Designing Care Spaces in Urban Areas

Abstract

This paper presents a novel approach to sustainable urban revitalization. Care Spaces are defined, and an area selected for revitalization is described. Transformation of city space is of fundamental importance for everyday life, its comfort as regards the functional aspect as well as the psychological and cultural ones. The presented projects are in accord with 2030 Agenda for Sustainable Development and the conception of Baukultur. Both approaches tend to create well designed environment that support health and well-being of people and other living creatures while taking into account cultural aspects in design and construction. Focusing on the combination of soundscape analysis with design elements. To monitor the soundscape, a custom database of urban sound recordings was constructed, and key analytical methods such as Mel-Frequency Cepstral Coefficients (MFCC), feature extraction, Recurrent Neural Networks (RNN), and permutation of feature importance were applied. The effectiveness of these algorithms was confirmed through field investigations.

1. Introduction

Debates concerning modernization and creation of new cityscapes are taking place all over the world [1,2]. Studies conducted in New York, Melbourne, and Tokyo by William H. Wythe [3] have indicated three elements of human activity in urban areas: people were seeking interaction, there occurred a phenomenon called the crowd’s attraction power, and people were happy to remain in the mid-zone between close persons and strangers. Nowadays, the fragmentary character of city planning systems and the inappropriate city infrastructure, which is incompatible with the climate changes of today, show that it is necessary to change the way of thinking about sustainable development. Planning requires understanding of the present and future needs and threats involving the people living today. In the current paper, the authors are trying to emphasize the fundamental importance of transformation of urban space for our everyday life as well as its comfort in the functional, emotional and cultural aspects. The idea of sustainable development requires combining the knowledge from various domains and creating an interdisciplinary team. The team should include interior designers and architects [4] directly responsible for designing and shaping of the space. The designers involved in sustainable development of cities should be assisted by sociologists [5], which would allow for efficient monitoring and assessment of social changes influencing this environment. On the other hand, spatial planning geographers [6] have a reverse task. They study in what way environmental changes transform modern societies and restructure future social and spatial imagery. The team also requires specialists in acoustics [7] who would examine the required soundscape and participate in developing it. The authors of the current work are members of such an interdisciplinary team. The paper presents the results of workshops during which the team members combined their knowledge of various domains, applying it from various perspectives in order to design a “Care Space”. Working in a multidisciplinary team, we determined our goal as recognizing of the tendencies and directions of development based on three values: beauty, sustainable development and togetherness emerging or/and getting reinforced in the society. Like other scientists [8,9,10], while designing cityscapes we take into account transformations in accord with the 2030 Agenda for Sustainable Development [11]. It is a resolution adopted by 193 countries—members of the United Nations Organization in 2015, defining the model of sustainable development. Sustainable cities are safe spaces, favourable to social inclusion and education, fighting poverty, caring for good health and life quality. We also work in accord with the concept of Baukultur [12]. Baukultur [13] signifies well designed cities, villages and buildings creating life environment that promotes the sense of well-being and good health in people, birds and other animals, accepting cultural aspects of protection, planning and construction.

2. Materials and Methods

2.1. Execution of Research Projects in the Form of Interdisciplinary Workshops New Space

The series of interdisciplinary workshops “New Space” is based on the collaboration of research workers from four Krakowian institutions of higher education: the Academy of Fine Arts (Faculty of Interior Design); Cracow University of Technology (Faculty of Architecture); Jagiellonian University (Institute of Sociology and Geography and Spatial Management); and AGH University of Krakow (Faculty of Mechanical Engineering and Robotics, Field of Acoustic Engineering). The series of workshops involved defining and designing Ethical Spaces, Oppressive Spaces, Alternative Public Spaces [14], and the Care Spaces discussed in the current paper. Cooperation between the higher education institutions includes the exchange of university teachers and researchers engaged in various forms of education, both within the obligatory syllabuses and beyond. The workshops are attended by students, postgraduates, and academics. While preparing the topic of the workshop, the academics aim to turn the collaboration into a form of professional training that includes skills necessary in the labor market. The students work in six different groups consisting of architects, artists, sociologists, urban geographers, and acoustics engineers. The ideas developed in the individual groups are discussed with PhD students and academics. Each workshop lasts one month, beginning with an introductory lecture. Then, the students get acquainted with the given space, discuss ideas concerning revitalization, and get to know the other members of their project team. In the following week, they work together on the project at one of the educational institutions, and in the final week, all teams present the initial results of their work. Subsequently, the results are prepared for a presentation to which economic operators interested in the topic are invited. The teams regularly collaborate with municipal officials responsible for zoning. Every year, our projects are discussed in the press.

