In recent years, land cover classification technology based on multispectral remote sensing images has been applied to many fields of environmental monitoring. The existing methods generally rely on the information of spectral bands. However, the spectral information of monotemporal multispectral images cannot be generalized to describe the spectral characteristics of the ground objects at different times. In addition, the time-series samples will slightly change over time, and it is difficult to maintain classification performance continuously. To address these problems, we propose a method of land cover classification for multispectral images based on the time-spectrum association feature and multikernel boosting incremental learning. Our method is conducted in two main stages. (1) We propose the time-spectrum association features to acquire the seasonal spectral characteristics different ground objects. (2) To design the classifiers, we propose multikernel boosting method and introduce a multikernel boosting classification learning, which uses continuous new samples to update the weights of the classifier by low-computational and small-scale incremental learning. We test the proposed method on a public multispectral dataset from Landsat-5. The experimental results show that the extracted time-spectrum association features can better characterize the differences of different ground objects, and the proposed classifier can reach more accurate classification with gradually increasing samples over time. |
ACCESS THE FULL ARTICLE
No SPIE Account? Create one
CITATIONS
Cited by 4 scholarly publications.
Multispectral imaging
Image classification
Remote sensing
Image segmentation
Data modeling
Environmental monitoring
Calibration