Abstract: In the face of the construction of natural resource remote sensing monitoring system and the annual update demand for key elements of provincial basic surveying and mapping,a method for constructing a provincial remote sensing interpretation sample database is proposed based on geographical and national surface coverage as the main reference data. Firstly,a sample classification system for remote sensing interpretation of natural resources in Gansu province suitable for deep learning and sample selection standards that take into account geoscience knowledge are developed. Based on this,three types of sample datasets including full element,single element,and change detection are constructed,including different scale units. Then,based on the self-developed Gansu province remote sensing interpretation sample library platform,a full link technical system from sample collection,model training,intelligent interpretation to quality assessment is constructed. The experimental results show that for different scale regions,the accuracy of feature extraction reaches 90%,and the accuracy of change detection reaches 74%,achieving adaptive interpretation of surface elements from coarse to fine granularity. The research results have been applied in the provincial basic surveying and mapping update,urban land space monitoring,non-agricultural monitoring and other fields in Gansu province,improving the accuracy and intelligence of natural resource management.

Key words: intelligent interpretation, change detection, platform architecture, deep learning, sample database