Training Sample Transfer Learning from Multi-temporal Remote Sensing Images for Dynamic and Intelligent Land Cover Classification

Abstract:

Reconstructing historical land cover dynamics by time series remote sensing images is the basis of geographic scene sensing and modeling. However， classification of historical archived remote sensing images is quite difficult due to the limited training samples and low interpretation capability for multi-temporal images. A spatio-temporal training sample transfer method from existing land cover products and corresponding remote sensing images was proposed based on geometric and attribute constrains. The transferred training samples were then embedded into a multi-temporal land cover classification framework to solve the problem of high sample labeling cost in multi-temporal image classification. The multi-temporal classification results in the Taihu Basin show that the proposed unsupervised sample transfer method can make full use of the prior geometric constraints and spectral properties of the products and images， and the classification accuracy in large scale area is over 89%. The results demonstrate that the proposed method is effective for land cover updating and geographic environment evolution modelling.