2.2. Defining a Care Spaces

In 2021, the post-Covid phase began; 24 February 2022, marked the beginning of the war in Ukraine, and the first wave of war refugees arrived in Poland. At the same time, Europe experienced dramatic weather changes caused by the climate crisis with increasing frequency. In response to these events, a research team consisting of academics initiated the topic “Care Spaces” [15], aimed at creating a new normal and counteracting a disturbed sense of security. Students expanded this definition. The projects conducted during the workshops are presented in the chapter Results. The understanding of care in the projects is multi-dimensional and manifests itself in various interconnected and overlapping perspectives. The project “Wesola Anew” includes nurturing the memory and identity of the area; “Green Empiria” involves the use of natural resources and vegetation. Another important aspect involved providing broadly understood security, psychological and/or physical well-being, through revitalization and reorganization of spaces that were supposed to perform therapeutic and care functions, as seen in the project called “Zen Tree”. Making the space free of noise and urban traffic was an essential element of the project “Com’on Space”. Moreover, activation of different senses, such as hearing, sight, and touch, in people using the space, thus expanding its therapeutic potential through sensory experience, was an element of the project

“SUNE ERGIA” The students also made a point of satisfying people’s needs, offering solutions that took into account various possibilities and expectations, including this in the project “Plan B” Inspiring and stimulating people’s activity through creating spaces facilitating cooperation was the goal of the project “Wesola Space”. “Green Empiria” involved ensuring accessibility of spaces and designing places ready to accommodate all users. The “Wesola Anew” project included thinking about the future and creating new possibilities and conditions for development with the use of organic materials friendly for the present users as well as the future generations. The described classification plays basically the ordering and cognitive role; it is neither finished nor separable.

2.3. Description of the Area Selected for Revitalization

Wesola quarter is situated in the center of Krakow city, not far from the railway station and bus station, close to the Old Town, which is the core of the City. The Botanical Garden of the Jagiellonian University, founded in 1783, is located within this area. Historically, the quarter performed various functions; in the 19th century, a complex of hospitals was constructed there. Recently, after more than 200 years, present doctors and patients have been moved to modern buildings constructed in another part of the city. The area in the center of the city became abandoned and required revitalization. Between 26 November 2020, and 26 February 2021, public consultations were carried out regarding the spatial planning of “Wesola—Kopernika Street area”. Hydrological guidelines, conservation recommendations, and spatial requirements were also developed. An inventory of greenery was carried out. The report described the image of the Wesola area as a green enclave. In accordance with the expectations of people taking part in the public consultation, it should become a space for the meeting and coexistence of residents, artists, and scientists.

Technology and ecology should intertwine here. The space should be transformed for the residents, taking into account the needs of various social groups. The quarter should also be the place of cooperation for social and professional groups, artists, businessmen, and activists, healthy persons and convalescents, representatives of modern professions and disappearing trades, followers of different religions, and atheists. The timeless attractiveness and original character of this place should serve as the common point uniting these diverse characteristics. Administrators of the area have suggested that modern buildings, such as the Krakow Library, a Culture Quarter, and Krakow Municipal Holding, should be located there. However, due to the fact that the area lies in the center of the city, all spatial plans are permanently involved in legal controversies. To date, no revitalization has started in this area. The current paper discusses alternative solutions offered by the students, postgraduates, and academics from Krakowian higher education institutions. Considering the fact that the project teams include acoustics engineers, the authors also consider the aspect of audio sphere, which has not been addressed in the existing spatial plans.

2.4. Soundscape Studies

When creating modern spaces or reclaiming formerly lost locations in cities, we approach the execution of plans with greater awareness of the soundscape. In the field of science, there is a new domain of soundscape studies [16,17] aimed at integrating positive emotions with the structure of sound that will be present in the designed space. The goal is to achieve balance between living organisms and their acoustic environment [18,19,20]. To start a debate on what soundscape analysis is possible when the revitalized area is merely a design, it is necessary to include data science engineers in the interdisciplinary team and start a study on Artificial Intelligence algorithms for monitoring quiet enclaves.

One of the goal of the paper is to combine the soundscape [15] analyzed in the context of the project with the actual area. Every project of urban area revitalization has its strategic spots that require an appropriate soundscape. Network training that classifies soundscape as, e.g., quiet or overwhelming allows for constant monitoring of these spots before the project is adopted. If the soundscape in the monitored spots becomes overwhelming, the situation can be changed as early as the project modification stage, by eliminating unfavorable acoustic events using appropriate soundproofing materials. Revitalizing an area requires eliminating all sorts of soundscape that may signal threat or overcrowding. In such places, people get isolated from one another and from their surroundings. Lack of any audible sounds also causes unease. Monitoring of sounds in urban areas is a significant element of soundscape management based on the analysis of complex sound patterns [21,22]. However, space design is not solely dependent on determining the sound level at a given spot, but also on recognizing which sound features facilitate the sense of peace and tranquility in the urban environment. Monitoring is becoming increasingly accessible thanks to the rapid development of neural networks. Research is being conducted on the recognition of cityscape sounds with the help of various neural network models, such as artificial neural networks (ANN) [23], convolutional neural networks (CNNs) [24,25], recurrent neural networks (RNNs) [26], and neural networks with long short-term memory (LSTM) [27]. These studies have shown that neural networks show promise as tools for monitoring and classification of urban sounds [28,29]. Integration of these methods with traditional signal processing techniques and modern technologies like Internet of Things (IoT) can bring about more efficient management of urban soundscapes [30]. Thus, the application of artificial intelligence methods in defining quiet spaces and overwhelming ones and implementing these results in urban revitalization projects has recently become possible.

