Big-scale Urban Land Cover Mapping From High-spatial Resolution Remote Sensing Images

Large-scale urban land cover mapping can provide essential data for urban ecological construction,disaster prevention,and urban management.High spatial resolution remote sensing images are the domain data for large-scale urban land cover mapping.However,large-scale urban land cover classification is still challenging:1)the problems of scale effect,the same object with different spectrum and the different objects with the same spectrum,re-sult in the low accuracy of image segmentation,seriously affecting the image pre-processing and land cover classification;2)the standard image segmentation methods are inefficient in large-scale image segmentation;3)it is difficult to collect samples for large-scale urban land cover classification task and the accuracy of land cover classification is low.To solve the above problems,I proposed Deep Merge to solve the problem of low accuracy of image seg-mentation,Dynamic pruning to solve the problem of low efficiency of image segmentation,and City Eye a using object-based weakly supervised deep learning land cover classification method to improve the sample collection of large-scale urban land cover classification ensur-ing the classification accuracy.Therefore,the novelties of this paper are as follows.(1)Proposed a deep learning-based region merging method Deep Merge.To address the problem of low accuracy of image segmentation,Deep Merge learns the similarity between image over-segmentation primitives using deep learning,merges the simi-lar primitives,and then iterates the above process to obtain a satisfactory image segmentation result.This is the first image segmentation method combining deep learning with region-merging.The study area is Phoenix city cluster covering 5,660 km~2from Google Earth,with a spatial resolution of 0.55m and 18.7 billion pixels.The experimental results show that Deep-Merge can segment objects in different sizes using only 0.2%of total segmentation primitives,and the optimal segmentation results have a stable scale parameter.Compared with the state-of-the-art methods,Deep Merge achieves the highest F value of 0.9446 and the lowest TE value of 0.0962 and ED2 value of 0.8989.(2)Proposed a general pruning framework Dynamic Pruning.To address the problem of low segmentation efficiency,I carefully analyze the tradi-tional bottom-up segmentation methods that rely on region adjacency graph(RAG)model,and found that the segmentation efficiency and accuracy are usually an irreconcilable contra-diction.The common solution is to add area and shape constraints.However,the segmenta-tion with these constraints is a compromised method between efficiency and accuracy.The RAG model has a large number of redundant edge weight updates in the process of merging nodes,resulting in low efficiency in large images.The method was tested in five methods and Deep Merge.The results show that the pruning framework can substantially improve the seg-mentation efficiency without performance dropping.In the multi-core mode test,all methods achieve super-linear speedups with a maximum improvement of 102.74.(3)Proposed an object-based weakly supervised land cover classification method City-Eye.It addressed the problems of difficult sample collection of large-scale urban land cover classification.Based on the results of the two previous methods,I spent less than 2 working days to label a small number of object samples.These object samples are then used to train a convolutional neural network model and generate more high-confidence pseudo-labels.Fi-nally,these pseudo-labels are used to train the Encoder-Decoder classification model.The method is tested on the Phoenix urban dataset.The test results show that a small sample la-beling time can obtain pseudo-labels with high confidence.The classification results reach an acceptable level among the Encoder-Decoder models.