Leveraging Recurrent Neural Networks for Lithology Identification and Chinese Rural Landscape Planning in Sustainable Design

Abstract

This paper explores the integration of ecological sustainability, human-centered design, and advanced computational techniques, with a particular focus on the use of Long Short-Term Memory (LSTM), Recurrent Neural Networks (RNNs), and Deep Recurrent Neural Networks (DRNNs) in urban landscape planning and rural landscape restoration. DRNNs, an advanced extension of traditional RNNs, are specifically designed to capture complex temporal dependencies in sequential data through deeper network architectures. These models are particularly effective in identifying intricate patterns in time-series data, making them well-suited for dynamic processes in landscape planning and ecological analysis. The study highlights the significance of applying ecological principles to urban design, aiming to create spaces that are not only visually appealing but also environmentally harmonious and socially inclusive. Additionally, the research investigates the role of installation art in public urban spaces, emphasizing its potential to foster community interaction, raise environmental awareness, and promote sustainability. By integrating data-driven approaches, such as LSTM-based lithology identification and DRNN-based ecological forecasting, the paper illustrates how advanced algorithms can optimize landscape features, predict ecological trends, and guide more informed planning decisions. Ultimately, this research underscores the need for a holistic and sustainable approach in urban landscape design that balances environmental, social, and technological dimensions, ensuring a harmonious coexistence between people and their environments.

1. Introduction

Chinese rural classical gardens have long been intricately intertwined with the field of art, fostering a dynamic interaction between these two domains [1,2]. In contemporary landscape architecture, particularly in the design of classical Chinese rural gardens using recurrent neural network (RNN)-based methodologies, this connection remains evident, albeit in a context that now integrates various multidisciplinary approaches [3,4]. Despite the advances in technological tools, the influence of artistic fields, particularly through rapid innovations, continues to permeate garden design. Whether characterized by the profound, subtle introspection of traditional gardens or the minimalist, ecologically focused aesthetics of modern landscapes, both approaches bear an enduring relationship with diverse artistic schools [5]. These schools, evolving through their own artistic paradigms, concurrently shape and redefine the practices within landscape architecture [6,7].

The proliferation of surrealist art, for instance, has notably influenced landscape design, including the work of the distinguished American landscape architect Thomas Church. Church’s integration of Cubism and Surrealism into his landscape designs—employing simple, flowing forms and an intuitive balance between geometry and nature—exemplifies the transformative impact of art movements on landscape architecture. His designs illustrate how surrealist principles can seamlessly blend dynamic visual forms with the natural environment [8]. Similarly, Brazilian landscape designer Braemax highlights the interconnectedness of art and landscape. He argues that the distinction between landscape design and art is largely semantic, as both share a common artistic impulse. In his practice, Braemax employs fluid, organic forms and utilizes large masses of flowers of uniform size to create expansive color fields, resulting in landscapes that evoke the earth itself rather than a mere canvas [9].

Turning to the realm of computational methods, the application of recurrent neural networks offers a promising avenue for advancing classical garden design. Unlike traditional feedforward neural networks, RNNs are specifically designed to process and retain sequential information, enabling the analysis of dynamic data over time [10,11]. This capability makes RNNs particularly effective in handling complex, sequential tasks where context and flow are paramount. An advanced variant of RNNs, Long Short-Term Memory (LSTM) networks, addresses common issues such as vanishing gradients and the degradation of long-range dependencies in sequence learning [12,13]. By maintaining memory over extended periods, LSTMs enhance the ability to process longer and more intricate sequences, thereby optimizing the performance of models used in design tasks that involve temporal or sequential data [14,15].