In every space, we are surrounded by a soundscape. Sound is spatial; therefore, acoustic engineers and those collaborating with them can hardly consider environmental sound as a two-dimensional landscape; it is easier to perceive it as an audio sphere. Sound loses its spatial character and its ability to evoke good emotions only if it is muffled by broadband noise on streets, in locations full of hustle and bustle, or by artificially emitted sound.

More recent studies also highlight the importance of quiet spaces in urban environments as a key component of sustainable city planning. Research has shown that natural sounds not only alleviate stress but also enhance social interactions, suggesting that quiet urban areas should be designed to maximize exposure to these sounds [31,32]. Additionally, the shift from merely reducing noise to creating intentional soundscapes emphasizes the role of sounds as valuable resources rather than undesirable by-products [33]. Approaches that integrate both subjective experiences and objective measurements have proven effective in identifying quiet zones, supporting the notion that soundscapes are complex, multi-layered phenomena requiring a nuanced approach [34]. This complexity is further demonstrated in studies utilizing AI models that simultaneously detect sound events and predict levels of annoyance, offering insights into the dynamic interplay of different sound sources in the urban context [35,36].These insights support the approach of integrating artificial intelligence with modern soundscape analysis to create urban revitalization projects.

In Care Spaces, the desired sounds are those connected with nature—it is a kind of silence that symbolizes tranquility without disturbing good relationships. The authors decided to define quiet spaces and overwhelming spaces from the perspective of sound analysis. Among the infinite number of combinations of sound events that may take place in a city space, in the following chapters, the authors search for features of acoustic signals that can define tranquility and its opposite.

2.5. Development of the Database

In the urban soundscape there occurs a huge but finite number of acoustic events and an infinite number of their combinations, co-occurrences, etc. The distance between the observer and the source of sound as well as the number and kind of acoustic event are of great importance in shaping the soundscape. It is really hard to generalize this problem, therefore it should be viewed from a different perspective. Nowadays, we are getting used to sounds in the city and we either do not listen to them or isolate from them with the help of headphones. If we started to listen to the soundscape, the anthropogenic sounds produced by air conditioning, generators, street traffic or warning sirens would soon become overwhelming and tiring. During the workshops we design potentially therapeutic Care Spaces in which broadly understood peace and quiet is necessary. The space which is to be revitalized is situated in the very centre of the city and surrounded by streets carrying public transport including railways, buses and trams, as well as private cars. We have two opposite acoustic environments: the one expected in the project, i.e., care spaces, and the overwhelming areas in the vicinity of the revitalized space. Hence, the study is focused on the binary classification of urban sounds as quiet spaces and overwhelming ones, which aims to answer the question of how different urban sounds influence the perception of the acoustic character of a specific environment [37,38]. Binary classification will facilitate checking which areas in the project are exposed to disturbances coming from the neighboring streets.

The discussed analysis is not focused on individual acoustic events but on the affective perception by human senses of the total soundscape of a specific place [39]. This is why most of the commonly accessible databases do not provide significant information for designing Care Space. When we have to do with projects of revitalizing of urban space, while designing the soundscape it seems necessary to search for the existing spots that satisfy the definition of a Care Space. We watched thousands of videos featuring strolls within cities. We were searching for peaceful urban spaces and overwhelming ones. The labels were verified to decide whether the acoustic events occurring in the videos could take place in the space we designed. Therefore, a database consisting of generally available recordings, which were downloaded from YouTube and are licensed under Creative Commons Attribution (CC BY), documenting soundscapes of various urban areas from all over the world (Figure 1), was developed in the Google Collaboratory environment. When the set of videos reached the size that allowed us to teach the network, we terminated our search.

The analyzed recordings registered people moving around different urban areas. In order to extract the audio track only, each recording was downloaded and transformed from .mp4 into .mp3 format with the help of Pytube [40] and Moviepy [41] libraries. In the next step, they were divided into sixty-second fragments which were then analyzed and, depending on the context, classified in one of the two categories as “quiet” or “overwhelming”. This division allows for a more precise analysis of the urban soundscape [42]. The total of all one-minute fragments with defined labels constitutes the database for further analysis. Depending on the length of the original recording, each considered location on the map has a specific number of segments. The database includes recordings from culturally and geographically diverse areas, which allows for studying the universal character of the model. The results of this process are presented in

Table 1. Summary of Recording Locations Used for Database Construction, Including the Number of Segments from Each Location and Their Classification Labels (Quiet or Overwhelming).

