運用機器學習整合衛星資料量化亞熱帶日間山地雲霧森林之雲霧時空分佈

Abstract

山地霧林廣義指被雲霧頻繁籠罩的山地森林，通常為濕暖空氣受到地形抬升作用形成雲霧。獨特的水文氣候特徵為森林提供足夠水源和養分，穩定高濕度環境減少太陽輻射透射量和森林蒸發散進而增加林相的多樣性。近期觀測研究表明氣溫的上升可能會影響雲霧帶的時空分佈，這將會對山地霧林的分佈造成威脅。為瞭解其分佈的變化機制和評估其帶來的影響，光學遙測技術視爲量化雲霧的重要觀測依據。然而，先前的研究受天氣影響限制了多雲區域的評估可行性。本研究針對不同的天氣特徵的山地霧林，提出以利用日本氣象衛星向日葵八號連續時間高密度之可見光、近紅外光、紅外光和太陽幾何角（高度角和天頂角）的觀測資料，搭配雙頻道亮溫差異、植被指數和地表特徵變數，如海拔、坡度、地形位置指數、地表粗糙指數和地形湿度指数（共31變數）作特徵萃取，結合監督式機器學習模型以建立日間（7時至17時）霧和低雲(FLS)的偵測算法。研究樣區為植被同質性較高的亞熱帶山地霧林-台灣棲蘭山區，於海拔1151-1810公尺之間安裝四個縮時攝像作爲現地雲霧觀測依據，以二元方式分類FLS事件。本研究在2018-2021期間共蒐集了53,358張影像，利用其作爲訓練數據建構三種不同天氣特徵（黃昏/黎明，晴空，多雲）之隨機森林（Ranger）模型,進行獨立測試結果，並以F1分數作爲衡量二分類模型精確度指標。結果顯示晴空條件下F1分數為0.864，黃昏/黎明和多雲條件下皆爲0.945，這顯示該模型可在不同天氣條件下穩定地偵測山地霧林之FLS事件。這一發現有助於系統性量化山地霧林FLS事件之時空變化。

Montane cloud forests (MCFs) are characterised by the presence of persistent, frequent wind-driven, horizontal belt shaped and orographic clouds also known as a cloud band. The hydroclimatic characteristics of MCFs often acts as a water supply to ecosystems due to the prevailing perhumid and dim light, where moisture introduced by depositing cloud can be more efficiently retained. Recent observation has shown that elevated temperatures may lift cloud band, which would cause colossal impacts on MCFs. The first step to assess the potential ramification is to quantify the occurrences of fog and low stratus (FLS) regionally; satellite remote sensing is an ideal tool for the task. However, previous research efforts may only be effective in limited weather condition, confining the feasibility in this cloudy region. In this study, we developed an algorithm to detect diurnal (defined as 07:00-17:00 in this study) FLS occurrence that was insensitive to weather conditions. We used the visible and infrared bands of the Advanced Himawari Imager on board Himawari-8 (H-8), sun geometry (solar zenith [SZA] and azimuth [SAA]) angles of each pixel, dual band differences (DBD), normalised difference vegetation index (NDVI) and local topographic variables (elevation, slope, topographic position index, vector ruggedness measure, topographic wetness index) as input data (31 variables) to model the FLS using “RANdom forest GEneRator” (Ranger), a recently developed machine learning approach derived from random forest). We carried out the study in subtropical MCFs of Chi-Lan Mountain in northeast Taiwan. We installed four ground FLS observation stations across an elevation range of 1151-1811 m a.s.l with 53,358 diurnal time-lapse photographs from 2018-2021 with or without FLS identified by visual assessment. We applied three different model (twilight/dawn, cloudy, clear sky) settings to model FLS occurrence because of the responses of the various bands depending on the time of day and the underlying surface characteristics. We randomly selected 80% of the data for Ranger development and the rest of data for validation. We found that it was possible to detect FLS occurrence in MCFs regardless of the weather conditions using the proposed method with the overall F1-scores  0.864, and 0.945 and 0.945 for clear-sky, twilight/dawn and cloudy conditions. The finding may facilitate systematic mapping of FLS occurrence in MCFs.