第七课 遥控感知图像.pdfVIP

IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 51, NO. 1, JANUARY 2013 329 Graph Matching for Adaptation in Remote Sensing Devis Tuia, Member, IEEE , Jordi Muñoz-Marí, Member, IEEE , Luis Gómez-Chova, Member, IEEE , and Jesus Malo, Member, IEEE Abstract—We present an adaptation algorithm focused on the acquired and similar images can lead to catastrophic results: description of the data changes under different acquisition con- even if the objects represented in the images are roughly the ditions. When considering a source and a destination domain, same, differences in reflectance, illumination, and atmospheric the adaptation is carried out by transforming one data set to the other using an appropriate nonlinear deformation. The eventually conditions may induce local changes in the probability distri- nonlinear transform is based on vector quantization and graph bution function (PDF) instead of simple global linear changes. matching. The transfer learning mapping is defined in an unsuper- Therefore, finding an efficient PDF matching strategy becomes vised manner. Once this mapping has been defined, the samples an urgent and complex task for new generation remote sensing in one domain are projected onto the other, thus allowing the processing chains. application of any classifier or regressor in the transformed do- main. Experiments on challenging remote sensing scenarios, such In machine learning literature, the problem of model adap- as multitemporal very high resolution image classification and tation is usually referred to as domain adaptation or transfer angular effects compensation, show the validity