Vegetation change detection using artificial neural networks with ancillary data in Xishuangbanna,Yunnan Province,China

Timely and accurate change detection of the Earth’s surface features provides the foundation for better planning, management and environmental studies. In this study ANN change detection was used to perform vegetation change detection, and was compared with post-classification method. Before the post-classification was performed the ANN classification was used to yield multitemporal vegetation maps. ANN were also used to perform a one-pass classification for the images in 2003 and 2004. DEM and slope were used as two extra channels. During the training stage, the training data was separated into 82 subclasses including 36 change subclasses and 46 no change subclasses. Moreover NDVI differencing methods were used to develop the change mask. The result showed that combining the NDVI differencing method with visual interpretation when identifying reference areas can produce more accurate change detection results for the ANN one pass change classification. Moreover, it is effective to use elevation and slope as extra channels together with PCA components, to perform ANN-based change detection in mountainous study areas. It is also important to separate the vegetation transition classes into subclasses based on spectral response patterns, especially for mountainous terrains. This processing can reduce the topographic effect and improve the change detection accuracy.