Reflections on remote sensing image interpretation

Remote sensing images are different from ordinary natural images , The interpretation of remote sensing images also has its unique features . When analyzing remote sensing images , The interpretation method is required not only to find the appearance of the object , We should also explore the internal relationship of features , Establish the law of change of objects . Although the manual interpretation method can realize simple interpretation tasks , But manual interpretation depends on a wealth of experience , And the interpretation efficiency is low , It is very disadvantageous to the processing of massive remote sensing images , It is not used for monitoring tasks with high timeliness requirements , Such as disaster warning 、 Emergency response 、 Traffic management, etc .
Remote sensing images can reflect various basic attributes of ground objects , Like shape 、 size 、 Color 、 texture 、 Location 、 Scope, etc . Shape usually reflects the basic form of an object , For objects with obvious shape features , Only using shape information can we identify the category of objects , For example, airplanes 、 ship 、 Buildings, etc . The size of the object is closely related to the resolution of the image , When the object size on the remote sensing image does not match the real scale , It is not easy to judge such objects . Color information is the most intuitive attribute of an object , The color information can be used to judge the color in remote sensing images “ green ” Color vegetation and “ blue ” Color ocean . However , For some objects with similar colors , Such as grassland and forest , It is difficult to distinguish the category of objects only depending on color information . in fact , The texture information reflects the frequency of color change of the object , It is a concentrated reflection of various basic attributes of an object , It is easy to distinguish by texture information “ smooth ” Grass and “ Rough ” In the forest . The position of an object reflects its coordinates in the graph , It also corresponds to the geographical coordinates of the object , For example, ships usually only appear in the water , So when judging this kind of goal , We should pay attention to its possible geographical range . besides , There are also many factors that affect the results of remote sensing image interpretation , for example , Illumination conditions for remote sensing image acquisition , Shadows present in the image 、 Cloud and fog block the ground .
It's important to point out that , Remote sensing image interpretation is a process of comprehensive utilization of various features , It is difficult to achieve the purpose of accurate interpretation simply by relying on a certain attribute . The internal relationship between attributes is an important aspect of various remote sensing image interpretation methods . In recent years , As an effective natural image processing method , Deep learning and deep neural network can realize the classification of natural images 、 Recognition and segmentation . The implementation process depends on the semantic information at all levels of the image , Not just by the basic properties of the object . in fact , The purpose of remote sensing image interpretation is to obtain the semantic information of objects . Combine deep learning with remote sensing image interpretation , The weight parameters of the neural network are adjusted by supervised learning , It can fit the nonlinear equation that reflects the internal relationship of various attributes . Deep learning method dimension automation 、 Intelligent translation provides a feasible technical route , It is conducive to the further development of remote sensing image interpretation technology .
notes ： The article is excerpted from 《 High resolution remote sensing image processing and application based on deep neural network technology 》 Zhang Qiang waits