Location	Number of Segments	Category	Location	Number of Segments	Category	Location	Number of Segments	Category
New York, Manhattan	35	Overwhelming	Madrid	43	Overwhelming	Madrid	Amsterdam Vondelpark	20
Vienna Stadtpark	21	Quiet	Barcelona	25	Overwhelming	Barcelona	Phoenix Park, Dublin	28
Vienna Türkenschanzpark	32	Quiet	Milan	52	Overwhelming	Milan	Benjakitti Park, Bangkok	22
Central Park, New York	52	Quiet	Englischer Garten, Munich	27	Quiet	Englischer Garten, Munich	Albert Park, Auckland	23
London	33	Overwhelming	Japanese Garden, Nagoya	23	Quiet	Japanese Garden, Nagoya
Paris	40	Overwhelming	Chicago	34	Overwhelming	Chicago
Total			Number of segments
Quiet			262
Overwhelming			248
Total			510

.

2.6. MFCC Feature Extraction

The next stage of the analysis was aimed at extracting the features of Mel-Frequency Cepstral Coefficients (MFCC) from each one-minute segment of the recording. This process allows for effective recognition of speech and sound in the recordings thanks to the ability of MFCC features to distinguish sound characteristics that are perceptually significant for the human ear [43]. In our study, we extracted 20 MFCC features (n_mfcc = 20), which provided a balance between capturing sufficient acoustic information and reducing the dimensionality of the dataset. First, audio files were downloaded with the help of the Librosa library [44], which enabled their transformation into numerical format. Then, the sound signal was transformed from the time domain into the frequency domain with the use of the Fourier Transform, which allowed the analysis of its frequency components. The power spectrum was then converted to the Mel scale, which is designed to mimic the human ear’s perception of sound frequencies. The Mel scale uses a logarithmic transformation, making it particularly suitable for analyzing urban soundscapes where low-frequency sounds, such as traffic noise, play a significant role. Subsequently, the amplitudes on the Mel scale were logarithmized, enhancing the differences between weak and strong signals. Finally, the logarithmized spectrum was subjected to a discrete cosine transform (DCT), which reduced the data to a small number of coefficients that well describe the spectral shape of the signal. The first 20 coefficients were selected as they provide a good representation of the spectral envelope while discarding the less informative high-frequency components. The processed MFCC coefficients were averaged along the time axis, resulting in a set of statistical features representing the entire one-minute segment [45]. This averaging helped reduce the impact of short-term variations in the sound signal and provided a more stable representation of the soundscape. The extracted features were recorded in numerical format, allowing for their further analysis using machine learning models. By focusing on these MFCC features, we leveraged their ability to capture both the spectral and temporal aspects of the urban soundscape, which are critical for distinguishing between “quiet” and “overwhelming” sound environments. Figure 2 shows Mel-spectrograms for an overwhelming space in Manhattan, New York (a), and a quiet space in Stadtpark, Vienna (b).

2.7. Recurrent Neural Network (RNN)

After collecting and classifying the data from various places all over the world, the next step involved designing and implementing a neural network model capable of effectively analyzing urban soundscapes. The architecture (Figure 3) chosen was based on recurrent neural networks (RNN) [46,47], which in this case are used to process dynamic, time-varying sequential data such as sounds [48]. The RNN model consists of two stacked SimpleRNN layers, each containing 50 neurons, followed by ReLU activation functions to introduce non-linearity and capture complex patterns in the data. The RNN model was designed so that the information included in MFCC features could be used in full. The input data were reshaped to match the RNN input format allowing the network to process sequences of MFCC features effectively. To mitigate the risk of overfitting, dropout layers with a rate of 0.2 were added after each SimpleRNN layer. These layers randomly exclude a certain proportion of neurons during the training, which forces the network to learn more general patterns in the data instead of relying on specific, potentially random training features. The network uses recurrence, meaning that each recurrent unit processes the input data while also taking into account information from previous time-steps. This design allows the model to capture temporal and contextual dependencies present in urban soundscapes.

The training data were split into 80% for training and 20% for testing using stratified sampling to maintain class balance. The model was trained using binary cross entropy as the loss function, and the Adam optimizer with a learning rate of 0.001 was chosen for its efficiency in updating the weights. The training process involved 30 epochs with a batch size of 64, during which both loss and accuracy metrics were monitored on training and validation sets. This approach enabled continuous evaluation of model performance and early detection of potential overfitting. The final Dense layer contains 50 neurons with a ReLU activation function, followed by an output layer using a sigmoid activation function for binary classification (quiet vs. overwhelming).

