Urban Landscape Evaluation Based on the Theory of Resilience

For a long time, the natural landscape has become an important way for people to relieve their illnesses and release their pressure. Creating urban green space through the use of the natural environment is an effective and economical way of choosing to improve the urban ecological environment as well as the social image. Urban landscape refers to the landscape function in the human settlement environment inherent and created by the beauty of the natural landscape, which can make the city with natural landscape art so that people have a sense of comfort and enjoyment in urban life [18]. People and the environment are interdependent, and co-development, the gradual increase in human material conditions and spiritual conditions of the demand makes us for the environment more and more strong, the environment has a certain role in regulating psychological fatigue and bad mood, contact with the natural environment is more beneficial to the psychological recovery. Emotions have an important impact on the operation of human daily life; human brain activity is divided into three levels of the overall function of the human body that play different roles, so the urban landscape design is crucial, indirectly affecting the daily life of the residents. The three operational levels of brain activity and the recovery benefits of the natural environment are shown in Fig. 1.

In order to make the urban landscape play the role of regulating and relaxing the residents’ emotions, the study proposes to apply the theory of restorative environment to urban landscape design. The healing environment refers to the environment that has the effect of restoring the residents’ physical strength and health level, including the two major theories of stress reduction theory, psychological evolution theory, and attention restoration theory. The theory of stress reduction refers to the fact that users who suffer from psychological and behavioral stress due to various negative stimuli in daily life can be cured in the restorative environment, and the condition of cure is that the users have full contact with the natural elements in the restorative environment. The theory of attention restoration refers to the fact that people’s attention is easily attracted to natural environments, and psychological fatigue can be reduced without consuming their own psychological functions [19]. In order to evaluate the restorative properties of a certain natural landscape environment in an all-encompassing and multi-dimensional way, four typical characteristics, namely, remoteness, extensibility, charm, and compatibility, are clarified under the framework of the attention restoration theory, which defines the ways in which restorative environments are able to bring about stress relief to the psyche of the users. Currently, the most commonly used evaluation method for attention restoration is to use the Perceived Restorativeness Scale (PRS) [20]. The specific evaluation indexes of PRS are shown in Table 1.

Fig. 1.

Three operational levels and the benefits of natural environment restoration.

However, a single restorative evaluation lacks a comprehensive examination of environmental elements. Therefore, the study evaluates the attentional restoration effect of the environment through PRS and draws conclusions from data analysis of the performance of different urban scenes in the four dimensions of attentional restoration. At the same time, the Semantic Differences (SD) method was used to evaluate the environmental elements of the urban landscape and analyze the environmental elements affecting the effect of attention restoration. The SD method refers to the screening of the language vocabulary of the residents around the parks, summarizing the summary results for the division of the indicators, and selecting appropriate adjectives to describe the indicators. Finally, professional students were asked to score all the park scenes according to the indicators. Professional students are then asked to score all the park scenes according to the indexes so as to dataize people’s psychological feelings. The relevant data are then statistically analyzed through the data analysis method, and the obtained data can be used as support for urban healing landscape renovation design, which provides a certain reference for the renovation design.

This quantitative evaluation method can assess the degree of attention restoration of individuals by urban landscapes and, to a certain extent, focus on the adaptability and sustainability of urban landscapes, thereby providing some reference for urban landscape design with psychological healing effects. However, this evaluation method may also lead to subjective bias, resulting in limitations in the evaluation results and an inability to provide important details about landscape composition. Therefore, it is crucial to accurately identify the specific landscape elements that affect the resilience of urban landscapes after evaluating their resilience in order to make targeted improvements in urban landscape design and enhance the city’s resilience.

CNN-Based Urban Landscape Element Recognition Model Construction

On the basis of conducting a restorative evaluation of urban landscapes, in order to provide a large amount of landscape data for urban landscape design, help designers better understand the structure and composition of existing landscapes, and provide a scientific basis for urban landscape design, research will further identify the elements of urban landscapes. In urban landscapes, the environment of historical core areas is usually more complex and difficult to identify. Therefore, research mainly takes historical core areas as an example to explore effective identification methods for urban landscape elements. Deep learning has been widely used in image recognition but performs poorly in narrow streets and complex environments. To this end, a CNN-based urban landscape element recognition model is proposed to accurately recognize landscape elements in the historical core of the city. CNN is a type of feed-forward neural network that contains convolutional operations and has a deep structure, which is widely used in the fields of image recognition, natural language processing, speech recognition, etc. [21]. CNN generally includes five types of network hierarchical structures: input layer, convolutional layer, activation layer, pooling layer, and fully connected layer. Fig. 2 displays the basic structure.

