Research On Remote Sensing Identification Method Of Construction And Demolition Wastes Based On Google Earth Engine And Machine Learning

With the acceleration of global construction and urbanization,the scale of demolition and new construction is expanding,and the amount of construction and demolition waste(C&DW)is increasing.Due to the serious threat of C&DW to society,economy and ecological environment,the management of construction waste has gradually been paid attention to.Monitoring the spatial distribution of C&DW is one of the important research on intelligent supervision of C&DW.At present,the main methods for monitoring C&DW are manual field investigation and remote sensing monitoring.The former has the problems of high cost and low efficiency.The latter identifies C&DW from remote sensing images by simply constructing remote sensing monitoring index.The complexity of spectrum and texture of C&DW makes the identification efficiency and accuracy unable to meet the requirements.Machine learning algorithm has excellent performance in remote sensing image classification,high accuracy and efficiency,and can bring new methods for C&DW identification.But machine learning algorithm parameter combination of different C&DW identification effect is different,the use process consumes a lot of data storage and computing resources.In order to explore the ability of different machine learning algorithms and different parameter settings to identify C&DW,realize the rapid identification and positioning of C&DW dumps,and reduce the limitations of machine learning in the use process,this paper explores the effect of four machine learning algorithms on the rapid identification of C&DW remote sensing combined with Google Earth Engine.The main research work and conclusions are as follows:(1)The high-precision Sentinel-2 remote sensing image data were obtained based on Google Earth Engine.Taking the megacity Beijing with high risk of construction waste pollution as the research area,the cloudless Sentinel-2 image was generated by image preprocessing(including clipping,declouding,median synthesis,etc.),and the remote sensing monitoring image database of the research area was established.Aiming at the generation,stacking and processing of construction waste,based on remote sensing images,a sample library of remote sensing morphology,texture,structure and spatial distribution of construction waste is constructed,which accumulates sample data for machine learning experiments and meets the needs of different machine learning algorithms for construction waste recognition.(2)The mainstream machine learning algorithm decision regression tree(CART),random forest(RF),support vector machine(SVM)and deep learning algorithm are selected to identify and calculate the C&DW,and to draw the spatial distribution map of C&DW dump in satellite images.The C&DW identification method based on shallow feature extraction model uses empirical selection method,experimental trial and error method and grid search method to explore the optimal parameterization scheme of decision regression tree,random forest and support vector machine.In the experiment of C&DW recognition based on deep feature extraction model,Tensorflow deep learning framework is used to process remote sensing images of C&DW,and these data are used to train U-Net model,and Google Earth Engine is used to output remote sensing recognition results of C&DW.(3)By evaluating the accuracy of the above four algorithms and field verification results,the machine learning method suitable for remote sensing recognition of C&DW is determined.In the identification of C&DW based on shallow feature extraction model,the overall identification accuracy of decision regression tree,random forest and support vector machine for C&DW is 73.12 %,98.05 % and 85.62 %,respectively.In the C&DW recognition based on deep feature extraction model,the semantic segmentation deep learning model is used to monitor the C&DW recognition.The loss function evaluation U-Net model is trained well and good recognition results are obtained.In terms of the accuracy of category pixels on the evaluation data,the recall rates are 64.1 % and 85 %,respectively.The trained model achieves60.3 % intersection between the predicted and real observations in the verification image.(4)The four algorithms show good recognition ability and high performance for remote sensing recognition of construction waste,and the recognition results are basically the same.The recognition efficiency and classification accuracy are generally in line with the requirements of remote sensing recognition of construction waste,which further proves the effectiveness of machine learning method in intelligent recognition of construction waste based on remote sensing images.Comparing the random forest algorithm in the shallow feature extraction model with the deep learning algorithm in the deep feature extraction model,both have advantages and disadvantages.In the process of construction waste identification,the random forest algorithm has fast recognition speed and high accuracy,but it has noise interference.It has poor recognition effect on small-scale construction waste storage points,and it is difficult to monitor the complete edge.It is easy to lose details.The deep learning algorithm can learn more deep features of construction waste from the combination of context information and detail information,and the results of extracting the edge information of construction waste are better.The contour of the identified construction waste dump is relatively complete,and more accurate recognition results of construction waste are output.But compared with the random forest algorithm,the running time is longer.Therefore,different machine learning identification methods can be selected according to different needs,that is,the random forest algorithm can be used to quickly locate the construction waste dump,and the deep learning algorithm can be used to obtain the accurate range of construction waste dump.The study constructs an extensible and efficient machine learning recognition framework for C&DW,which uses data and computing resources of Google Earth Engine to identify and supervise C&DW.The method of remote sensing identification of C&DW based on machine learning is explored,which can identify C&DW timely and accurately,and provide scientific basis for intelligent supervision and resource utilization of C&DW.