Quantification of the spatiotemporal dynamics of diurnal fog and low stratus occurrence in subtropical montane cloud forests using Himawari-8 imagery and topographic attributes

• Published in: International journal of applied earth observation and geoinformation 2024-11, Vol.134, p.104212, Article 104212
• Main Authors: Chong, Jie-Yun, Lo, Min-Hui, Huang, Cho-ying
• Format: Article
• Language: English
• Subjects: Machine learning; Montane cloud forest; Random forests generator; Taiwan; Time-lapse photography; Topography
• ISSN: 1569-8432
• DOI: 10.1016/j.jag.2024.104212

•Machine learning modeled dFLS using satellite meteorological and topographic data.•Model performances were validated using ground time-lapse photographs.•Models with or without considering weather conditions were effective to detect dFLS.•The NDVI was the most significant variable in modeling dFLS.•It is feasible to model dFLS using machine learning solely with satellite imagery. Montane cloud forests (MCFs) feature frequent, wind-driven cloud bands (fog and low stratus [FLS]), providing crucial moisture to the ecosystems. Elevated temperatures may displace FLS, impacting MCFs significantly. To evaluate the consequences, quantifying FLS occurrences is vital. In this study, we employed “RANdom forest GEneRator” (Ranger), an advanced machine learning algorithm, to detect diurnal (07:00–17:00) FLS (dFLS) occurrence from 2018 to 2021 in MCFs in northeast Taiwan using 31 variables, including the visible and infrared bands of the Advanced Himawari Imager onboard Himawari-8, pixel solar azimuth and zenith angles, band differences, the Normalized Difference Vegetation Index (NDVI) and topographic attributes. We applied simple (lumping all data) and three-mode (sunrise/sunset, cloudy and clear sky) models to predict dFLS occurrence. We randomly selected 80 % of the data for model development and the rest for validation by referring to four ground dFLS observation stations across an elevation range of 1151–1811 m a.s.l with 53,358 diurnal time-lapse photographs. We found that it was possible to detect dFLS occurrence in MCFs using both simple and three-mode models regardless of the weather conditions (F1 ≥ 0.864, accuracy ≥ 0.905 and the Matthews correlation coefficient ≥ 0.786); the performance of the simple model was slightly better. The NDVI was more important than other variables in both models. This study demonstrates that Ranger may be able to detect dFLS in MCFs solely using a comprehensive array of satellite features insensitive to varying atmospheric conditions and terrain effects, permitting systematic monitoring of dFLS over vast regions.