Semantic segmentation is an image segmentation technique in the field of computer vision, and the core objective is to assign each pixel in an image to a predefined category, thus realizing a refined understanding of the image. Existing mainstream semantic segmentation algorithms are composed of an encoder and a decoder structure, where the encoder can map the input data into a high-dimensional feature matrix, which facilitates subsequent processing and analysis, and the decoder can distinguish the semantic differences between different landscape elements and then accurately extract various landscape elements. In order to adapt to the recognition of complex urban landscape elements, the study introduces a polarized attention mechanism to strengthen the model’s ability to discriminate between different landscape elements. ResNet-r50, an image classification network with fully connected layers removed, is selected as the encoder. resNet is a deep CNN that uses residual blocks to construct a deep network and to overcome the problems of gradient vanishing and gradient explosion caused by the increase in the depth of the network. resNet introduces a leapfrog link on each residual block and adds the output value in each residual block to the original input value, and the obtained output is as shown in Equation (1). The output obtained is displayed in Equation (1).

In Equation (1), F(x) denotes the output value of the residual block, x denotes the original input value, σ denotes the linear rectification function and W denotes the weight parameter of the convolutional layer. In order to solve the effect of the scale difference of landscape elements, the study introduces the inverse convolution solver dock combining Atrous Spatial Pyramid Pooling (ASPP) with null convolution as a semantic segmentation decoder, as shown in Fig. 3.

In order to further improve the extraction accuracy of urban landscape elements, the study introduces a polarized self-attention module in the encoder and decoder. The polarized self-attention module, as a more refined dual-attention mechanism, is able to maintain internal high resolution in spatial and channel dimensions, thus further reducing the information loss caused by the reduced dimensionality. The Softmax function is also used to enhance the range of attention in the branch where the dimensions are compressed, and finally, the Sigmoid function is used for dynamic mapping. The polarized self-attention module is obtained by fusing the spatial and channel self-attention branches in series. In this case, the process of the spatial self-attention module to get the feature map possessing spatial self-attention is shown in Equation (2).

Fig. 2.

The structure of CNN.

Fig. 3.

Schematic diagram of ASPP.

In Equation (2), φ2, φ2, and φ1 denotes the weights of different layers, P^sp(X) denotes the spatially polarized self-attention map, fSG denotes the activation function, σ denotes the Softmax function, fGP denotes the spatial pooling layer, W denotes the 1×1 convolution, and X denotes the input feature map. The channel self-attention module compresses the spatial information while fully retaining the channel information, which can efficiently extract the features of different categories of elements, and the study retains half of the total number of channels to compute the channel self-attention in order to improve the efficiency, as shown in Equation (3).

In Equation (3), P^ch(X) denotes the channel polarization self-attention map. The study inputs the original feature maps to the spatial self-attention and channel self-attention modules, respectively, and sums the obtained feature maps to enhance the category characteristics and spatial distribution information of the landscape elements, as displayed in Equation (4).

In Equation (4), F(X) denotes the final attention map and ⊙ denotes the matrix multiplication operator. The polarized self-attention module is shown in Fig. 4.

To improve the training efficiency, migration learning is introduced to train the network to learn landscape element features based on the street view information embedded in the existing dataset. The migration learning process of the encoder is shown in Equation (5).

In Equation (5), Pfe denotes the encoder parameters trained using generic visual information, Dimg denotes the visual image information, Tf denotes the image categorization task, Dhcsv denotes the landscape element features, and P'_fe denotes the encoder parameters with landscape element information and generic visual information. The migration learning process of the decoder is shown in Equation (6).

In Equation (6), P and P' denote the decoder parameters before and after migration, respectively, and Pfe denotes the street feature recognition task. To assess the performance, the study uses Average Accuracy (AA); Overall Accuracy (OA); Intersection over Union (IoU), and Mean Intersection over Union (mIoU) as evaluation metrics. mIoU are used as evaluation metrics. AA is the average value of model recognition accuracy, as displayed in Equation (7).

In Equation (7), n denotes the total number of categories, X denotes the number of correct identifications, and Ni denotes the number of pixels whose true category is the i category. The calculation of OA is shown in Equation (8).

In Equation (8), M denotes the full number of pixels. The IoU is displayed in Equation (9).

Fig. 4.

Polarized self attention module.

In Equation (9), Xij denotes the number of pixels of class i that are recognized as class j and Xji denotes the number of pixels of class j that are recognized as class i. The calculation of mIoU is shown in Equation (10).

