Analysis of influence factors and strategies of long-time series remote sensing image classification using transfer learning

KANG Yi;MA Yong;YAO Wutao;LONG An;REN Yingfeng;ZENG Yi;School of Information, Beijing Forestry University;Aerospace Information Research Institute,Chinese Academy of Sciences;Sanya Remote Sensing Research Institute, Chinese Academy of Sciences, Hainan Key Laboratory of Earth Observation;Environmental Emergency and Accident Investigation Center of Guangxi Zhuang Autonomous Region;State Owned Forest Farm Named Yugong of Jiyuan City;

long-time series data;;land use;;transfer learning;;remote sensing image

Classification of long-time series remote sensing images is an important foundation for the study of spatial and temporal changes in regional natural resources and land uses. There are some problems in the traditional classification of regional long-time series remote sensing images, such as low reusability of samples, much manual annotation work, etc. As a machine learning method that applies existing knowledge to different tasks, transfer learning can meet the need of sample reuse in remote sensing images. However, in the investigation and monitoring for specific regional changes in a long-time span and large spatial range, the phenological changes caused by seasonal differences have an impact on sample reuse. At present, there is a lack of relevant research on the scope and influencing factors of sample feature migration. In this study, Jiyuan City, Henan Province, China, was selected to be the research area, and Landsat long-time series remote sensing images were used to realize the migration experiments. Through the design of transfer experiment, we explored the influence of different factors, the influence degree and application scope of phenology difference, time span, and other factors in the process of transfer learning. Support vector machine(SVM) and random forest models were developed to apply the samples of single remote sensing image to the 20-year span of long-time series images for the classification based on the method of direct transfer learning. According to the results of migration experiments under different phenology and time span conditions, the influence of two factors on the migration experiment was analyzed in order to improve the reusability of samples and reduce manual work for the long time series remote sensing image classification. Among them, phenology is the most important factor, followed by the time span. In different land use types, the migration effect of non-vegetation sample plot was better and more stable, followed by the vegetation sample plot. The results of SVM model and random forest model were similar and had high reliability. In this experiment, the effect of SVM model was better. Finally, the strategies and methods were summarized to provide reference for using transfer learning to apply to the long-time series remote sensing images classification.