Towards multi-temporal hyperspectral images classification based on 3D signature model and matching
Résumé
Multi-temporal data processing is being actively explored in the remote sensing community for robust land cover recognition and other similar applications. Such an approach exploits multiple, independent observations of a phenomenon and extract more detailed information. Thus, the recent advent of more sophisticated sensors leads us to perform a decision level for scene interpretation. In this paper, we propose a new 3D model characterizes all the pixels in a scene by considering their reflectance values as a function of time of imaging and spectral waveband. Then, the classification task is based on similarity distance of each pixel with the 3D spectral data base. The obtained distance is then used for classification task. Experiments reported in this paper show that proposed approach overcome the state-of-the-art methods problems. These experiments are based on synthetic data, studying the effect of noise absence of pure pixels, and on a real dataset. The proposed framework provided very promising recognition performance even in small sample size conditions.