3. Results

3.1. Projcets Developed During the Workshops

During the “New Space” workshops [15], different conceptions of Care Space were developed. These projects included places favorable for reviving social relationships, supporting mutual care and cooperation, or focused on security and coziness as well as promoting psychological and physical wellbeing through contact with nature. They were a response to the current needs of the area, its users and residents, and to the modern challenges connected with climate changes, post-pandemic health problems, and social tensions issuing from conflicts and mass migration. Care was executed in them through appropriate (organic) shapes of the designed objects and elements, the use of natural materials and colors, planting of vegetation, and the application of noise-masking technologies and air-purifying filters. The spaces were designed so that they could also use the therapeutic potential of the existing hospital complex by creating a refuge from excessive traffic noise. The projects discussed are presented below.

Figure 4 presents the Com’on Space project by Pat Czepiel, Anna Faltyn, Kacper Góra, Katarzyna Hetmańczyk, Julia Juros, Marta Romankiv, and Maria Talaga.

Figure 5 shows the Green Empiria project by Julia Bąk, Patrycja Jarzębak, Weronika Jaszczyk, Kinga Sapieja, Magdalena Świdnicka, and Sylwia Wojdan.

Figure 6 illustrates the Wesola Place project by Marcin Baumann, Piotr Brauntsch, Aleksandra Gierszewska, Justyna Kopacz, Paweł Michalak, Bartłomiej Ormaniec, and Paulina Polaczek, Ludwika Pysz.

Figure 7 displays the Wesola Anew project by Wiktoria Bajda, Maria Brzyska, Kinga Bukowiec, Olga Gizelska, Aleksander Sielecki, and Grzegorz Tryba.

Figure 8 presents the Plan B project by Urszula Kraszewska, Aleksandra Rosek, Marzena Cieniawska, Joanna Jaros, Nikodem Chodakowski, Gabriela Mermon, and Jakub Naspiński.

Figure 9 shows the SUNE ERGIA project by Kamila Łuczak, Anna Leśniak, Paulina Pasztalaniec, Aleksandra Kotyła, Weronika Wierzgacz, Weronika Wolnik, and Alicja Sędzikowska.

Figure 10 illustrates the ZenTree project by Agnieszka Gajewska, Mateusz Dziuba, Dagmara Jędrychowska, Przemysław Poręba, Eden Dejena, Emilia Stefanowska, and Magdalena Masłowska.

3.2. Designing a Soundscape

Designing a soundscape will be discussed on the basis of the Green Empiria project. This project involves the development of a Care Space facilitating mental recovery, that is, a sphere in which persons who are not members of a given community would feel invited and could move freely within it. Based on the executed measurements, the space arrangement was designed so that its conditions could be put to the best use. At the spot where the registered level of noise was the lowest, a space for meditation called the Enclave of Silence was created. At this spot, the traffic noise is hardly audible, but to protect the enclave better from undesirable noise, wooden sound-absorbing panels were installed. The project also includes bicycle lanes and a preventive healthcare center that would promote health screenings. For example, a coffee shop in this area could offer special discounts like ‘Coffee for health check-ups’, thus supporting preventive healthcare. The Enclave of Silence is a place for calming down and meditation, equipped with sound-absorbing materials. On the other hand, an information center equipped with interactive boards and models would provide information about the whole area and the forms of mental recovery. Additionally, in the sensory reset garden, one could find special therapeutic booths that would ensure full sensory regeneration. The area also provides horticultural therapy zones, art therapy spheres, and aromatherapy spots, which offer diverse forms of therapy through contact with nature, art, and fragrances. The design was visualized in detail in Figure 11.

In Figure 11, the locations where acoustic measurements were conducted are indicated and labeled as follows: A—42.8 dB, B—40.3 dB, C—37.8 dB, D—42.8 dB, and E—55.2 dB. These spots were selected so that they could cover various zones of the planned area. At measurement spot A, where 42.8 dB was recorded, there is a bridge, bicycle lanes, and a coffee shop. The spot was selected due to the need for monitoring of the intensely used area where pedestrian and bicycle traffic could generate noise requiring permanent control. Measurement spots B, C, and D, with the results 40.3 dB, 37.8 dB, and 42.8 dB, respectively, are situated in the central part of the project, known as the Enclave of Silence. This is where the planners designed a meditation space that was supposed to offer the minimum level of noise for the best relaxation and calming down [15]. Measurement spot E, with the highest recorded level of noise equal to 55.2 dB, is situated on Grzegorzecka street, close to the bridge and the entrance. This location is pivotal for monitoring and controlling the impact of city traffic on the surrounding area as well as for verifying the effectiveness of sound-absorbing barriers [49,50]. The measurements were executed in order to precisely determine the intensity of sound and identify the quietest areas, which allowed for appropriate planning [51]. Thanks to this, it was possible to create spaces facilitating relaxation and calming down, which is vital for the execution of the Care Space concept.

