Machine learning detection of fog top over eastern Taiwan mountains from Himawari-8 satellite true-color images

Abstract

During the cold season, fog frequently blankets the eastern Taiwan mountains areas, a region known for the montane cloud forest (MCF). The mountain fogs play crucial roles in supporting local biodiversity. However, it remains challenging to identify the appropriate temporal and spatial variation of the mountain fogs due to limited ground observations in the complex terrain. To address this issue, this study combines machine learning and satellite remote sensing to provide continuous, terrain-unlimited temporal and spatial detection of fog occurrence over complex topography. From the nadir view of the satellite, the mountain fog top can form a clear edge closely following the topographical features to represent the maximum height of the fog-occurring area. This provides the opportunity to detect the top edges of mountain fog from satellite true-color images by applying the U-net machine learning technique. The training data consists of the three visible bands observed at 8 a.m. local time (LT) from the Himawari-8 satellite between the years 2019–2021, with the fog edges in the cases of fog occurrence serving as the training labels. The model performance is evaluated using true-color images at the same and different LTs. The result shows that the model can capture the climatology of the fog edge hotspots at specific elevations, with an accuracy of nearly 89% for predicting cases of fog occurrence at the same LT and over 70% for different LTs. These fog edge hotspots are highly consistent with the known locations of the MCF from past field surveys. The current results provide valuable guidance on identifying potential fog-prone ecosystems. Notably, the current model uses relatively simple input data, indicating the flexibility of the data source from other satellite observations and the potential application to other regions with topographical-related fog and low cloud.