Recent advancements in the application of RNNs and other deep learning techniques in lithology recognition and rural landscape planning in China have yielded promising results [16,17]. In lithology recognition, methods such as RNNs and Convolutional Neural Networks (CNNs) have been utilized to analyze seismic, borehole, and remote sensing data, facilitating the identification of subsurface lithological properties [18]. However, challenges persist, including the reliance on large labeled datasets, limited model interpretability, and difficulties in handling complex geological conditions. In rural landscape planning, deep learning models are applied to predict land use changes, assess landscape features, and support cultural heritage preservation. While these models hold considerable potential, issues related to data quality and consistency—particularly in rural areas—pose significant barriers. The high regional variability in landscapes further complicates model application, requiring adjustments to account for local ecological, cultural, and socio-economic contexts. Additionally, integrating diverse data sources remains a persistent challenge.

2. Research Background

The rapid urbanization occurring in China has catalyzed increasing attention to the design and optimization of classical Chinese rural landscapes within the fields of planning and architecture [19,20]. Notably, rural geoscape has gained considerable prominence in recent years, reflecting a broader cultural and environmental shift. Rural landscapes are not only integral to the livelihood of China’s population, but they also serve as focal points for public interest and concern. These landscapes are multifaceted, encompassing both objective elements—such as geographical features, topography, hydrology, weather, fauna, flora, and anthropogenic structures—and subjective factors, including the socio-economic development, cultural values, and customs of the local population [21]. As such, the rural landscape is a dynamic construct, shaped by natural processes and human influence over time. The study of rural landscapes, therefore, is inherently linked to the concept of rural habitat, with rural landscape gardening emerging as a central topic in this research [4,22].

Urbanization planning based on village landscape gardening plays a pivotal role in safeguarding rural habitats while fostering sustainable development [23]. This planning framework influences the growth and character of rural settlements, guiding the establishment of distinctive architectural styles and the preservation of the natural environment [24]. The aesthetic, scale, and functionality of rural gardens are vital components that determine the resilience and sustainability of these communities. A well-designed rural park, while seemingly simplistic in form, often embodies complex principles such as “adaptation to local conditions”, “material usage in accordance with local availability”, and “sustainable development”. These gardens are rich in ecological and cultural values, which resonate with urban dwellers who are increasingly drawn to the natural charm of rural landscapes.

Amidst the accelerated urbanization of China, rural gardens are assuming an increasingly critical role in shaping the urban-rural interface and the overall quality of living environments. For rapidly urbanizing rural communities, there is an urgent need for strategic planning to define their identity and guide their development [25]. As many rural areas undergo a process of urbanization and potential degradation, the introduction of thoughtful landscape architecture can serve as a transformative tool, reinvigorating these spaces and preserving their cultural essence. The construction of modern living environments in rural areas, driven by the high-speed urbanization of China, aims not to turn rural landscapes into urban replicas but to preserve and enhance the quality of life for rural populations. This process, however, comes with challenges: while some historic rural gardens are being lost to urban expansion, new forms of rural landscape gardens are emerging. These new designs must strike a balance between honoring traditional folk culture and embracing modernity, ensuring that the new rural landscape is distinct, innovative, and sustainable.

While extensive literature on Chinese garden theory exists, both domestically and internationally, research into traditional Chinese rural gardens often combines literature review with empirical investigation. This approach involves examining historical records, local chronicles, and genealogies, which provide insights into the selection, layout, and development of rural gardens. As rural landscape design relies on comprehensive field research, it is necessary to conduct large-scale empirical investigations, analyzing the social, economic, and cultural dynamics that shape these landscapes [26]. Traditional rural gardens, which are deeply embedded in the natural landscape, are less constrained by external artificial norms, making them ideal candidates for in-depth site analysis and mapping. Moreover, understanding the local populace’s lifestyle, values, and resource demands is crucial for creating contextually appropriate designs.

In the context of modern urbanization, rural landscape architecture is also influenced by village scale standards and land-use regulations, necessitating a nuanced understanding of relevant policies and guidelines. By conducting systematic research into both traditional and contemporary rural landscapes, this study seeks to identify the fundamental changes in rural landscape architecture’s form and meaning. Through comparative analysis, the research aims to categorize the basic types of rural landscapes and elucidate their essential characteristics. The study also seeks to uncover underlying patterns in rural garden design, advancing the scientific understanding of rural landscape architecture.

