On many land cover maps using optical satellite images, parts of mountain shadows (shadows caused by terrain) are often misclassified as water areas. This is because a target in a shadow looks darker and bluer, due to illumination by scattered sunlight only. In the present study, we developed a land cover classification method that is hardly affected by mountain shadows, using multi-temporal optical satellite images. Experiments were carried out using ALOS/AVNIR-2 images (after atmospheric correction, orthorectification correction, and slope correction; years 2006 to 2011; 63 time periods), from 36°N to 37°N and 140°E to 141°E (1 degree×1 degree). First, for each image at each time period, the likelihood of each land cover category was estimated by kernel density estimation (KDE). Next, the positions of mountain shadows were estimated from the elevation data (AW3D DSM) for each period. The parts of the mountain shadows were then adjusted to reduce differences in likelihood among categories. The contributions of mountain shadows were thus decreased and the contributions of non-shadowed areas increased. As a result, most erroneous classification, which seems to arise due to the influence of mountain shadows, disappeared, particularly in the case of fake waters appearing on the north face of the mountains on the land cover map created by integrating the likelihoods of all the times.