Individual measurements of the level of sound may be used only in the preliminary analysis. Designing enclaves of tranquility that are identical to Care Spaces is a complicated task. Soundscape is influenced not only by the levels of sound but also by its structure. During the day and night, sources of sound and their number change. What is more, the simultaneous presence of multiple sources of sound is also harmful, even if the level of sound is low. We could hardly neglect the fact that urban areas have an infinite number of acoustic events. It is the neural networks that deal with combining these issues with the help of one algorithm. In the next chapter, the authors discuss the evaluation of such an algorithm. Good training of the network allows for monitoring of strategic spots in the project. If the soundscape becomes overwhelming during certain periods of night or day, changes should be introduced in the project or the necessity of eliminating such acoustic events should be taken into consideration.

3.3. Evalution of the Recurrentneural Network Model

Evaluation of the recurrent neural network (RNN) model was executed with the help of Learning Curves and Validation Curves. They provide essential information regarding the effectiveness of the model in the classification of urban sounds. An analysis of these curves allows for an evaluation of both the learning process of the model and its ability to generalize on unknown data. Figure 12 and Figure 13 show the Learning Curves of the trained RNN.

The model is quick to achieve a high level of precision on the training and validation data. The initial leap in precision, particularly in the first epochs, is a sign of correct pattern learning. Precision on validation data, though it is a bit lower than on training data, remains constant, which indicates good generalization. The loss curve shows how the model improved over time. The rapid decrease in the loss value in the first epochs shows that the model quickly adjusts to the task. Further, the gradual decrease in loss on both the training and validation data indicates that the model continues learning and improving in the more subtle aspects of sound classification without losing the ability to generalize. The stable character of the precision curve and consistent decrease in loss on validation data proves that the model is well-adjusted and that the phenomenon of overlearning does not occur. Good results on validation data show that the model is capable of effectively working with new data, which is vital in urban applications where sound conditions may be highly diverse.

Another stage of the evaluation of the RNN model’s effectiveness is an analysis of the Classification Metrics and Confusion Matrix. The analysis included metrics such as precision, recall, and F1 Score. These indices help assess how well the model copes with the classification task.

In this case, the following results were obtained:

Precision: 0.96—This metric indicates that when the model predicted a case to be in the “overwhelming” class, it was correct 96% of the time. In other words, out of all the instances that the model identified as “overwhelming”, 96% truly belonged to this class. This high precision reflects the model’s ability to avoid false positives, ensuring that the majority of its positive predictions are accurate.

Recall: 0.93—This metric reflects the model’s sensitivity in detecting the “overwhelming” class. It shows that 93% of all actual “overwhelming” cases in the dataset were successfully identified by the model. While this is also a strong performance, it is slightly lower than the precision, indicating that the model missed some true cases of “overwhelming” (i.e., false negatives), but still managed to capture the vast majority.

F1 Score: 0.95—The F1 Score represents the harmonic mean of precision and recall, providing a single metric that balances both. With an F1 Score of 0.95, the model demonstrates a very high level of overall classification quality, effectively managing the trade-off between precision (minimizing false positives) and recall (minimizing false negatives). This score suggests that the model is both accurate in its predictions and comprehensive in identifying true cases of the “overwhelming” class.

Figure 14 represents the Confusion Matrix showing the number of correct and erroneous classifications, divided according to the actual and expected classes.

Analyzing the above matrix, the model achieved very high effectiveness, correctly classifying most cases. However, the small number of classification errors, particularly false positives and false negatives, may come from areas with mixed acoustic characteristics. Classification errors may occur in areas where peaceful sounds intertwine with overwhelming ones, for instance, on the borders of parks close to busy streets. Such cases pose a challenge for the model and may cause the few errors in the classification. Recognizing and understanding these cases is vital for further improving the model and developing its ability to differentiate more complex acoustic environments.

To ensure reliability and robustness of the model’s results, Cross-Validation was conducted. It is a technique for evaluating model efficiency, which involves dividing the data into several subsets (folds) and iteratively training and testing the model on various combinations of these folds. In this case, a 5-fold cross-validation was employed, meaning that the data were divided into five equal subsets. In each iteration, the model was trained on four given folds and then tested on the remaining fold. This process was repeated five times, with a different fold used for testing each time.

The results of cross-validation show that the model achieved high precision on each test fold, indicating its ability to generalize and its effectiveness in analyzing urban sounds. Precision for the individual folds is shown in

Table 2. Results of Cross-Validation.

Subset Number	Result
Fold 1	0.9314
Fold 2	0.9412
Fold 3	0.9608
Fold 4	0.9216
Fold 5	0.9412
Mean for all subsets:	0.9392

.

