application of deep learning in the field of natural images,many researchers have applied it to remote sensing image classification,and also achieved great success.However,the classification method of deep learning uses supervised learning,i.e.,samples that have been manually labeled need to be learned before automated classification can be performed.The performance of the network depends on the accuracy of the training sample labels.Common training sample labels are labels inherited through human visual surveys or from historical land cover maps or land use maps,which usually contain label noise based on subjective human knowledge and the timeliness of historical maps.How to help the network deal with noisy label samples during the training process is particularly important.To this end,this paper carries out the research on the noise-resistant optimization method of convolutional neural networks under noise-containing labels based on remote sensing data.The main work and innovations of this paper are as followsFirst,a noise-resistant labeled convolutional neural network framework,Weight Loss Net(WLN),is proposed to solve this problem.WLN contains three main components:(1)segmentation sub-network,which is used to generate pixel-by-pixel classification results of images and can be replaced using other segmentation models;(2)loss weight parameter λ,which is used to assign weights to each training sample,assigning high weights to clean samples and low weights to noisy samples to reduce the influence of noisy samples on the network training process and improve the network the noise-resistant performance;(3)the category balance coefficient α,which helps the network to learn each class equally and avoid overfitting the model due to the imbalance between different classesSecondly,the segmentation task was set to extracting building in Inria Aerial Image Labeling Dataset.Four types of label noise(the insufficient label,the redundant label,the missing label,the wrong label)were simulated by the dilate and erode processing to test the network’s anti-noise.The experimental process adds training samples with different noise rates and noise levels to the training dataset and tests them on a clean dataset to compare with the original U-Net network and evaluate the noise immunity performance of the networkFinally,the result shows the proposed anti-noise framework(WLN)can maintain high accuracy while the accuracy of the U-Net model dropped.Specifically,when the noise rate and noise level are relatively low,U-Net is not affected by the noise labels,which may be due to the deep learning model’s certain noise resistance,and when the noise rate and noise level gradually increase,the accuracy of U-Net network decreases significantly,while WLN can always maintain high accuracy and outperform U-Net This result shows the anti-noise framework proposed in this paper can help current segmentation models avoid the noisy training labels’ impact.