Identification Of Construction Waste From Multi-source Remote Sensing Image Based On Transfer Learning

In recent years,with the continuous progress of urbanization in China,there are more and more projects such as demolition,shantytown transformation,municipal engineering and so on.The problem of construction waste faced by the city is more and more prominent.The irregular stacking of construction waste will not only bring security risks,but also cause serious environmental pollution.It is one of the key tasks of construction waste management to supervise and check the construction waste storage.At present,it mainly adopts the form of field visit and field measurement.But this method has some problems,such as poor timeliness,large consumption of human,material and financial resources.Remote sensing earth observation technology has the advantages of long-distance detection,large area coverage,and short revisit cycle and so on.At present,it has been applied in the field of construction waste storage monitoring.Moreover,with the development of deep learning technology represented by neural network,it has achieved good application results in remote sensing image processing and recognition.Therefore,in this paper,the deep learning method is applied to the remote sensing identification of construction waste storage,and the research is carried out around two aspects Firstly,in view of the shortage of building remote sensing data set samples,the improved generation countermeasure network is used to realize the expansion of building waste remote sensing image samples;secondly,the migration learning method is used to realize the semantic segmentation of building waste remote sensing image,providing technical support for the remote sensing identification of building waste stacking.This paper mainly carried out the following work:(1)Firstly,the research status of three classical remote sensing recognition methods is reviewed.The remote sensing object recognition based on pixel only uses the spectral information of image,and can’t make full use of the context information of image,so the extraction accuracy is poor;the object-oriented remote sensing object recognition,the segmented image often can’t accurately represent the whole object,thus affecting the recognition accuracy;the remote sensing extraction method based on deep learning semantic segmentation is a hot research direction To solve the problems of the above two methods,but a large number of data samples are needed to provide support.(2)Through the summary of the research on three kinds of remote sensing recognition methods,this paper selects the remote sensing recognition methods based on semantic segmentation and makes a sample set for the research experiment.After pre-processing and clipping,956 images with a size of 512 × 512 pixels are selected as the original sample set of the 5 South Beijing images of GF-2 on September 5,2018.After field investigation and image feature interpretation,six kinds of recognition experiments were established.Through manual annotation,the corresponding labels of the original image were obtained,which accumulated sample data for future experimental research.(3)Based on the improved generation countermeasure network,the sample generation is studied.The generative countermeasure network consists of generative model and discriminant model.In this paper,the loss function of generative model is improved according to the LBP Operator which can extract features,and the feature transfer learning idea is introduced in the model training to improve the learning ability of key features of the model.In the experiment,956 original images of the sample set were used for training,and the maximum MSE(Mean Squared Error)and PSNR(Peak Signal to Noise Ratio)of the experimental results were 12.9873 and 5838.7155,respectively.Compared with the original network,they were improved,and the color and texture of the generated samples were in line with the objective reality,so they can be used as samples for data set expansion.(4)Research on semantic segmentation based on transfer learning.In this paper,DeepLabV3 model is used as the pre training model,which includes void convolution,ASPP pooling,and backbone network ResNet.There are three experiments in the experiment: The deep lab model with ResNet-50 as the backbone network is chosen as the pre training model for the construction waste semantic segmentation experiment;In Experiment 1,the model after training was used to transfer the parameters of different layers;In Experiment 1,the model after training and the generated samples were used for sample migration experiment.In the first experiment,the mIoU(Mean Intersection over Union)is 75.32%,which is consistent with the requirements of construction waste identification.The results of Experiment 2 show that the mIoU can be increased to 78.12% by the parameter transfer method.If the sample size is insufficient,the parameter transfer learning can be properly carried out to achieve a certain accuracy;The results of Experiment 3 show that using the method of generating samples in this paper to get the samples to participate in the training of the segmentation network can improve the recognition accuracy of the construction waste segmentation algorithm to a certain extent.For the research of the construction waste segmentation,the ratio of the generated samples to the real samples is 1:3.