Cross-validation was conducted to ensure that the model is not excessively adjusted to the training data and that its effectiveness is satisfactory on unknown data. This method allowed for a more reliable quality assessment of the model, which is particularly important in the context of urban applications, where sound conditions may be highly diverse. The results of cross-validation prove that the model is well-adjusted and is capable of effective classification of sounds in different urban contexts.

3.4. Permutation Feature Importance

In the next stage of the study, Permutation Feature Importance was executed using the Sklearn library [52]. This technique is used to assess which input variables exert the greatest influence on the predictions of the model. In the context of the discussed RNN model, the application of this technique helps to understand which MFCC coefficients are the most important for the differentiation between quiet spaces and overwhelming ones. At the beginning, the basic precision of the model was calculated on the test data. This result serves as a reference point for comparing the results after feature permutation. Precision is measured using accuracy metrics, which evaluate the conformity of the model predictions with actual labels. Then, each feature is individually permuted. To this end, the values of a given feature are randomly reshuffled, which allows breaking its connection with the remaining features and labels. The model is reevaluated on the altered data, and the obtained precision is recorded. This process is repeated several times, in this case ten times for each feature, to obtain stable results. The difference between the basic precision and the mean precision obtained after the permutation of a given feature is the measure of its importance. The greater the difference, the stronger the influence of a given feature on the model’s efficiency.

Figure 15 shows the importance of specific MFCC features for the effectiveness of the classification of our recurrent neural network [52,53].

In the discussed RNN model, the first three MFCC coefficients (MFCC1, MFCC2, MFCC3) demonstrated the highest importance. They represent the largest amplitudes in the power spectrum of the sound signal, which reflects general characteristics of sound, such as its pitch and texture.

Why are they so important?

MFCC1: Represents the general power of the sound, which is a critical factor in differentiating between quiet and noisy environments. High power may indicate an overwhelming environment, while lower power is often connected with quiet spaces.

MFCC2, MFCC3: These coefficients help capture more subtle nuances in the sound profile, which may differentiate between types of environments, for instance, differences in the texture of sounds generated by traffic compared to the rustle of leaves.

To understand how sound features differ depending on the environments under consideration, it is necessary to create diagrams showing the variability of MFCC features over time, which allow for the visualization of their dynamics. For instance, in overwhelming spaces, we expect to see greater changes in MFCC1 values, reflecting increased sound intensity connected with traffic congestion. In contrast, in quiet spaces, MFCC1 values will be low and more stable, suggesting a less intense sound background. This phenomenon is evident in the diagrams of MFCC feature variability over time, as shown in Figure 16 for an overwhelming space—Manhattan, New York (a) and a quiet space in Stadtpark, Vienna (b).

Trends for overwhelming spaces are characterized by greater amplitudes and more frequent fluctuations of MFCC values, reflecting increased acoustic activity and a variety of sound sources. On the other hand, quiet spaces exhibit smaller fluctuations and lower values, which is typical of tranquil environments.

3.5. Field Study

Evaluation In order to verify the applied algorithm, a field study was conducted in the area of Wesola in Krakow. The study was aimed at analyzing the urban soundscape with the help of the previously trained RNN model. The study focused on identifying and classifying different types of urban environments based on their sound profiles recorded by two voice recording applications on Xiaomi Mi 10 (Xiaomi Corp., Beijing, China) and Sony Xperia XA2 Ultra (Sony Corporation, Tokyo, Japan) smartphones. The devices were selected to check independent results as well as compare and ensure repeatability of sound classification in different conditions [54]. Xiaomi used the system application “Dictaphone” (version 5.0.23.8) while Xperia used the Dictaphone application provided by Splend Apps (version 3.29). This approach allowed the researchers to analyze whether different sources of sound recording influence the results of classification. Using commonly accessible equipment instead of specialized measuring apparatus made it possible to check how an average user might experience the soundscape in different urban areas [55]. The measurements were conducted by a person walking along designated routes in the Wesola quarter of Krakow. They started at the Botanical Garden (marked by letter A), continuing to the Warsaw Uprising Avenue (Aleja Powstania Warszawskiego) (I), moving along Grzegorzecka Street (II), Blich Street (III), and Kopernika Street (IV), later turning towards the Market Square (V) and then in the direction of Mogilskie Roundabout (I). Each recording lasted one minute, which corresponds with the structure of the data used in the training of the neural network. The choice of this methodology allows for a direct comparison with the original training data, which were used to calibrate the model analyzing the urban soundscape. The map showing the routes where recordings were conducted as well as the aforementioned characteristic spots is shown in Figure 17. The letter B marks the conceptual visualization of the Green Empiria project placed on the map of the Wesola district.

