PROJECT TITLE :

Spectral and Spatial Proximity-Based Manifold Alignment for Multitemporal Hyperspectral Image Classification

ABSTRACT:

Multitemporal hyperspectral images provide valuable info for a wide selection of applications related to supervised classification, including long-term environmental monitoring and land cowl modification detection. But, the desired ground reference information are time-consuming and expensive to accumulate, motivating researchers to analyze choices for reusing limited coaching data for classification of alternative temporal images. Current studies that address high dimensionality and nonstationarity inherent in temporal hyperspectral data for classification are restricted for the case where vital spectral drift exists between pictures. During this paper, we have a tendency to adapt and extend 2 manifold alignment (MA) strategies for classification of multitemporal hyperspectral pictures in an exceedingly common manifold area, assuming that the local geometries of two temporal spectral images are similar. The first methodology exploits a domestically based manifold configuration of a source image (thought of to be the “previous” manifold), and the second approach links local manifolds of 2 pictures using bridging pairs. Besides exploiting manifolds estimated with spectral data for MA, we have a tendency to also demonstrate how spatial info will be incorporated into the MA strategies. When evaluated using 3 Hyperion information sets, the proposed ways outperform four baseline approaches and two state-of-the-art domain adaptation strategies. The advantages of the proposed MA strategies are additional evident when vital spectral drift exists between 2 temporal pictures. Plus the promising classification results, the proposed strategies establish a site adaptation framework for analysis of temporal hyperspectral information based on information geometry.