TL;DR
This paper explores the aesthetic appeal of check dams, crucial for soil and water conservation, by combining human perception (Scenic Beauty Estimation - SBE) with AI-driven image analysis. They found that the proportion of vegetation and water, along with orderly arrangement, significantly influences aesthetic preference, offering valuable insights for designing more visually harmonious conservation structures.
What Problem Does This Paper Address?
Check dams are essential engineering structures for mitigating natural hazards like debris flows and landslides in regions with steep topography and heavy rainfall, such as Taiwan. While their primary function is safety and sediment control, there's a growing need to integrate these structures harmoniously with the surrounding landscape. Current technical regulations and design manuals often overlook the aesthetic dimension, leading to a gap in understanding how to design these vital facilities to be not only functional but also visually appealing. This paper aims to bridge this gap by quantifying the aesthetic preferences for check dams and identifying key landscape features that contribute to their scenic beauty.
Methods
The study employed a two-pronged approach:
- Questionnaire Survey and Scenic Beauty Estimation (SBE): A questionnaire was distributed to collect data on respondents' perceptions and preferences for various soil and water conservation facilities, specifically focusing on check dams. The survey evaluated landscape elements (natural vs. artificial), characteristics (color, shape, arrangement), and emotional responses. These subjective evaluations were then used to calculate SBE values, a quantitative measure of scenic beauty, using the RMRATE software.
- AI-driven Semantic Segmentation: To objectively analyze the visual composition of the landscapes, the researchers utilized AI-powered semantic segmentation. This technique classifies each pixel in an image into predefined categories such as vegetation, structure, sky, land, and water. Two deep learning models, Swin-L and DiNAT-L, were employed for this task. The proportion of each segmented element was then calculated for each photo.
By combining the subjective SBE scores with the objective proportions derived from semantic segmentation, the study aimed to uncover the relationships between visual landscape composition and perceived scenic beauty.
Datasets
The study utilized a dataset of 10 representative photos of check dams and other soil and water conservation facilities. These photos were selected for their diverse designs and performance.
For training the semantic segmentation models, the researchers leveraged the ADE20K dataset. This dataset contains 150 categories of objects in indoor and outdoor scenes and was used to identify landscape elements.
Key Insights and Ideas
- Vegetation and Water are Key: Landscapes with a higher proportion of vegetation (exceeding 30%) and water received more favorable aesthetic evaluations. This suggests that incorporating these natural elements is crucial for enhancing the visual appeal of check dams.
- Order Matters: Even with abundant vegetation, an unorganized or disorderly arrangement detracted from the aesthetic appeal. This highlights the importance of thoughtful landscape design and arrangement, not just the presence of natural elements.
- Land is Less Preferred: Conversely, areas with a higher proportion of land (soil surface) were less preferred, especially when the proportion of water was low.
- AI Models Show Promise: The DiNAT-L model demonstrated superior accuracy (80%) compared to the Swin-L model in identifying vegetation, structures, and land, indicating its potential for more robust landscape analysis.
- Blurring Boundaries: The study noted that modern designs often blur the lines between artificial and natural elements, making it harder to distinguish them, which can impact aesthetic perception over time without maintenance.
Key Results
- SBE Values: The study calculated SBE values for 10 photos, ranging from -135.66 to 144.94, indicating a wide spectrum of perceived scenic beauty.
- Proportion Analysis: Photos with the highest SBE values generally had over 30% vegetation. Conversely, photos with the lowest SBE values had a smaller proportion of water and over 40% land.
- AI Model Performance: The DiNAT-L model outperformed the Swin-L model in semantic segmentation accuracy for 80% of the photos.
- Qualitative Findings: Respondents preferred neutral or cool-toned colors and orderly arrangements, which contributed to a sense of calmness and were associated with higher SBE values. Positive emotional responses (tidiness, harmony, pleasure) were strongly correlated with higher preferences, while negative perceptions (disorder, depression) led to lower preferences.
Limitations and Future Work
- Limited Dataset Size: The study used a relatively small dataset of 10 photos for the questionnaire survey and analysis. This limits the generalizability of the findings.
- AI Model Limitations: While DiNAT-L showed better performance, both Swin-L and DiNAT-L models sometimes misidentified vegetation, structures, and land.
- Future Work: The authors suggest that future research could expand the dataset and explore more advanced AI models for semantic segmentation. Further investigation into the specific design elements that contribute to orderly vegetation arrangement and the long-term aesthetic impact of check dams with varying maintenance levels would also be beneficial. The findings provide a foundation for developing design guidelines that integrate aesthetic considerations into soil and water conservation engineering.
Summary updated