It is important to note that the scientific inquiry into rural landscape architecture cannot be confined to a single discipline. It requires an interdisciplinary approach, drawing on landscape architecture, ecology, geography, urban planning, and architecture. However, the theoretical foundations of rural landscape architecture remain in the developmental stages and require further refinement. Unlike urban gardens, which often replicate nature through artificial means, rural gardens are inherently tied to the natural landscape, preserving its authenticity. Aesthetic principles in rural gardens diverge from those in urban settings—while urban gardens may be designed to appear natural, rural gardens tend to evolve organically from the landscape itself, reflecting a harmonious balance between human intervention and nature. From a natural beauty perspective, rural gardens surpass urban gardens in their unspoiled aesthetic qualities. By studying these rural landscapes, we can derive invaluable insights into sustainable, context-sensitive design practices. Despite this, landscape planning and design in this field remain underdeveloped in China, making the study of rural gardens not only timely but also of profound significance.

2.1. Classification of Disseminators of Intangible Cultural Heritage Videos

A hierarchical structure model was developed using hierarchical analysis software to evaluate the landscape construction and aesthetic qualities of rural village settings. This model relied on a comprehensive questionnaire assessing the ecological aesthetics and other relevant criteria, from which the weights of various data points were derived. The results revealed that, at the macro level, village folk customs accounted for the largest proportion, representing 29% of the total weight. This finding underscores the importance of preserving traditional rural elements—such as landforms, mountains, water systems, and green spaces—within the context of rural landscape planning and construction. The primary objective is to maintain the integrity of the landscape while minimizing the frequency of natural disasters and preserving surface water environments, particularly in ecologically sensitive areas. These measures are essential for ensuring the ecological foundation of rural landscapes, optimizing the structure of rural ecological spaces, and fostering an ecological aesthetic that is both sustainable and resilient, as depicted in Figure 1.

Following folk customs, the next most significant categories were “the pursuit of the beauty of life and the extension of survival struggles” and “environmental security ecosystems”, accounting for 22% and 20%, respectively. These categories highlight the interrelationship between villagers’ understanding of their environment and their role as both creators and users of rural landscapes. As key stakeholders in rural revitalization, villagers’ contributions to the design and improvement of their surroundings are vital for the successful implementation of sustainable rural development initiatives. Both of these categories are aligned with the broader goal of high-level rural landscape planning, which aims to strengthen the operational foundations of rural industries and improve the living standards of rural populations through efficient industry management. This, in turn, contributes to the long-term sustainability of rural revitalization efforts.

The findings from the analysis of the criterion layer corroborate previous literature and underscore the practical applicability of the data. Furthermore, while planning and design are often influenced by the designer’s subjective perception, these personal factors are most effective when grounded in reasoned, legally sound principles. In this context, the integration of both objective ecological data and subjective cultural insights creates a comprehensive framework for the sustainable development of rural landscapes, ensuring that both environmental and social dimensions are addressed.

Lithology identification and classification are integral to urban landscape design, particularly in understanding the soil composition and geological features of the area. These geological insights inform key decisions regarding plant selection, land usage, and the overall design of the landscape. By identifying the specific lithological characteristics, designers can choose the most suitable vegetation and materials, ensuring that the landscape is not only aesthetically pleasing but also sustainable and aligned with the natural environment. Furthermore, lithology data helps predict how the land will react to various environmental factors, such as moisture and temperature fluctuations, which is essential for creating resilient and functional urban spaces. This approach supports more informed, ecologically balanced, and context-specific landscape planning.

2.2. Advancements in Lithology Identification Using Machine Learning and LSTM Networks

Accurate lithological understanding is crucial for the precise characterization and comprehensive evaluation of complex carbonate reservoirs [27]. Lithological and structural information of subsurface strata is typically obtained through methods such as drilling, core sampling, cuttings logging, and borehole wall coring. However, due to the high costs associated with drilling service coring and the relatively low accuracy of cuttings logging, significant attention has been directed toward the development of key technologies for lithology identification, particularly in exploration wells. The lithological classification of exploration wells is generally based on the mapping relationships between well parameters and lithological types. These relationships are then applied to predict the lithology of unsampled wells.

