Abstract
Artificial intelligence (AI) integration into environmental analysis has revolutionized various fields, including the construction and application of gardens, by enabling precise classification and decision-making for sustainable practices. This paper presents a strong AI-driven framework that uses convolutional neural networks (CNN) and pretrained models like VGG16 and InceptionV3 to classify eight distinct environmental classes. The custom CNN achieved superior performance among the tested models, reaching an impressive 98% accuracy with optimized batch sizes. This demonstrates its effectiveness for precise environmental condition classification. This work highlights the crucial role of AI in advancing the construction and application of gardens, offering insights into optimizing garden design through accurate environmental data analysis. The diverse dataset used ensures the framework’s adaptability to real-world applications, making it a valuable resource for sustainable development and eco-friendly design strategies. This paper not only contributes to the field of AI-driven environmental analysis but also provides a foundation for future innovations in garden management and sustainability, paving the way for intelligent solutions in the evolving landscape of ecological design.
Introduction
Landscape architecture is currently transitioning from purely intuitive, experience-based methods toward data-driven ecological planning. While traditional artistic expression produces aesthetically pleasing spaces, it often overlooks complex environmental factors such as precise soil quality, water availability, and long-term ecological impact. As urbanization accelerates, gardens must serve as functional ecosystems that mitigate climate change and support biodiversity. This paper addresses the "intelligence gap" in site identification and simulation by proposing an AI-driven framework for automated land-use classification.
Methodology
The researchers developed a multi-stage pipeline for environmental classification. The process begins with preprocessing, where satellite images are resized to a uniform 224x224 resolution and normalized to ensure consistent pixel intensity, facilitating faster model convergence.
The study compared four distinct architectures:
- Artificial Neural Network (ANN): Serving as the baseline.
- VGG16 & InceptionV3: Established pre-trained models utilized via transfer learning.
- Custom CNN: A bespoke architecture featuring five max-pooling layers and optimized fully connected layers designed specifically for the spatial features of land-use imagery.
Dataset and Experimental Setup
The study utilized a curated dataset of 4,000 satellite images across eight categories relevant to garden planning: Agriculture, Beach, Dense Residential, Forest, Golf Course, Medium Residential, Parking Lot, and River. The dataset is perfectly balanced with 500 images per class, preventing algorithmic bias toward any specific land type. Hyperparameter tuning focused on batch sizes (4, 8, and 16) to identify the optimal balance between learning stability and generalization.
Key Results and Discussion
Quantitative analysis revealed a clear progression in model capability. The Custom CNN emerged as the superior model, achieving 98% accuracy with a batch size of 16 and a validation loss of only 0.01. This performance suggests that niche applications like landscape classification benefit significantly from models tailored to domain-specific spatial features rather than generic "off-the-shelf" architectures.
As shown in the comparative analysis, while transfer learning models like InceptionV3 (96%) and VGG16 (93%) performed robustly, they were outperformed by the optimized custom CNN. The custom model also achieved high reliability metrics, including a 0.97 Precision and a 0.95 F1-Score, making it suitable for real-world deployment.
Conclusion
This framework demonstrates that AI can serve as a powerful "simulation engine" for landscape architects, allowing for the simulation of sunlight exposure and plant compatibility before construction begins. While limitations such as fixed satellite resolution and temporal blindness (seasonal changes) remain, the study provides a foundation for future innovations. Future research will explore the integration of edge computing for real-time monitoring and the expansion of datasets to include more diverse geographical regions.
Summary updated