Municipal Solid Waste Detection Of High Resolution Remote Sensing Images Based On Machine Learning

As the acceleration of urbanization,the problem of solid waste being piled up everywhere has become increasingly serious.So the demand of rapid detection on municipal solid waste is becoming stronger.Most previous studies focus on the identification of large-scale landfills using manual visual interpretation and supervised classification methods however these methods face many problems such as low degree of automation,hard to establish diagnostic criteria and also hard to detect the small-scale open-air local solid waste.In order to solve the limitation of existing methods,this paper introduce region-based convolution neural network(RCNN)and traditional machine learning to apply an object detection on solid waste.And in order to solve the problem of lack of sample for Faster R-CNN,this paper also provide a convolutional neural network framework based on sample composition for object detection on solid waste(SC Faster R-CNN).The main research contents and results are as follows:(1)The solid waste detection based on traditional machine learning.This part introduces the thought of traditional machine learning for object detection and realizes object detection on solid waste which is based on selective search(SS),local binary pattern(LBP),scale invariant feature transform(SIFT),histogram of oriented gradient(HOG)and support vector machine(SVM)artificial feature extraction.The comparison experiment shows the object detection method based on SS+LBP+SVM has the superiority of precision and efficiency.Applying SS+LBP+SVM on two wide range of remote sensing images,the precision rate and recall rate are both below 40%.Which shows that the model can't meet the needs of practical engineering.(2)A sample composition based convolutional neural network for objection on solid waste.This part introduces the thought of convolutional neural network for object detection.To resolve the shortcoming of existing model,which is insufficient information input of sample,it provide a convolutional neural network framework based on sample composition for object detection on solid waste--SC Faster R-CNN.After that object detection based on R-CNN,Fast R-CNN,Faster R-CNN and SC Faster R-CNN are realized.The comparison experiment shows the absolute superiority of SC Faster R-CNN in precision and efficiency.SC Faster R-CNN has significant increases in precision and efficiency compared with traditional machine learning.On two wide range of remote sensing images,the precision rates of SC Faster R-CNN are 96.29%and 77.45%,the recall rates are 98.11%and 92.94%,and the total detection time are 356s and 287s,respectively.From this result,features which are learned actively by convolutional neural network are more suited to express solid waste sample than manual feature,and sample composition can effectively restrain wrong detection and missed detection.