運用影像統計量預測景觀偏好之研究

Abstract

當面對令人讚嘆的美景，無數的光粒子進入眼中，心情頓時感到愉悅，這時我們是如何解讀這些光粒子? 景觀偏好(Landscape Preference)即是探討這個問題的研究。 過去的研究回答這個問題有兩種主要的模式，心理物理模式(Psychophysics Paradigm )與認知模式(Cognition Paradigm)。心理物理模式著重於了解外界的環境物理性特徵對景觀偏好的影響，認知模式著重於觀賞者本身對於環境的認知如何影響景觀偏好。那麼是否可以直接從光粒子中的訊號，來了解景觀偏好的形成。因此本研究嘗試模擬大腦處理視覺訊號的過程，建構一個可以透過影像訊號來預測景觀偏好的計算模型。 視覺訊號由相位譜與能量譜構成，能量譜紀錄了這些光粒子的強度與整體結構，對應到實際的環境中是整體的環境質感與空間結構，此特性稱為空間包絡特徵(Spatial Envelope Property, SEP)是一種影像統計量，由於空間包絡特徵與影響景觀偏好的因素(例如開放度、複雜度)相當類似，因此本研究假設可使用影像中的空間包絡特徵構成計算模型以預測景觀偏好。 本研究收集了八種景觀類別(公路、高樓、街道、市區、海岸、森林、鄉村、高山)的影像共480張，運用主成分分析法由影像的能量譜中萃取了三個主要的空間包絡特徵，分別是空間質感、空間方向與空間深度，並依此架構計算模型。同時以網路問卷的方式進行影像景觀偏好調查。 為探討模型的預測力，以空間包絡特徵對景觀偏好進行多元回歸分析，區分為兩種模式，綜合類別分析是混合不同環境類型影像進行測試，單一類別分析是以單一環境類型影像進行測試。結果顯示，在綜合類別分析中，計算模型對景觀偏好的預測力是顯著的，然整體模式解釋力較低，在單一類別分析中，計算模型可以預測三種環境類別的景觀偏好，分別為公路、市區與森林，其餘五類預測效果不佳。影響市區景觀偏好的特徵是空間質感，影響公路與森林景觀偏好的特徵是空間方向與空間深度。 然而空間質感的變化為何會產生愉悅的感受呢?本研究推論其關係可能是建立在空間質感所具有的碎形結構上，由於大腦可以快速處理具有碎形結構的訊號，因此產生愉悅的感受。而空間方向、空間深度與景觀偏好的關係可能建立在眺匿理論上，空間開放度越高生物越能掌握環境對其生存威脅的情況，因此景觀偏好度提高。在應用面上，空間質感與植物元素的變化具有相關性，可運用於植被變化的計算而進行環境監測。亦可運用景觀知覺實驗，探討質感與景觀偏好的關係。 本研究是一個探索性的嘗試，試圖了解環境中光粒子中包含了何種讓我們感到愉悅的訊號統計特質，未來仍有待更多的研究，一起繪製更為完整的景觀偏好圖像。

When we see beautiful scenery, light particles enter our eyes and induce pleasant feelings. Landscape preference is the study of this phenomenon. Two paradigms are used to study landscape preference: the psychophysical paradigm and the cognitive paradigm. The psychophysical paradigm focuses on understanding the relationship between the physical attributes of the environment and landscape preference, while the cognitive paradigm focuses on understanding the relationship between the cognitive attributes of the environment and landscape preference. That being the case, what is the relationship between the light particles and landscape preference? The purpose of this research was to establish a landscape prediction model based on the visual signal computational process of the brain. Visual signals are composed of the phase spectrum and the power spectrum. The power spectrum represents the modulation of light intensity, which contains the global spatial structure of a scene. The property, known as spatial envelope property (SEP), is a type of image statistics. Because these properties are similar to factors that affect landscape preference, we therefore suppose these properties could predict landscape preferences and be used to construct a computational model of landscape preferences. The 480 images used in this research were composed of eight types of scenes including highways, tall buildings, streets, inner cities, coasts, forests, the countryside, and mountains. Principle component analysis was applied to extract the SEPs from the power spectrum of these images. Three components were extracted in this research: spatial texture, spatial direction, and spatial depth. The prediction model of landscape preference was constructed from these SEPs. We used multiple regression to test the prediction ability of the model. Two modes were used for testing: the mixed category mode, which tested all the images, and the single category mode, which tested the category of the image. The results show that the model could predict landscape preference in the mixed category mode but the R-square of the model was lower. In the single category mode, the computational model could predict landscape preference in the following three categories: highways, inner cities, and forests. Spatial texture was found to affect landscape preference in the inner city categories; whereas, spatial direction and spatial depth affect landscape preference in the highway and forest categories. However, why does spatial texture induce landscape preference? We suppose that the reason is the fractal structure of spatial texture. Because our brain can process signals composed of fractal structures easily, we feel pleasure when seeing such scenes. The relationship between spatial direction and spatial depth to landscape preference may be explained by prospect-refuge theory, because creatures could protect themselves from being eaten in an open scene. According to the relationship between spatial texture and the variation of vegetation, it could be used as the index for monitoring the environment. In addition, spatial texture could be used in a landscape perception experiment for examining the relationship between preference and distribution of spatial texture. This exploratory research aimed to understand the signal information contained in the environmental light particles, which induce our landscape preference. We found a small component of the answer but further research is required to construct the complete picture of landscape preference.