The advent of statistical learning, machine learning, and pattern recognition has led to the widespread use of mathematical theories and computational algorithms in lithology identification [28]. Techniques such as principal component analysis, decision trees, support vector machines (SVM), Naive Bayes, self-organizing maps (SOM), and fuzzy logic have been employed in the construction of lithology identification models. These approaches typically assume that lithology and its associated parameters are independent of each other in depth, often neglecting the spatial and temporal correlations of rock strata during sedimentation and diagenesis. This limitation is particularly notable when depth series fail to capture the geological characteristics of rock sequences or their compositional variations.

Elfeki et al. [29] utilized Markov chain theory to model the early characteristics of lithological sequences, representing them through transition probability matrices between different lithological classes. Hidden Markov Models (HMMs) have also been widely applied, as they combine the sequential correlation of lithology with measurement methods at a deeper level. However, these approaches still fall short in accurately delineating the transformation sequence of lithological categories and capturing the intricate relationships between different strata sequences.

Recurrent Neural Networks, particularly LSTMs, offer a more robust solution for lithology identification by preserving the sequential dependencies within the data. The self-loop structure inherent in RNNs and LSTMs allows these models to capture the internal accumulation of lithological structure and the scale of lithology determination under various exploration parameters. LSTMs, in particular, address key issues present in standard RNNs, such as the vanishing and exploding gradient problems, thus offering a more effective framework for lithological classification and prediction in exploration wells. The recursive nature of LSTM networks makes them an ideal tool for modeling the sequential and spatial correlations inherent in geological data, enabling more accurate and reliable lithology identification.

3. Materials, Methods and Object

3.1. Research Object

Currently, China’s agricultural development model remains relatively simplistic, with newly built towns often lacking distinctive features, resulting in monotonous and uniform landscapes. In many regions, the pressure exerted by local cultural resources on the natural landscape is intensifying, leading to the gradual erosion of traditional regional scenic spots. This issue is exacerbated by a lack of scientific and comprehensive urban planning, resulting in chaotic urban landscape layouts. The prevalent “Fordist” model of working and living, characterized by a focus on “modernization”, has further impacted the fundamental material environment in which people live [30]. Presently, the dominant values of human existence—solidarity, norms, and mass production—have contributed to a crisis of identity, homogeneity, and a lack of diversity, which is undermining the cultural and ecological fabric of rural communities.

Although most villages and towns in China have developed master plans, these plans are often of substandard quality, as many villages and towns follow the same design principles. Architectural design tends to prioritize function over aesthetics, resulting in an absence of individuality. With the acceleration of urbanization in China, cities are expanding and consolidating rapidly, leading to the disintegration of rural areas and the shrinkage of regional cultures. The relationship between regional urban centers and surrounding rural areas has shifted from a model where rural areas surround cities to one where cities (and parks) encircle rural areas. This transformation requires a careful consideration of how these two entities can achieve coordination and compatibility.

For a long time, China has adhered to the design principle of “economy, practicality, and aesthetics” in rural construction. Unfortunately, many rural construction projects have resorted to imitating urban housing or villas, often placing excessive emphasis on the functionality of buildings at the expense of regional architectural characteristics and individuality. This phenomenon, epitomized by the so-called “matchbox” architecture, became a symbol of low-cost, adaptable general spaces, yet failed to incorporate distinctive cultural or environmental elements [31].

The disconnect between rural housing and actual agricultural production is another pressing issue. China’s rural agricultural production remains relatively underdeveloped, with many areas still relying on traditional farming practices. This creates a significant disparity between the functional requirements of rural houses and those of urban homes [32]. As part of the new rural construction initiatives, it is essential to integrate rural housing with functional spaces such as storage yards, solar farms, and other production-related functions, maximizing agricultural productivity and work efficiency. However, the prevalent use of urban architectural models for rural construction projects has exacerbated the disconnection between rural production methods and the distinctive survival modes of traditional agriculture. Furthermore, new rural construction projects often face severe environmental and landscape challenges, lacking cohesive, scientifically grounded planning. This has resulted in imbalanced rural landscapes, fragmented natural environments, and the inability to foster healthy rural ecosystems. Over-exploitation of resources, particularly in large-scale environmental remediation efforts such as land leveling or the filling of lakes, has become a critical issue.

