Takumi Wada


Study on Estimating Snow-Covered Areas Beneath Cloud Cover in Mountainous Regions Using High-Frequency Optical Satellite Imagery


Kazuyoshi Takahashi

Snow distribution in mountainous regions during the snowmelt season must be monitored continuously for both snow-disaster mitigation and water resource management. However, optical satellite imagery is strongly affected by cloud cover, and cloud-obscured areas frequently become unobservable. As a result, it remains difficult to capture the spatial distribution of snow cover continuously over wide areas.
This study aimed to develop a method for estimating snow-covered areas beneath cloud cover during the snowmelt season using a snow history map derived from high-frequency Sentinel-2 optical satellite imagery, and to clarify the estimation characteristics and applicability of the method. The study area was a mountainous region around the border between Niigata and Gunma Prefectures, Japan. Sentinel-2 data acquired during the snowmelt season were analyzed, and suitable scenes were selected by excluding clouds and cloud shadows with additional visual interpretation.
The snow history map was generated by classifying each pixel into snow, mixed, or vegetation classes for each observation date and accumulating the classification results pixel by pixel. Multiple normalized difference indices, mainly the Normalized Difference Snow Index (NDSI), were used for the classification. For estimation beneath clouds, both single-cloud and multiple-cloud conditions were examined. A ring region was defined around each cloud, and the threshold value best_z was determined based on the consistency between the snow history score and the land-cover classification image.
The results showed that best_z remained approximately 30 under both single-cloud and multiple-cloud conditions and exhibited little variation even when the ring width changed from 1 to 100 pixels. Although no clear improvement in the Kappa coefficient was observed with increasing ring width, overall accuracy remained relatively high, approximately 0.92 to 0.94, under all conditions. These results indicate that the proposed method is effective for complementing missing observations caused by cloud cover and for estimating snow-covered areas during the snowmelt season.