Construction Of Remote Sensing Image Sample Database For Deep Learning In Gansu Province

Fully automatic interpretation of remote sensing images is a major technical problem that China's land and resources science and technology committed to tackle.In recent years,with the rapid development of machine learning technology represented by deep learning,the application of deep learning to automatic interpretation of remote sensing images and the realization of land use type recognition as automated as possible is an important research goal and direction for Chinese researchers.Deep Learning Corresponding Depth Neural Network(DNN)works on the premise that the deep network is fully trained,which requires a large number of samples as training data.Traditionally,training images are acquired manually,labeled manually,time-consuming and laborious,with huge workload,and are vulnerable to the influence of operator's working mood and negligence;at the same time,the lack of source data and the difficulty of sample production restrict the extensive application of deep learning in remote sensing image interpretation.Based on the diversity and expansibility of crowd-source data OSM,current land use data,manual annotated vector data and remote sensing image data,this thesis used the Export Training Data for Deep Learning Tool in ArcGIS Pro to produce sample images and used Microsoft SQL Server Management Studio 2008 to construct sample database.Finally,the accuracy of the sample database was tested by using the in-depth learning model.The main contents and results of this study are as follows:(1)According to the basic requirements of in-depth learning for remote sensing image samples,this thesis put forward the principle of sample construction: a.There are abundant data in each category;b.Object category is the combination of the standard classification system of land use and land cover classification and OSM classification system in China,the level of each category is designed to further improve the diversity and comprehensiveness of samples;c.Land use has different types,which are established according to China's land classification standards.There are significant differences among different classes;d.The recognition rate of the whole system is high,which can avoid blurring objects and improve image quality;e.Each class has different imaging angles,sizes,shapes and colors to increase sample diversity,which can improve the generalizationperformance and robustness of the model.(2)Based on the existing OSM(Open Street Map)classification system,China's land use classification criteria and geographic survey classification system,a remote sensing image sample database classification system was constructed,including arable land,garden land,woodland,grassland,buildings(areas),structures,railways and roads,man-made dug land,desert and bare land,waters,geographical units 11 first class,28 second class and 25 third class.(3)The sample database has two sub-databases of 128*128 pixels and 256*256 pixels,respectively.Among them,128*128 sample database has 42 sub-categories of about 19500 images,with an average of about 490 images per class,and 256*256 sample database has37 sub-categories of about 16000 images,with an average of about 420 images per class.(4)Deep learning model was trained by using the constructed sample data,the training accuracy is about 99%.Based on the training model,the classification accuracy is98.9%,the result is encouraging.