The Research Of Land Use Classification Of Time-series Remote Sensing Images Based On Transfer Learning

Land cover and land use in China are changing rapidly with the development of economy and the acceleration of urbanization,which brings a series of problems,such as the decrease of cultivated land,the destruction of forest land,the depletion of water resources and so on.The demand for ecological protection is increasing.At the same time,with the vigorous promotion of China's ecological civilization construction,rational use and effective protection of limited land resources has become a topic of concern.The use and protection of land resources cannot do without the study of its ecological background and spatiotemporal changes.And the classification of time-series remote sensing images is an important basis for the study of regional natural resources and spatiotemporal changes of land use.The traditional classification of regional long time series remote sensing images mostly uses visual interpretation,and selects samples from each image for classification.The reusability of selected samples is low,and the manual work is complicated.As a machine learning method that can apply the existing knowledge to different tasks,transfer learning can realize the reuse of sample features and give full play to the value of manual selection of samples.And then the complicated manual work can be reduced.However,there is less systematic research on the reuse effect of sample feature in remote sensing images of long time span,different phenology and even different satellites,as well as the influencing factors in the migration process.This paper uses Landsat time-series images and sentinel-2a images,introduces the theory of transfer learning,and explores the application effect of sample features in long time span and different satellite images,so as to reduce the manual work in the process of long time-series land use classification.The research mainly includes:(1)Long time-series remote sensing images migration classification of the same satellite.Using Landsat long time series remote sensing image,based on the method of direct transfer learning and using SVM and random forest model,the sample of single scene image is applied to 20-year long time series image for classification.According to the results of migration experiments under different phenology and different time span,the influence range of phenology and time span is summarized,and the migration strategy is given.That is,when the sample is migrated and reused in the satellite long-time sequence images,it is necessary to ensure that the seasonal phase difference between the target image and the source image is within one month,the time span is short than ten years.And then,the accuracy of sample migration classification can achieve about 80%.(2)Images migration classification of the same satellite after consistency processing.For Landsat long time-series images,the pseudo invariant target method is used for relative radiometric correction to achieve the consistency of multi scene satellite images.The IR-MAD algorithm is used to select the invariant target,and the least squares method is used to fit the linear relationship between the images.Compared with the results of the migration experiment before processing,the consistency processing with satellite images can improve the classification accuracy by 1% ? 4%,and the overall accuracy can reach up to 85%.(3)Migration classification between different satellite images.consistency processing is processed to enhance the radiation consistency in order to realize the migration experiment between different satellite images.It is concluded that the data must be processed consistently when reuse sample and features in different satellite.The accuracy of classification can reach 78%.The Sentinel-2A image of 10 meter resolution can be used as a supplementary data source,and can increase the details of the classification results.Compared with the Landsat Image of 30 meter resolution,the Sentinel-2A image of10 meter resolution can extract construction land,cultivated land and other features more accurately.According to the above migration experiments and results,the influencing factors and scope are analyzed,and the corresponding migration classification strategy is obtained,which provides a reference for reducing manual intervention and improving the reuse rate of samples in long time-series images classification.