探討景觀視覺特徴對美感體驗與認知的影響

Abstract

本研究探討人們對自然環境的景觀偏好，結合認知科學與深度學習模型分析景觀偏好矩陣神經基礎。研究比較了人們對於偏好矩陣的行為評分（包括一致性、易讀性、複雜性、神秘性和整體偏好）與大腦特定功能區域的活化，以及使用深度學習模型的特徵來預測偏好矩陣的相關指標。研究結果顯示，不同景觀偏好指標涉及不同的的腦神經機制，高階認知區域與一致性和易讀性相關，而複雜性和神秘感則與場景識別及記憶相關區域呈現較高的相關性。此外，整體偏好似乎是多層次視覺訊息的整合結果，無法單獨由AI模型進行預測。本研究初探了人類對於景觀偏好的神經機制，並對環境心理學、景觀設計及機器學習用於景觀偏好的預測提供實證支持及應用價值。

Landscape preference is shaped by a complex interplay of perceptual, cognitive, and affective processes. This study investigates the behavioral, neural, and computational mechanisms underlying landscape preference. By integrating human behavioral ratings, functional magnetic resonance imaging (fMRI) data, and deep learning-based scene analysis. Participants rated natural landscapes based on coherence, complexity, mystery, legibility, and overall preference while undergoing fMRI scanning. The BOLD activation of preference indicators and representational similarity (RSA) was applied to examine the cognitive correlation of landscape preference and alignment between brain activation patterns, human aesthetic judgments, and feature representations from a deep neural network. The results revealed distinct neural correlates for different aesthetic dimensions. Coherence and legibility engaged higher cognitive regions involved in spatial organization and decision-making, while complexity and mystery were associated with brain areas related to emotional processing and memory retrieval. The RSA findings indicated that mid-to-late layers of AlexNet captured human perceptions of complexity and legibility, but preference remained difficult to predict, suggesting that aesthetic appreciation extends beyond visual processing to include cognitive and emotional factors. These findings offer insights into the neural basis of landscape aesthetics and the extent to which AI models approximate human aesthetic judgments. While deep learning models capture certain visual features, they do not fully account for the subjective and affective dimensions of preference.