| The rapid development of industrialization and urbanization in China has greatly accelerated the growth rate of solid waste.If solid waste lacks effective management and is stacked at will,a large number of disordered solid waste storage yards will be formed,which will seriously damage the ecology,affect human settlements and aggravate the difficulty of management.Therefore,how to efficiently and accurately monitor urban solid waste yards is crucial for the timely regulation of disordered solid waste yards,and is of great significance for improving the comfort of residents lives.However,traditional monitoring methods based on field investigation are difficult to meet the needs of efficient monitoring,the macroscopicity and real-time performance of remote sensing technology can provide a new way for largescale,rapid,objective and dynamic monitoring of urban disordered solid waste storage sites.Traditional machine learning interpretation methods based on artificial design features have limited intelligence and generalization ability,so they can’t be accurately applied to the complex identification task of disordered solid waste yards.Deep learning algorithms represented by convolutional neural network can abstract advanced semantic features of images,so they have unique advantages in the field of image processing.In view of this,according to the task requirements of this thesis,the idea of instance segmentation in deep learning was introduced in order to identify the disordered solid waste yards automatically and efficiently.Specifically,this thesis carried out the following research work in view of the possible problems in the identification of solid waste yards based on instance segmentation:(1)This thesis constructed the first high-resolution remote sensing interpretation sample set of disordered solid waste yards,which is used for instance segmentation.The sample set was made by manual labeling,and a copy-paste data augment strategy based on the significance constraint of spectral residual was designed subsequently,which can sufficiently avoid the blocking of significant targets in the pasting process,thus effectively expanding the number and richness of solid waste yard samples.The experimental results showed that the F1-score and Io U of the model trained on the augmented train set could reach 75.6% and 60.8%respectively when applying on the test set,compared with the model trained on the basic train set without augmentation,the F1-score and Io U improved by 1.6% and 2.1% respectively,and it obtained the best test performance surpass other data augment strategies.(2)Taking the size distribution characteristics of disordered solid waste yards into account,this thesis adjusted the size and aspect ratio of the anchor of Mask R-CNN by KMeans clustering,so that the adjusted anchor parameters could better fit the identification task of disordered solid waste yards.The experimental results showed that the F1-score and Io U of the model which trained on the adjusted anchor parameter could reach 76.2% and 61.6%respectively when applying on the test set.Compared with the original anchor parameter,the F1-score and Io U could improve by 0.6% and 0.8% respectively,indicating that the adjusted anchor in this thesis is more suitable for the identification of disordered solid waste yards.(3)In view of the difference in sensor resolution between google images and GF-2 images,there is a difference in data distribution between them,so that the model trained on the train set cannot be efficiently applied to the test set.Toward this problem,this thesis introduced the domain adaptive algorithm to implement the consistency optimization on the data distribution of train set and test set,so that it is able to identify disordered solid waste storage yards more accurately when appling the trained model on the optimized GF-2 images.The experimental results showed that the F1-score and Io U could reach 55.8% and 38.7% respectively when applying the trained model on optimized GF-2 images,and increased by 8.7% and 7.9%respectively comparing with original GF-2 images,indicating that the domain adaptive can significant improve the identification performance of disordered solid waste storage yards on the GF-2 images. |