Human attempts to transform nature into a mechanized, controlled environment have often overlooked the inherent value of the land’s topography, landforms, flora, and fauna. These natural resources form a relatively stable ecosystem that has developed over time, offering significant ecological benefits. Disrupting these natural systems often leads to detrimental effects on both human health and environmental sustainability. From an urban planning perspective, the principle of “anti-planning” could be a more effective approach, wherein the design of rural landscapes prioritizes the preservation of unbuilt communities and green spaces. This principle would enforce strict controls on the scope of development and renovation, ensuring the protection of valuable ecological assets within rural areas.

3.2. Research Methods

3.2.1. Modeling Temporal Dependencies with RNNs

Recurrent neural networks, as a class of deep neural networks, are particularly suited for capturing temporal feedback loops within data. In the context of rural landscape and environmental planning, RNNs can be employed to model the complex, sequential relationships inherent in rural development and ecological systems. These networks, particularly when folded in time, can function as deep neural networks with infinite layers, effectively storing and transmitting context-sensitive information across multiple updates. By leveraging hierarchical structures and time-based feedback, RNNs ensure that each layer in the network retains critical data from previous iterations, enhancing the model’s ability to process and analyze complex, time-dependent patterns [33].

As shown in Formula (1):

$$T_i=W_{hx}{X}_i+W_{hh}{h}_{i-1}+b_i$$

(1)

As shown in Formula (2):

$$h_i=e\left(t_i\right)$$

(2)

As shown in Formula (3):

$$S_i=W_{yh}{h}_i+b_y$$

(3)

As shown in Formula (4):

$$y_i=g\left(s_i\right)$$

(4)

In Equations (1)–(4), the following variables are defined: i represents the index of the current input or time step in the sequence, indicating the position of the data point being processed. T denotes the total number of time steps or the length of the sequence being fed into the network. S refers to the system’s state at a particular time step, which could represent either the hidden state or the memory cell state, depending on the context of the equations. y is the output at a specific time step, representing the prediction or result generated by the network at that point in the sequence. Lastly, s refers to the set of weights or parameters in the network that govern the transformations of the hidden states and the system’s update mechanisms. These variables are essential to the recursive nature of the RNN and LSTM models, where each time step involves the interaction of input data, system state, and weights to produce the final output. The variable t represents the input to the output unit, which is a k-dimensional vector derived from the current time step’s hidden state. This input plays a crucial role in generating the final output of the model. The variables E and G are pre-defined nonlinear vector-valued functions applied to the hidden states and input data. These functions introduce the necessary nonlinearity to the model, allowing it to capture complex patterns and dependencies within the data. By utilizing these functions, the model can better process and interpret the sequential data, leading to more accurate predictions.

3.2.2. LSTM Networks for Sequential Modeling and Lithology Identification

Traditional Recurrent Neural Networks often suffer from issues such as gradient explosion and vanishing gradients, which can significantly impair the efficiency of network predictions over long sequences. To address these limitations, Hochreiter et al. [34] introduced Long Short-Term Memory networks, which were later enhanced by Graves et al. [35]. These networks incorporate gating mechanisms such as the forget, input, and output gates, which regulate the flow of information and enhance memory retention. This enables LSTMs to capture long-term dependencies more effectively. The ability of LSTMs to process long sequences of data has made them highly effective for tasks such as time-series forecasting and natural language processing. However, LSTM models still encounter challenges, particularly the issue of vanishing gradients in long sequences. This problem is mitigated through techniques such as gradient clipping and the use of advanced optimizers. Further advancements by Gers et al. [36] led to the development of Gated Recurrent Units (GRUs), a variant of LSTM that reduces model complexity while maintaining comparable performance in many cases. These improvements have addressed several of the shortcomings of RNNs, particularly their difficulty in modeling long-range dependencies in sequential data.

