Automated urban landscape design: an AI-driven model for emotion-based layout generation and appraisal

Basic modeling

In the URDDL scheme, the whole experimental study is mainly divided into three processes, as shown in Fig. 1:

(i) Training pix2pix model. Firstly, the selected samples are divided into a training set and a test set by manual annotation, both of which are paired images. The input image of the training set is the landscape land use and surrounding road conditions. The output image is the landscape primary entrance selection and internal spatial function layout. The pix2pix model is trained for the internal function layout ability of different land use and road conditions.

(ii) The trained generator is used to generate false samples. At the same time, the paired images in the test set are input to the generator g. The generator automatically generates the matching landscape interior layout results according to the training rules. It juxtaposes the results with the original paired images to intuitively compare the differences between the generated results and the actual results.

(iii) Qualitative and quantitative analysis of the generated false samples and actual samples. From the point of view of a conventional building plan investigation, the created products are compared with the genuine products, and the quality of the machine produced comes about is assessed.

Data preparation is used to produce training data sets. The quality of training data sets is essential for deep learning training. Only by providing enough good data sets can the pix2pix model produce satisfactory results. The model training involves the input of the prepared data set into the established training environment, and the computer automatically completes the training process by adjusting a few parameters.

The generator G in the URDDL scheme is designed to convert input images into generated outputs that approximate the desired results. It utilizes an encoder–decoder architecture, with the encoder progressively down-sampling the input image to extract high-level features. At the same time, the decoder up-samples these features to produce the output image. Specifically, the pix2pix model employs a U-Net architecture characterized by its inclusion of skip connections that preserve spatial information by linking corresponding encoder and decoder layers. The input to the generator is a paired image representing landscape land use and surrounding road conditions. At the same time, the output is a synthesized image depicting the internal spatial function layout and landscape entrance selection. The generator’s training involves minimizing the discrepancy between generated images and real-world images from the training set. This process employs a combination of pixel-wise loss, such as mean squared error (MSE) or L1 loss, and an adversarial loss. The adversarial loss is provided by the discriminator, guiding the generator to produce images that are increasingly indistinguishable from real samples through iterative optimization using gradient descent techniques.

The discriminator D serves the essential function of distinguishing between real images and those generated by the generator. It is structured as a deep convolutional neural network (DCNN) with successive convolutional layers that reduce spatial dimensions while enhancing feature map richness. The discriminator processes both real images derived from the test set and generated images to assess their authenticity. The output of the discriminator is a probability score that indicates the likelihood of the input image being real. This score plays a crucial role in training the generator, as it provides feedback on the realism of the generated images. The discriminator’s loss function is usually binary cross-entropy, which measures the accuracy of the discriminator in classifying images as real or fake. During training, the discriminator and generator are optimized concurrently. While the generator strives to produce more realistic images, the discriminator improves its capacity to distinguish between real and generated images, driving both components toward an equilibrium where the generator’s outputs are nearly indistinguishable from genuine images.

The final result evaluation is the qualitative and quantitative analysis of results. The production of data sets also includes three steps:

• Data collection. The final purpose and effect of our experiment should be considered. The data collected in the first step should be filtered, and the samples that meet the screening principle should be retained as the basis of the third step of data annotation processing.

• Data processing to ensure the unity of training data style, on the basis of the second step of data filtering, the data is further labeled, and the labeled data is used as the final training data set of deep learning. Formula (1) is the final loss function, and Formula (2) is the position code for evaluating word order. (1)${\displaystyle \sum \mathrm{y}=\mathrm{n}{\mathrm{b}}_0+{\mathrm{b}}_1\sum {\mathrm{x}}_1+{\mathrm{b}}_2\sum {\mathrm{x}}_2,\sum {\mathrm{x}}_{1}y={\mathrm{b}}_0\sum {\mathrm{x}}_1+{\mathrm{b}}_1\sum \underset{1}{\overset{2}{\mathrm{x}}}+{\mathrm{b}}_2\sum {\mathrm{x}}_1{\mathrm{x}}_2,\sum {\mathrm{x}}_{2}y={\mathrm{b}}_0\sum {\mathrm{x}}_2+{\mathrm{b}}_1\sum {\mathrm{x}}_1{\mathrm{x}}_2++{\mathrm{b}}_2\sum \underset{2}{\overset{2}{\mathrm{x}}}}$ (2)${\displaystyle \mathrm{P}{\mathrm{E}}_{\mathrm{pos},2\mathrm{i}}=sin\left(\frac{\mathrm{pos}}{10002\mathrm{i}}\right){\mathrm{PE}}_{\mathrm{pos},2\mathrm{i}+1}=cos\left(\frac{\mathrm{pos}}{10002\mathrm{i}}\right).}$

Emotion perception

Emotional perception analysis can be regarded as a simple modeling of human emotional understanding (Dzedzickis, Kaklauskas & Bucinskas, 2020; Gu et al., 2024). In the evaluation of urban landscape design, a domain dictionary is created based on the existing semantic resource vocabulary, and the positive and negative emotional words are compared with corresponding emotional values, which are stimulating, sense, and action, as shown in Fig. 2.

