Exploring the potential of generative AI to complement multi-stakeholder landscape preference assessment

Abstract

As urbanization and agricultural intensification continue to reshape rural landscapes, understanding and incorporating diverse stakeholder preferences has become crucial for sustainable land use and management. Traditional landscape preference assessments remain constrained by limited scalability, high cost, and time intensity, highlighting the potential of artificial intelligence to complement human evaluation. This study employs two multimodal large language models (MLLMs), GPT-4o and Qwen3, to simulate and analyze the landscape preferences of farmers, tourists, and experts in the Mulberry-Dyke and Fish-Pond agricultural landscape in China. Extreme gradient boosting and Shapley additive explanations were applied to examine discrepancies between MLLMs’ predictions and human judgments, and to examine how specific landscape characteristics shape stakeholder preferences. Furthermore, stakeholder-derived importance weights of landscape characteristics were incorporated into the prompts to improve model alignment with human perception. The results show that GPT-4o outperformed Qwen3 in predicting human preferences. While humans emphasized the dyke-pond ratio and fishpond shape, GPT-4o tended to prioritize built-environment features such as local buildings. Incorporating stakeholder evaluations into the prompting process substantially enhanced model-human correlation by approximately 38%, 85%, and 54% for farmers, tourists, and experts, respectively. These findings demonstrate that MLLMs can serve as adaptive tools for multi-stakeholder landscape preference evaluations, offering new opportunities to integrate diverse human perspectives into landscape planning and decision-making.