Detection Method Of Urban Land Cover Change Based On Combining Domain Knowledge And Deep Learning

Land cover change detection,as an effective measure of land resource use monitoring,plays an important role in rapidly developing cities.The process of urbanization has accelerated,infrastructure construction has been increasingly perfected,land resource utilization has been changing all the time,land values have risen,and there has been a large amount of illegal use of land.The contradiction between the continuous tightening of the national land policy and the unabated demand for land used by local governments has been continuously deepened.The government needs to establish a sound land resource management system,plan and monitor it to reduce the occurrence of illegal land use and maintain social harmony and stability.The increasing spatio-temporal resolution of remote sensing satellite images has reached a sub-meter level so that the use of remote sensing satellite imagery to detect changes in surface coverage has become a major trend,which means using different phases of remote sensing satellite images to compare whether the types of land cover in the same region are different.There are two main purposes for the detection of changes in two time phases images:detecting changes in surface coverage(extracting areas of change),determining categories before and after changes in surface coverage(extracting the change types of the change area).At present,for the extraction of surface cover type variation patches,operators determine the spatial location of land cover change through visual interpretation of time-phased high-resolution satellite remote sensing images based on their own subjective empirical knowledge.Therefore,the correctness of the changed patches depends on the visual interpretation experience of the operators,and it is prone to errors.The general detection process can obtain good detection results for specific remote sensing image data.However,in the face of large-area,multi-dimension,and high-resolution urban remote sensing images,this detection method seems a bit stretched and the application effect is not good.In this paper,the urban fringe development zone of Shenzhen is taken as the research area,we study the integrated detection framework of urban land cover change according to the uncertainty of urban high spatial resolution image data.The main tasks are:(1)Integrate current mature and stable detection algorithms,Optimizing the traditional detection methods from the inspection process,Combining the field knowledge,overcome the problem that traditional detection methods do not consider the uncertainty of image data and produce unsatisfactory detection results.Design the urban land cover change.detection framework.(2)According to the detection framework proposed in this paper,extracting geography knowledge and landscape morphology knowledge from the data related to land.Combining with expert experience,this paper have a research on the methods for extracting detection knowledge related to urban land cover change are studied.Developmenting a visual tool used by experts to extract knowledge.Achieving the aims of quickly obtaining and updating knowledge in this area.(3)In view of the small number of image samples and inconspicuous features,data augmentation is performed on the samples,according to the temporal characteristics of the image samples,the training samples and verification samples are divided.And in combination with transfer learning,several classical deep convolutional neural networks in the field of natural images classification are selected and their pre-trained models are used.The sample selected in this study area fine-tuned the pre-trained model,combined with the model's convergence and accuracy,and analyzed the feasibility of deep convolutional neural network in the classification of ground cover.(4)Based on the change detection framework that takes into account the uncertainty of the image data,the domain knowledge and deep learning are integrated into the detection process.And the land cover change detection system is developed.The system is successfully applied to the main area of Guangming Street in Shenzhen.It have improved the detection efficiency in accuracy and degree of automation,which have satisfied the needs of engineering application.