URDDL mainly focuses on the interaction and influence mechanisms among space users, place, and emotion. The reason is that with the acceleration of urbanization and the improvement of life quality, the attention to spatial emotion helps to enhance residents’ happiness and build a people-oriented smart city. The popularity of social media networks provides a channel for the expression of individual opinions. The mechanism of emotional spatial pattern is of great significance. By solving the above key problems, it can make suggestions and policy implications for decision-makers of urban public safety, public health, and design management.

Users’ emotions exist and can derive expectations, optimism, trust, and other emotions. These emotions cannot be summarized, and even complex emotions generated by the fusion of basic emotions are challenging to identify and determine by traditional methods. Because of this situation, when we embed emotional intelligence for URDDL, we assume that β is the feature weight and X Evaluate the word frequency in the data for users. Y is the number of neurons in the output layer.

(3)${\displaystyle {\beta }^1={\left(X^{T}X\right)}^{-1}{X}^{T}y={\left(\sum x_i{x}_{ii}^T\right)}^{-1}\left(\sum x_i{y}_i\right).}$

In addition, in each kind of sentiment tendency classifier, the user comment text data represented by a word vector is used as the input of the model. The forward and backward semantic information of microblog comment Chinese text data is obtained through pix2pix, and the deep-seated feature vector is mined v. The probability distribution is shown in Formula (4) (Takada, Wang & Yamasaki, 2021). (4)${\displaystyle p_i=pix2pix\left(w_{C}v+b\right)}$

where P indicates the probability of user comments belonging to emotion i.

Network data processing

For the layout generation of professional architecture and landscape, it is essential that the characteristics of graphics and the corresponding relationship between each other, rather than the pursuit of similarity between real samples, which is a significant difference between the two. Through the final generation results, we can also find that most of the generated results are different from the real samples, and these differences are just the performance of the machine to master its basic laws through learning a large number of samples. The specific network diagram is shown in Fig. 3.

The 280 pairs of samples obtained from data processing cannot be directly inputted to the computer for learning, but they need further processing. The training image size of the pix2pix model is 256 pixels * 256 pixels, and the result of data processing is 300 pixels/inch high-resolution image (Harris et al., 2022). The input to the model is a pair of corresponding images rather than a single image, so the data set preprocessing is also divided into two steps: image compression and image combination. Different training purposes or training objects will have different optimal learning rates, and it is not a constant learning rate but a constant learning rate.

The essence of the training process is to update parameters, which requires each parameter to have a corresponding initial value. The selection of the initial value has a significant influence on the convergence speed and quality of the model. The number of iterations is the calculation times of each training sample. An iteration will calculate all the samples in the training set once. For example, if there are 25 training samples in the training set, each iteration will calculate the 25 samples one by one. Each calculation will update the generator G and sum discriminator D parameters in the first experiment; the training learning rate of the pix2pix model was 0.0002, the number of iterations was 2,200, and the training time was 1.5 h. Finally, the computer professional perspective to evaluate the quality of the generated results of the generated countermeasure network, one is based on the size of the loss function, the other is to compare the similarity between the generated results and the real samples, the more similar and more accurate, the better the training effect.

Dataset

Given a target landscape land boundary and surrounding road conditions, the computer can automatically learn the layout rules of the general layout plan from the data set of the real landscape master plan and then get the design results. We use UBGGset (https://zenodo.org/records/8352777), which serves as a large-volume sample dataset specifically tailored to support and foster URD research endeavors. The UBGGset consists of 14,627 sample images (without data augmentation), with dimensions of 256 pixels in length and width, covering an urban area of approximately 2,272 km2. The UBGGset was constructed with co-registered pairs of 3 m planet images and fine-annotated urban landscapes labeled on 1 m Google Earth image. Input images are resized to 256 × 256 pixels and normalized. Normalization techniques such as mean subtraction and standardization are employed. Mean subtraction involves removing the mean pixel value from each color channel (red, green, blue), which is typically calculated across the entire training dataset. Standardization is achieved by dividing the pixel values by the standard deviation for each channel, helping to normalize pixel values to a consistent range.

