Evaluating the environmental perception of urban parks is highly significant for optimizing urban planning. To address the limitations of traditional evaluation methods, a multimodal deep learning framework that integrates pre-training and reinforcement learning strategies for the comprehensive assessment of various park types (seaside, urban, mountain, and wetland) across three dimensions—accessibility, usability, and aesthetics—is proposed herein. By combining image data and user review texts, a unified architecture is constructed, including a text encoder, image visual encoder, and multimodal fusion module. During the pre-training phase, the model captured latent features in images and texts through a self-supervised learning strategy. In the subsequent training phase, a reinforcement learning strategy was introduced to optimize the sample selection and modal fusion paths to enhance the model’s generalization capability. To validate the cross-type prediction ability of the model, the experimental design uses data from three types of parks for training, with the remaining type as a test set. Results demonstrate that the proposed method outperforms LSTM and CNN architectures across accuracy, precision, recall, and F1 Score metrics. Compared with CNN, the proposed method improves accuracy by 5.1% and F1 Score by 6.6%. Further analysis shows that pre-training enhances the robust fusion of visual and textual features, while reinforcement learning optimizes the sample selection and feature fusion strategies during training.
Parks are an important component of urban ecosystems, yet traditional research often relies on single-modal data, such as text or images alone, making it difficult to comprehensively and accurately capture the complex emotional experiences of visitors and their relationships with the environment. This study proposes a park perception and understanding model based on multimodal text–image data and a bidirectional attention mechanism. By integrating text and image data, the model incorporates a bidirectional encoder representations from transformers (BERT)-based text feature extraction module, a Swin Transformer-based image feature extraction module, and a bidirectional cross-attention fusion module, enabling a more precise assessment of visitors’ emotional experiences in parks. Experimental results show that compared to traditional methods such as residual network (ResNet), recurrent neural network (RNN), and long short-term memory (LSTM), the proposed model achieves significant advantages across multiple evaluation metrics, including mean squared error (MSE), mean absolute error (MAE), root mean squared error (RMSE), and the coefficient of determination (R2). Furthermore, using the SHapley Additive exPlanations (SHAP) method, this study identified the key factors influencing visitors’ emotional experiences, such as “water”, “green”, and “sky”, providing a scientific basis for park management and optimization.