The process of creating, analyzing, and formatting sound recordings started with converting the audio files into a format more suitable for analysis. The library ffmpeg was used to convert the files from .m4a and .aac formats into .wav format. The conversion was necessary because the .wav format is standard in sound analysis, offering lossless processing. After the conversion, extraction of sound features was performed using the Librosa library. Each audio file was divided into one-minute segments, and then MFCC features were extracted from them. This process allowed the sound data to be converted into a format suitable for further analysis using machine learning models. During data processing, file names were normalized to remove extensions and format differences, making comparison and identification of files easier. The normalized data were then used in the classification process, where each sound segment was analyzed and classified using the previously trained RNN model.

The unification of the process of converting the recorded audio files allowed for compatibility with the tools for sound analysis and ensured that all data were processed in the same way as the data used to train the recurrent neural network. These transformations resulted in obtaining uniform data consisting of a definite number of segments for particular locations. The obtained classification results depending on the analyzed location are presented in

Table 3. Results of classification of recordings conducted in the field study.

Category	Number of Segments	Number of Segments Classified as Quiet	Number of Segments Classified as Overwhelming
Botanical Garden	30	30	0
Grzegorzecka Street	10	0	10
Blich Street	10	2	1
Kopernika Street	3	10	6
Powstania Warszawskiego Avenue	16	0	16

.

All 30 segments recorded within the Botanical Garden were classified as peaceful (quiet). The consistency in the classification indicates that the Botanical Garden is an area free from intense traffic, offering seclusion and tranquility, which makes it a friendly and relaxing space for visitors. All ten segments recorded on Grzegorzecka Street were classified as overwhelming. This result reflects the high level of urban hustle and bustle and the intensity of traffic on this thoroughfare, which is typical of main city streets with intense pedestrian and road traffic. On Blich Street, three segments were recorded, one of which was classified as busy while the other two as quiet. This indicates diverse acoustic conditions in this location, where intervals with intense traffic as well as more peaceful moments can be observed. Kopernika Street showed more diversity in the classification results. Out of the 16 segments, 10 were classified as quiet while the remaining 6 were classified as overwhelming. Such results may suggest variable acoustic conditions in different parts of the street. This diversity serves as an example of how changing conditions can influence the perception of the urban soundscape. On the other hand, all 10 segments recorded on Powstania Warszawskiego Avenue were classified as overwhelming. This result reflects the intensity of traffic on this important Krakowian thoroughfare, which is consistent with its function as one of the main city routes with heavy traffic.

The field study results are related to the acoustic measurement spots identified in the Green Empiria project. The measurement at spot A, located near Kopernika Street, indicated variable acoustic conditions, which is consistent with the results from Kopernika Street, where both peaceful and overwhelming segments were recorded. This sound variability suggests the need for further optimization of the space for better noise management. Conversely, the measurement at spot E near Grzegorzecka Street reflected a high level of sound intensity, confirming the results from this thoroughfare classified as overwhelming. These results indicate the need for additional sound-absorbing solutions and noise reduction strategies in these areas to effectively implement the objectives of the Green Empiria project regarding the creation of an acoustically safe enclave of silence in the central part of the modernization project of this urban space.

The classification results obtained from two different sound-recording devices—Xiaomi Mi 10 and Sony Xperia XA2 Ultra—were also compared. The goal was to assess whether the type of device and the software used would influence the classification results. The analysis shows that the total number of sound segments for each device was identical, amounting to 69 segments for each. Of those segments, 27 were classified as quiet while 42 were recognized as overwhelming, regardless of the device used. Importantly, no differences in segment classification were observed between the devices, suggesting that neither the type of recording device nor the software used had any impact on the neural network’s classification results.

4. Conclusions

The article discusses the results of projects carried out by interdisciplinary teams consisting of students and scientists from four Cracovian higher education institutions. This paper has defined the Care Spaces, and discussed the area that was selected for revitalization. The paper also discusses the projects that were executed during the workshops. The presented projects are in accord with the 2030 Agenda for Sustainable Development. We also took into account the conception of Baukultur, the approach that aims at creating well designed environments which are supportive to health and well-being of people and other creatures, and respects cultural aspects in design and construction.

One of goals of the research carried out by the team of academics, including data science engineers, was integrating the soundscape analyzed in the project context with that of the actual area The studies regarding soundscape design with the use of Recurrent Neural Network showed that it is possible to monitor and classify urban soundscapes based on human affective perception of sound. Using the features of Mel-Frequency Cepstral Coefficients (MFCC) to analyze one-minute segments of recordings, the authors managed to precisely distinguish quiet zones from overwhelming ones. The model of RNN using recurrent and dropout layers demonstrated high accuracy both on the training data and on validation data. The results of cross-validation confirmed the model’s ability to generalize, which is vital for its application in actual urban conditions. The analysis of permutation of feature importance showed that the MFCC1, MFCC2, and MFCC3 coefficients exert the greatest influence on classification, which helps to understand what sound features are the most important in the identification of quiet spaces and overwhelming ones. Field studies in Wesola quarter in Krakow confirmed that the RNN model can be effectively used in different urban conditions.