An artificial neural network approach for optical-SAR imagery fusion and landslide mapping

In the mountainous areas of Taiwan, affected by active tectonics, frequent typhoons and human activities, landslides are commonly induced by heavy rainfalls especially during typhoon seasons. During Typhoon Morakot (2009), the heavy rainfall induced large-scale landslide and caused severe damages in the mountainous region. This study focuses on landslide detection in Laonong River watershed in southern Taiwan. Image fusion can visually or statistically enhance the characteristics of land-objects. Usually, investigators mainly detect bare surface of landslides by using optical images instead of using synthetic aperture radar (SAR) images to identify the erosion, transportation and deposition patterns, which can be critical to landslide susceptibility assessment. Thus, this study aims to develop an Optical-SAR image fusion for an advanced landslide mapping task. In this study, SAR data is used to detect the change of land surface by distinguishing backscatters of images before and after the Typhoon event. In addition, the machine-learning method, artificial neural network (ANN), was operated for landslide pattern fusion and mapping practice. With applying image segmentation, Normalized Difference Sigma-naught Index (NDSI) from SAR images and Normalized Difference Vegetation Index difference (NDVIdiff) from optical images were generated, and calculated their texture statistics, such as mean, standard deviation, contrast, entropy, homogeneity and dissimilarity. Landslide detection results were assessed by overall accuracy (OA) and kappa coefficient, with comparing to a manually interpreted landslide inventory. Result shows the OA of optical-SAR image fusion is 0.896 and the kappa coefficient is 0.547, which has better performance than results with only using optical or SAR images for landslide detection.