LSTM networks, a specialized type of RNN, have garnered considerable attention for their ability to model and predict sequential data. Unlike traditional RNNs, LSTMs are specifically designed to address the vanishing gradient problem, enabling them to capture long-term dependencies within time series data more effectively. This capability is made possible by their unique architecture, which includes gates—such as the forget gate, input gate, and output gate—that control the flow of information. These gates allow LSTMs to retain important information over extended periods while discarding irrelevant data, making them particularly well-suited for tasks involving sequential or temporal data. Applications such as time series forecasting, natural language processing, and even lithology recognition or landscape planning benefit from LSTM’s ability to model complex patterns in data over time. As a result, LSTMs are increasingly being adopted across various fields where recognizing temporal patterns is crucial [37].

In an LSTM network, each time step involves the interaction of the cell state and hidden state across multiple gates. The input gate at a given time step receives inputs from the previous hidden state and the current input, which are combined to update the memory cell. The forget gate determines which portions of the previous memory should be retained or discarded, while the input gate regulates the flow of new information into the memory. Subsequently, the output gate filters the updated memory cell to generate the hidden state for the current time step, which is used for the final output. This structure enables LSTMs to maintain and propagate relevant information over long time intervals, addressing the vanishing gradient problem and significantly improving the network’s ability to capture temporal dependencies. The flexibility and effectiveness of LSTM units make them a powerful tool for sequence prediction tasks, allowing for the preservation of important information across time while mitigating the issues that limit traditional RNNs. Through this architecture, LSTMs are capable of learning more robust and accurate representations of sequential data, facilitating their widespread application in various domains, from natural language processing to time series forecasting [38].

The development of the Long Short-Term Memory lithology identification model is based on the inherent complexity of LSTM networks, particularly due to the sequence dimension in their structure. Unlike traditional recurrent networks, LSTMs do not require excessive stacking of recurrent layers to capture sequence dependencies effectively. The first step in the process involves selecting log parameters that are sensitive to lithology, which are then standardized using mean-variance normalization to eliminate dimensional influences. Concurrently, a thermal encoding technique is employed to digitize lithological data. The lithology identification model incorporates LSTM layers in conjunction with fully connected layers, as illustrated in Figure 2, which outlines the data flow for the encoding process. In this process, Figure 2 illustrates how the data flows through the encoding mechanism of the model, where LSTM layers play a critical role in handling sequential information. The LSTM layers are designed to preserve long-term dependencies, which is particularly important in tasks like lithology identification where historical data points are crucial for accurate predictions. These layers are coupled with fully connected layers, which help process the information by transforming the LSTM outputs into the final classification results. The integration of LSTM layers ensures that both short-term and long-term dependencies are effectively captured, improving the model’s ability to analyze complex, time-dependent data and make more precise predictions. This approach allows the model to maintain relevant information over extended sequences, addressing common challenges like vanishing gradients found in traditional RNNs. The overall architecture thus enhances the predictive power of the model, making it more robust and effective in identifying lithological patterns.

3.2.3. Multi-Layer DRNN Architecture for Sequential Data

The DRNN model is built upon a multi-layer architecture of RNNs, which significantly enhances its ability to capture complex, nonlinear relationships in sequential data. By stacking multiple RNN layers, DRNNs increase the model’s depth, enabling it to learn hierarchical representations of time-dependent features. This deeper architecture allows DRNNs to better capture intricate patterns and long-term temporal dependencies compared to traditional RNNs, making them particularly well-suited for tasks involving complex and dynamic sequential data. Additionally, the increased number of layers enables the model to store and process more parameters, thereby improving its capacity to generalize from larger datasets. As illustrated in Figure 3, this multi-layer structure enhances the DRNN’s effectiveness in modeling complex, dynamic systems with nonlinear interactions.