After a large number of studies, it has been found that most of the landscapes in China are of similar regularity. With the requirements of the code, most construction projects in China have unilateral layouts, and the layout design results caused by the influence of local climate conditions are very similar to those in the Lingnan area. In the second experiment, the regional restrictions were relaxed, and the source of sample collection was extended to the whole country. At the same time, due to the large span between East and West, North and South, and the diversity of terrain, the scope is finally limited to the eastern and southern coastal and central areas of China. The types of landscape nodes are shown in Table 1, where “-” denotes unable to sum.

Experimental settings

We use Python to run and train the learning network. All experiments are completed in the hardware environment of gtx2080 (video memory 6 g). Table 2 shows the training parameters for the Pix2Pix model.

Results

Figure 4 demonstrates the weight diagram. The horizontal line denotes the emotional weights, and the vertical line indicates the corresponding samples. In the process of generating the initial value of the green line, the eight points on the plot represent the best value of the initial value of the plot. With the increase in the number of iterations, the computer-generated data is approaching the real data, and the layout pattern is gradually becoming mature.

The pix2pix model used in this study adopts the method of random initialization, that is, random selection rather than human designation. Once correctly selected, the model can quickly converge and generate ideal results. On the contrary, it will lead to the dilemma of model network optimization. From the relevant content of Fig. 4, we know that the three essential parameters of the pix2pix model are strictly uncertain. In the process of the experiment, we cannot accurately know how much the learning rate is the best, how to control the number of iterations that can achieve the goal quickly and accurately, and how to give the initial weight value to optimize the results. These parameters are only one of many variables which require a lot of experience and randomness. The experimental results show that the absolute average error of URDDL is increased by 30.14%, the average absolute percentage is reduced by 11.34%, and the root mean square error is increased by 13%. Similarly, when the type or quantity of training data changes, the changing trend of the loss function is shown in Fig. 5. The horizontal line denotes the landscape flows. The vertical line indicates the change trends.

Regarding the two parameters—learning rate and number of iterations—the minimum value of the loss function may vary with an increase in the number of training samples, and the optimal learning rate can also change. If training continues with a fixed learning rate of 0.0002, one potential issue could be that this rate is too low, preventing the global loss function from reaching its minimum by the end of the iteration. In such cases, potential solutions include either adjusting the learning rate while keeping the number of iterations fixed or increasing the number of iterations while maintaining the current learning rate.

Alternatively, if the learning rate is too high, it may cause the optimization process to skip over the global minimum, resulting in a high loss function or even model instability. To address this, one might reduce the learning rate while fixing the number of iterations. Given that the optimal learning rate is not easily determined and that increasing the number of iterations extends the experimental time, we ultimately chose to optimize the experiment by fixing the number of iterations and adjusting the learning rate accordingly. In conclusion, the absolute average error of the URDDL model is 91.31%, and the maximum value is 96.87%. The fitting degree is very suitable for evaluating the emotional prediction of urban landscapes.

Discussion

The implementation of the URDDL model holds considerable promise for architecture and landscape design, as evidenced by its performance metrics and adaptability. The model’s capacity to achieve an absolute average error of 91.31%, with a maximum error of 96.87%, underscores its precision in predicting emotional responses to urban landscapes. This precision is of paramount importance for architects and landscape designers who seek to create environments that elicit specific emotional reactions. Accurate prediction of how various landscape elements affect emotional states enables designers to make more informed decisions, thereby enhancing user satisfaction and engagement with their designs.

Another notable feature of URDDL is its adaptability to different types and quantities of training data, as illustrated in Fig. 5. The model’s responsiveness to variations in data types and quantities demonstrates its flexibility and versatility, allowing it to be applied across diverse urban contexts and landscape designs. This adaptability makes URDDL a valuable tool for designers working on a wide range of projects, enabling them to tailor their approaches based on the specific characteristics of the landscape and user preferences.

The challenges associated with optimizing the URDDL model, particularly concerning learning rates and iteration numbers, highlight the complexities involved in model tuning. The decision to fix the number of iterations while adjusting the learning rate addresses the trade-offs between computational efficiency and model accuracy. For architectural and landscape design applications, this optimization strategy ensures that URDDL provides robust and reliable predictions, allowing designers to experiment with various design parameters and refine their approaches based on the insights provided by the model. Furthermore, URDDL’s ability to predict emotional responses can be seamlessly integrated into the design process, facilitating continuous evaluation and improvement of urban landscapes. By employing the model to test different design scenarios, architects and planners can gain a deeper understanding of how modifications in landscape features impact public perception and emotional well-being. This iterative approach ensures that design decisions are informed by data-driven insights, aligning them more closely with user expectations.