利用分群與語意分割對夜間雲圖與淨長波輻射進行量化研究

Abstract

雲在調節地球大氣能量收支中扮演著關鍵角色，然而，由於觀測技術的限制，不同雲種在夜間的輻射效應仍難以精確量化。本研究旨在開發一個結合地面觀測與機器學習的創新框架，以深入探討夜間雲的物理特性（如雲種、雲量）如何影響地表淨長波輻射。 本研究於台灣國立中央大學的觀測站點，同步收集了夜間雲影像、長波輻射量及溫溼度等相關氣象數據。並提出一個擷取雲影像特徵的混合式分析流程：首先，利用基於 MobileNetV3 的模型對高維度影像進行降維，萃取與長波輻射相關的關鍵特徵，並透過分群演算法歸納出雲與輻射之間的潛在關聯模式，取得影像的分群特徵。同時，開發了一個基於 U-Net 的語意分割模型，專門用於夜間雲種的辨識與雲量估算。該模型能將夜空分為晴空及七種雲型（卷雲、卷層雲、卷積雲、積雲、層雲、層積雲、雨雲），在低光照條件下達到了 71.7% 的加權交並比與 88.8% 的像素準確率，展現了其優異的辨識性能，提供了影像中雲種與雲覆蓋量特徵。 為解析雲特徵與輻射之間的關聯，本研究分別採用了線性與非線性模型進行分析。在線性模型部分，我們使用多元線性迴歸模型，分析由語意分割得出的雲種與雲量資訊。結果明確指出，低層厚雲對於淨長波輻射有最強的增溫效應，其中雨雲、層積雲與層雲的迴歸係數最高，分別達到 27.53、24.50 與 21.54 W/m2。此結果驗證了雲的物理屬性（光學厚度、雲底高度）是影響地表輻射的關鍵因子。在非線性模型部分，本研究進一步整合了從降維與分群模型中萃取出的分群特徵。研究發現，這些特徵能有效捕捉線性模型難以描述的複雜交互作用。將分群特徵、雲種資訊與雲覆蓋量共同輸入非線性模型後，模型的預測準確度獲得顯著提升，證明了分群資訊在捕捉雲與輻射之間非線性關係中的獨特價值，同時模型仍保有一定程度的解釋力。 本研究成功證明，結合地面夜間影像與機器學習技術，是量化雲形態並估算夜間淨長波輻射的可行方案。透過線性與非線性模型的綜合分析，本研究不僅驗證了已知的雲物理效應，更揭示了影像深層特徵在解釋複雜輻射交互作用中的潛力。此框架為建構持續、低成本的夜間大氣與地表輻射監測系統提供了一條嶄新的途徑。

Clouds play a critical role in regulating the Earth's atmospheric energy budget. However, the radiative effects of different cloud types under nighttime conditions remain challenging to quantify due to observational and measurement limitations. This study aims to develop an innovative framework that integrates ground-based observations with machine learning to investigate how the physical properties of nighttime clouds, such as cloud type and cover, influence surface net longwave (NLW) radiation. Nighttime imagery, longwave radiation measurements, and corresponding meteorological data (e.g., temperature, humidity) were synchronously collected at a station located at the National Central University in Taiwan. A hybrid analysis workflow was proposed to extract cloud features from the images. First, a MobileNetV3-based model was employed for dimensionality reduction on the high-dimensional image data to extract key features correlated with longwave radiation. A clustering algorithm was then applied to these features to identify latent patterns in the cloud-radiation relationship, yielding “clustering features.” Concurrently, a U-Net-based semantic segmentation model was developed for nighttime cloud type identification and cover estimation. This model classifies the night sky into clear sky and seven distinct cloud types (cirrus, cirrostratus, cirrocumulus, cumulus, stratus, stratocumulus, and nimbus), achieving a weighted Intersection over Union (wIoU) of 71.7% and a pixel accuracy of 88.8% under low-light conditions, which demonstrates its robust performance and provides “cloud type” and “cloud cover” features. To elucidate the relationship between cloud features and radiation, both linear and non-linear models were employed. In the linear analysis, a multiple regression model was applied using the cloud type and cover information derived from semantic segmentation. The results specifically indicate that low, thick clouds exert the strongest warming effect on NLW radiation, with nimbus, stratocumulus, and stratus clouds exhibiting the highest regression coefficients of 27.53, 24.50, and 21.54 W/m2, respectively. This finding validates that the physical properties of clouds, such as optical thickness and cloud base height, are critical factors influencing surface radiation. For the non-linear analysis, the clustering features extracted from the dimensionality reduction and clustering models were further integrated. It was found that these features effectively capture complex interactions that are not easily described by linear models. The prediction accuracy of the non-linear model was significantly enhanced upon the inclusion of the clustering features alongside cloud type and cover data, demonstrating the unique value of these deep features in capturing the non-linear dynamics of cloud-radiation interactions while maintaining a degree of interpretability. Overall, this research successfully demonstrates the feasibility of quantifying cloud morphology and estimating NLW radiation by combining ground-based imagery with machine learning techniques. Through a comprehensive analysis using both linear and non-linear models, this study not only validates known cloud physical effects but also reveals the potential of deep image features in explaining complex radiative interactions. This framework offers a novel and cost-effective pathway for continuous monitoring of the nighttime atmosphere and surface radiation budget.