Weakly supervised classification of time-series of very high resolution remote sensing images by transfer learning

ABSTRACT In multi-temporal remotely sensed data analysis, labeling samples from each image is often a required process, this is however very tedious and time-consuming. In many cases the ground objects do not change significantly through time, and one can reuse some of the labels with appropriate consistency verification. In this letter, a novel weakly supervised transfer learning framework is proposed to classify multi-temporal remote-sensing images with only one labeled image. By utilizing the consistency of time-series images and a domain adaptation method, our framework is able to classify all the other multi-temporal images chronologically without any labeling effort for these images. Our framework achieves a similar level of classification accuracy as if it were through supervised learning. Our framework is shown to be effective for processing multi-temporal remote-sensing images when training samples are only available for one temporal dataset.