Kota Anaushi

Examination of the snow area automatic distinction technique using the satellite index image

Kazuyoshi Takahashi

Niigata is a world eminent heavy snowfall area, and disasters caused by the snow suffer serious damage. However, on the other hand, there are utilization to sightseeing, recreation and the demand as important aquatic resources. From such a snowy two-facedness, an estimate of the quantity of snow is necessary from not only preventive measures against snow damage but also the point of view called the snowy use.
Great human resources and economical resources are necessary that most of snow are distributed over the mountainous area and that a person goes to the mountainous area and observes the snow regionally. In addition, the danger such as a snowslide and the accident is accompanied. The study that utilized the remote sensing technology such as satellite images from such a scene is pushed forward. In the study of Nishihara, Tanize and others (2019), it is reported that an area ratio and the progress rate that did snow-melting from the snow area which I extracted from artificial satellite data of the melting snow almost agree, and it is thought that it is connected in it grasping the snow quantity to grasp the snow area.
The cloud has a big influence in the grasp of the snow area using the satellite image. Because the visible light level is similar to snow in the spectrum properties of the cloud, distinction is difficult. Therefore, in this study, it is intended that I examine automatic distinction technique of the snow area corresponding to plural artificial satellites using an index image.
Because pixel_qa band given the quality information every pixel was provided to Landsat8, and snow and clouds were classified, using satellite data of Landsat8, I made teacher data. The teacher data treat the image which composed three indexes of NDVI, NDSI, WhiteNess as teacher data. NDVI, NDSI, WhiteNess is the image which emphasized a plant, snow, a reflection of the whiteness level each.
In consideration of the influence that input size gave to learning precision and distinction precision, I built the distinction model of the form to input at 64×64 size on building a model to distinguish it. As a result, 98%, 96%, the learning precision of 78% were provided in the two level classification, the three level classification, the four level classification each.
Result Snow and Cloud which I gave the distinction model who built it unknown data in Landasat8 and distinguished were distinguished almost definitely, and distinction in Snow and Cloud was performed a result of the three level distinction more exactly. In addition, as a result of having performed an evaluation in the visual distinction, precision in the three level distinction was the highest. From this, it is thought that precision improves because I can learn detailed information by increasing labels to give, but it is thought that the number that is most suitable to the number of the labels exists.