Extraction Of Farmland Rocky Desertification Base On Deep Learning And Its Landscape Ecological Risk Assessment

Cultivated land is the basic guarantee of food security and agricultural production.Accurately obtaining the spatial distribution of cultivated land resources and its ecological risks plays an important role in promoting the construction of cultivated land spatial quality,monitoring and production of agricultural resources,improving the quality of cultivated land function,protecting cultivated land environment and developing modern agriculture.At present,the only way to obtain large-scale land surface information quickly is satellite remote sensing technology,which provides an effective way for monitoring and extracting large-scale cultivated land information.However,the traditional remote sensing technology has some disadvantages in the extraction of surface features in karst mountainous areas.The reason is that it is difficult to obtain high-quality remote sensing images in mountainous areas with complex and rainy surface,and there are some differences in the visual features,spectral features and texture features of surface features in remote sensing images,resulting in a large number of phenomena of "same object with different spectra","same spectral foreign bodies" and mixed pixels There is.Therefore,in the karst plateau mountainous area with complex surface environment,broken cultivated land resources and strong spatial heterogeneity,it is difficult to extract cultivated land boundary information quickly and automatically using traditional remote sensing technology.It is urgent to use new methods and technologies to improve the level of rapid and accurate extraction of cultivated land information.In recent years,the advanced learning technology which has been gradually emerging in the fields of speech recognition,image classification,feature extraction,target detection and artificial intelligence has been applied to the extraction of large-scale remote sensing image and ground objects recognition,and has been gradually studied and promoted.The basic idea is to let the computer learn the rich features of images through continuous iteration,and simulate the human visual perception mechanism to extract the actual figure.At present,the deep learning model of image and figure extraction studied by scholars has a good advantage in accuracy and calculation speed compared with traditional remote sensing image classification methods.However,due to the poor processing effect on the features of diversity of terrain features and high spatial heterogeneity in the actual geographical environment,the accuracy of the extraction of the image is lower than the result of manual visual interpretation.In conclusion,the experimental area of this paper is selected as a typical karst plateau mountain area in Guizhou Province.Considering the heterogeneity of geographical space in the experimental area,under the guidance of the idea of geographical zoning,the high-resolution remote sensing image features and deep learning technology are highly integrated,and a new method of rapid and accurate extraction of cultivated land form information in layered areas is proposed.Based on the extraction of cultivated land,the average value of stone desertification classification index is calculated by dividing the boundary of cultivated land as the constraint condition,and the identification map of rocky desertification cultivated land is formed by associating it with the attribute table of cultivated map spot,and the classification discrimination of rocky desertification cultivated land is realized by "map coordination".Finally,the ecological risk difference of different types of cultivated land landscape in southwest mountainous area with complex surface was evaluated by different risk sources,and the development of cultivated land resources under different levels of rocky desertification environment was analyzed in depth.The following research results have been achieved:(1)The overall classification accuracy of the test results is 92%,the lowest and the highest GIOU intersection ratio is 0.81 and 0.96 respectively,and the GIOU intersection ratio of 75% of the validation samples is greater than 0.9,which indicates that the extraction results of the proposed method can maintain good integrity and fit the actual farmland boundary well,and can effectively avoid the "same spectrum foreign body" and "same object different spectrum" in karst mountainous areas It is proved that the method has good adaptability in the extraction of cultivated land in plateau mountainous area with high spatial heterogeneity.(2)On the basis of accurate extraction of cultivated land morphological information,it is deeply integrated with rocky desertification classification index in karst area to form a collaborative database of cultivated land atlas in rocky desertification area,and build a "zoning and hierarchical classification" Rocky Desertification cultivated land discrimination system.In order to provide important basic data for risk assessment,trend monitoring,effect assessment and policy design of rocky desertification cultivated land in karst mountain area,this paper studies the extraction and classification of rocky desertification cultivated land in karst mountain area.(3)Based on the "classification by zones" of rocky desertification cultivated land,the paper comprehensively considers the impact of the interference factors of farmers,soil pollution stress factors,road stress factors,water stress factors and terrain stress factors on the landscape ecology of cultivated land,and explores the ecological risk differences of different types of cultivated land in Guanling County,and accurately distinguish the spatial distribution characteristics of landscape ecological risks of different grades of rocky desertification cultivated land.Then,the development of cultivated land under different rocky desertification environment is studied,which provides a strong support for the scientific agricultural planning and government decision-making in Karst mountainous areas.