Evaluation Results of Urban Landscapes Based on the Theory of Resilience

Five different spaces were selected as samples based on type, geographic location, and other factors, namely, plaza space, waterfront space, road space, rest space, and play space. In order to facilitate the statistical processing of the data, the evaluation scale was chosen to be 5 levels, with the highest value of 2 points and the lowest value of -2 points. For the five selected spaces, three dimensions and 15 evaluation indexes are formulated according to their landscape characteristics, and the evaluation indexes of the SD method are shown in Table 2.

Six scenes were randomly selected from five spaces for evaluation, and 40 landscape design students were selected as research subjects for the study. A total of 50 questionnaires were issued, and 50 valid questionnaires were recovered, with a validity rate of 100%. The data obtained from the questionnaire were comprehensively statistically analyzed using Excel. The PRS evaluation results of the 6 scenes and 5 spaces are displayed in Fig. 5. From Fig. 5a), Scene 4 has the highest rating overall and Scene 5 has the lowest rating overall. From Fig. 5b), the waterfront space has the highest PRS rating, and the plaza space has the lowest PRS rating. The results indicate that the plaza space has the worst recovery of attention and needs to be optimally designed.

The SD evaluation results for the five spaces are displayed in Fig. 6. In Fig. 6, the plaza space landscape, which has the lowest restoration effect of attention, has the highest score on the openness evaluation indicator at 2.51. The waterfront space has the best performance overall, with no indicators with the lowest evaluation scores. Open space had the highest evaluation score on the public infrastructure indicator and the lowest evaluation score on the hydrophilicity indicator.

Effectiveness Analysis of the Urban Landscape Element Identification Model

To verify the CNN-based urban landscape element recognition model, the study uses the Cityscapes dataset for experiments and compares the improved CNN algorithm combined with the ASPP proposed (CNN-ASPP) in this study with three models, namely, the Fully Convolutional Network (FCN), the Asymmetric Non-local Neural Network (ANNN), and the Pointwise Spatial Attention Network (PSANet). The comparison of the metrics results of the four models is shown in Fig. 7. From Fig. 7a), it can be seen that compared to the other three models, the CNN-ASPP model has the highest AA metrics of 72.78%. From Fig. 7b), it can be seen that the CNN-ASPP model also has better performance in terms of OA metrics, with an OA value of 90.49%. From Fig. 7c), the CNN-ASPP model has the highest mIoU metrics with 63.8%. The proposed model has better performance in AA, OA, and OA indicators and has a better extraction effect of landscape elements.

Fig. 5.

PRS evaluation results for each scene and space.

The picture extraction results of FCN, PSANet, and CNN-ASPP are shown in Fig. 8. Comparing Fig. 8a), Fig. 8b), Fig. 8c), and Fig. 8(d), the FCN model has the worst performance, and the extracted pictures are more blurred. The PSANet model extracts clearer outlines, but the details are not clear enough. The CNN-based cityscape element recognition model extracts the most complete picture elements, demonstrating good picture element extraction.

To further qualitatively analyze the element recognition performance of the models, the study recognizes six different landscape elements. The IoU values extracted by the four models for different landscape elements are displayed in Fig. 9. From Fig. 9a), the FCN model has better results in recognizing the sky with an IoU value of 88.11%. From Fig. 9b), the ANNN model has the best results in recognizing buildings with an IoU value of 86.81%. It can be seen in Fig. 9c) that the PSANet model has the best recognition for buildings with an IoU value of 86.81%. In Fig. 9d), the model in this study has the best recognition for buildings with an IoU value of 90%. The IoU values for recognizing vegetation, roads, pedestrians, motor vehicles, and buildings are higher than the other three models, and the IoU value for recognizing the sky is 86.55%, which is 1.56% lower than the FCN model. The proposed model has the best recognition effect in general and has better performance for recognizing different landscape elements.

Fig. 6.

SD evaluation results of five spaces.

Fig. 7.

Comparison of indicator results of four models.

Fig. 8.

Image extraction results of three models.

Fig. 9.

The IoU of four models for extracting different landscape elements.

In order to verify the running efficiency of the CNN-based urban landscape element recognition model, the study uses the significant memory occupation, the number of parameters, the GFLOPs, and the prediction speed as evaluation indexes for testing. The comparison results of the operational efficiency of the four models are displayed in Table 3. From Table 3, the CNN-ASPP model has a memory occupation of 9178 Mb, the number of parameters is 43.41 million, and the computational complexity is 1619.98 GFLOPs, which is slightly higher than that of the FCN and ANNN models, but it achieves higher efficiency, and the prediction speed is 1.2 FPS. The memory occupation and the number of parameters are substantially lower than those of the PSANet. The efficiency of the proposed model is competitive compared to large networks.