Research On Remote Sensing Automatic Identification Of Construction Waste Based On Multi-source Data

With the acceleration of urbanization and the rise of high-consumables and high-emission construction industry,the production of construction waste is increasing year by year.The disorderly stacking of a large amount of construction waste has caused a series of urban environmental problems.Therefore,the supervision of construction waste has become one of the major problems that countries around the world need to solve urgently.It is a time-consuming and labor-intensive task to supervise construction waste through manual field investigation,and it is difficult to achieve timely supervision of large-scale construction waste.The development of remote sensing technology provides data support for the automatic monitoring of construction waste,which makes it possible to realize the regional automatic monitoring of construction waste.With the wide application of deep learning in the field of remote sensing image analysis,remote sensing recognition of construction waste based on deep learning provides an alternative method for efficient monitoring of construction waste,but the accuracy of its recognition depends largely on the quality of the training sample set.Due to the complex composition and irregular distribution of construction waste,there is no typical texture and spectral characteristics.The traditional method of establishing sample data set based on visual interpretation of single data source is difficult to support the construction of large sample size construction waste sample data set.Therefore,according to the characteristics of construction waste and multi-source remote sensing data,the accurate and efficient construction of sample data sets is of great significance to the identification of construction waste based on deep learning algorithm.Based on the above problems,to obtain high-efficiency and high-precision construction waste extraction results,Pingdingshan City,Henan Province and Jining City,Shandong Province are used as the experimental research areas.In this study,by integrating multi-source data and object-oriented methods,combined with deep learning algorithms,the construction waste area is automatically identified,and an efficient remote sensing automatic identification method for construction waste is constructed.The main research contents and results of this paper are as follows:(1)For the problem of mixed classification of construction waste and other ground objects when labeling construction waste samples based on visual interpretation,this paper is based on GF-2,Sentinel-2,land cover data,and assisted by geographic information data(Open Street Map).Based on the remote sensing cloud computing platform Google Earth Engine(GEE),the Sentinel-2 data and land cover data in the study area were processed to complete the accurate extraction of vegetation areas and bare soil areas with similar spectral characteristics of construction waste.Based on the Open Street Map data,the registration of the above extraction area and the GF-2 data in the study area is completed,and the background removal operation of the image is completed through the mask.(2)Aiming at the problems of time,manpower and material consumption in labeling sample data sets by traditional deep learning algorithms,and the lack of open and available deep learning remote sensing data sets of construction waste,this paper proposes to integrate multi-source data to remove the background of complex objects and combine multi-scale object-oriented method for automatic labeling of construction waste sample data sets.In terms of construction waste sample data set labeling,compared with the method based on traditional visual interpretation,the automatic labeling of construction waste sample data set based on multi-scale object-oriented method has significantly improved the labeling efficiency,which provides data support for subsequent deep learning model training.(3)Using the constructed construction waste sample data set based on multi-source data,two mainstream deep learning algorithms Deep Labv3 + and U-Net are selected to construct an automatic identification method for regional construction waste.In order to test the effectiveness of the construction method of this study and the effect of this method on the improvement of the accuracy of automatic identification of construction waste,the study also constructed a visual interpretation of construction waste sample data set based on GF-2 and an object-oriented construction waste sample data set based on GF-2,and combined with deep learning algorithms(Deep Labv3 +,U-Net)for comparative verification.The results show that the construction waste sample data set based on multi-source data is better than the sample data set based on a single data,and the U-Net model is more effective than the Deep Labv3 + in improving the recognition accuracy of construction waste.The m Io U of this method is 2.72 % higher than that of the construction waste sample data set based on single data visual interpretation,and the overall accuracy OA is increased from 89.56 % to 91.7 %,and the recall rate is increased from 90.39 %to 92.46 %.Compared with the recognition results of sample data sets constructed based on single data and object-oriented methods,the sample data sets constructed based on multi-source data have improved the overall accuracy OA and intersection over m Io U by 0.85 % and 1.58 %respectively,and achieved good recognition results.It is further verified that the construction of construction waste sample data set based on multi-source data and object-oriented method,combined with network model U-Net,is effective in realizing remote sensing automatic identification of construction waste and improving recognition accuracy.The automatic identification method of construction waste based on multi-source data sample data set combined with U-Net proposed in this paper has certain innovative value in the construction method of automatic identification sample library of construction waste,enriches the training sample library of deep learning of construction waste,and improves the efficiency and accuracy of regional automatic identification of construction waste.The research results can effectively support the dynamic supervision of construction waste,improve the efficiency of construction waste resource utilization,and provide technical support for urban sustainable development.