Research On Land Use/cover Change Detection In Karst Areas Using Remote Sensing Images Based On Deep Learnin

With the development of aerospace technology and high precision aerial photography,earth observation technology has broken through the limitations of data collection and entered a new stage of comprehensive and in-depth application.Remote sensing image change detection based on deep learning is the main method to study the earth surface change.Change detection plays an important role in disaster monitoring,environmental protection and human settlement monitoring.At present,earth observation satellites continuously transmit data from space to the ground.In the face of massive data,we hope to find the changes of the objects of interest by exploring the surface changes of regions between two different time points,and obtain the changes or trends of the types of objects of interest.However,remote sensing image change detection is not only a simple technology.Satellite sensors,solar altitude and atmospheric conditions will play an important role in remote sensing image change detection.Areas with low ground economic value,such as the Karst geomorphology area of Guizhou Province,are rarely studied and solved due to the complex ground situation and limited data in such areas,which will not generate much economic benefit and have a long environmental recovery cycle.In recent years,deep learning algorithms gradually play an important role in the field of artificial intelligence.Remote sensing images are characterized by high dimension,magnanimity and diversity.Deep learning algorithm has great advantages in processing the data with these features,and can accurately extract the abstract features of complex objects,thus greatly improving the accuracy of remote sensing image interpretation.All kinds of deep learning methods have been widely applied in computer vision tasks.Similarly,deep learning theory can also be applied to remote sensing image change detection tasks.Therefore,it is urgent for us to carry out research on remote sensing image change detection based on deep learning.At the same time,in order to achieve the current more accurate remote sensing image change detection needs to open up a more intelligent application method.In this study,the change detection method of remote sensing image based on deep learning is divided into small sample change detection method and multi-class change detection method according to the key problems and detection difficulties.The main work includes the following aspects:(1)A change detection method based on semantic graph difference for small sample remote sensing image is proposed.Although there are many earth observation satellites at present,we are rarely able to obtain timely data of the region of interest due to the limitation of the orbit and the large ground area.The method based on semantic graph difference is derived from the difference graph method in conventional change detection methods.Combining the characteristics of difference graph method and deep transfer learning algorithm,this paper proposes a high-dimensional feature graph processing method integrating semantic graph and difference graph.In order to reduce the difference of feature distribution between data fields,In this paper,adaptive batch standardization is added to reduce the influence of domain feature distribution differences.Through experimental comparison and analysis of model fusion effect,it is found that the proposed method has strong adaptive ability of data domain and can accurately identify ground changes.(2)A multi-class change detection method based on class rebalancing in remote sensing images is proposed.Single type change detection can express many changes on the ground,and has good applicability for short-term disaster monitoring,but limited applicability for long-term ground monitoring.Multi-class change detection can express more ground changes,but it is difficult to detect multi-class change due to the unbalanced phenomenon of conventional data sets.In this paper,the class rebalancing algorithm is used to train the model and fill the data set at the same time to improve the class balance and reduce the sensitivity of the model.Through specific experimental analysis and comparison of model accuracy,it is verified that the proposed method can effectively reduce the sensitivity of the model to data class imbalance,and the proposed method can effectively improve multi-class accuracy.(3)In view of the above two methods,this paper carried out an empirical analysis on the remote sensing image of Wangmo County,a typical karst landform area in Guizhou Province,and the experimental results show that the proposed method can accurately identify the single type of land surface changes,and can identify the multiple types of land surface changes.