In the input layer of the model, the variable v represents the input vector at each time step. It is a k-dimensional vector that contains the data features fed into the network. These features are processed through the network to update the hidden states and generate the output. The input vector v is crucial for feeding information into the model, which then undergoes transformations through weight matrices $W_{hx}$ and $W_{hh}$ as outlined in Equations (1)–(4). The interaction between v and the weight matrices helps update the hidden state and compute the output at each time step, making it an essential component for the model’s functionality.

Moreover, the system design for the application should align with specific objectives and requirements. A unified labeling system is essential for categorizing short video content for user recommendation. The algorithm must select relevant labels from a comprehensive label library, ensuring that content is accurately described. Given that manually defined labels, such as “funny” or “humorous”, cannot be discerned from visual or audio cues alone, more specific, behavior-related labels should be established. Furthermore, the labeling system must account for the multi-dimensional nature of video content, which cannot be fully encapsulated by a single label. Therefore, videos should be described with multiple labels to provide a comprehensive understanding of their content.

The system must efficiently preprocess video data by segmenting it into frames. Unlike geophysical logging or borehole imagery, the video data used for the CNN model consists of pre-processed frames, specifically selected to assess the model’s ability to handle video data relevant to landscape planning. These frames are chosen for their capacity to capture key features that align with the model’s objectives, as well as to reflect the critical temporal and spatial dynamics required for the task.

Given that adjacent frames often exhibit considerable repetition, it is unnecessary to process every single frame. Instead, a subset of eight frames is selected to represent the overall content of the video. The separation points for these frames are determined based on the video’s temporal characteristics. The video is divided into meaningful segments to ensure that each subset captures a relevant portion of the sequence, while maintaining manageable data sizes. This strategy enables the model to focus on the most significant frames without losing important contextual information.

Once the frames are selected, the CNN extracts both spatial and temporal features. This approach improves the model’s ability to accurately classify video content, establishing a robust framework for video data analysis. The framework is designed to address the complexities inherent in both the temporal and spatial dimensions of video, ensuring that the model can effectively analyze the dynamics of the video while preserving its ability to perform detailed, context-sensitive analysis.

To improve the interpretability of our machine learning models, we performed an analysis using SHAP (SHapley Additive exPlanations) values, which provide insights into how each feature contributes to the model’s predictions. SHAP values offer a unified measure of feature importance by attributing each prediction to the individual features in a fair and interpretable way. This method is particularly useful for complex models like DRNNs and LSTMs, as it helps identify which variables have the most significant impact on the model’s decision-making process. By using SHAP values, we were able to visualize the contribution of each input feature, such as environmental factors and urban planning variables, to the prediction of lithology classification and landscape optimization. This transparency allows for better understanding and trust in the model’s predictions, enhancing its applicability in real-world decision-making. The SHAP analysis revealed that certain features, such as soil composition and air humidity, had a more significant influence on the model’s output, which was consistent with prior knowledge of ecological dynamics in landscape planning.

3.3. System Flow

Artificial neural networks (ANNs) are capable of automatically learning the nonlinear relationships between input and output data by constructing hierarchical structures [39]. The backpropagation (BP) neural network, one of the most widely used types of artificial neural networks, typically consists of three layers: an input layer, one or more hidden layers, and an output layer [40]. The BP network’s architecture involves neurons within each layer being fully connected to all neurons in the subsequent layer, while there are no connections within the same layer. Neurons in the hidden layers receive linear signals from the neurons in the preceding layer, which are then processed through activation functions before being passed to the next layer. This forward and backward propagation allows for iterative adjustments of the network weights, optimizing the network’s performance. However, the BP neural network’s ability to process sequential data of varying lengths is limited, as the network typically outputs a fixed-size result, which can hinder its effectiveness in tasks that involve time-dependent or sequence-based information.

In contrast, recurrent neural networks, which connect the internal nodes of each layer across time dimensions, can effectively learn and process sequential data. By maintaining temporal feedback loops, RNNs capture dependencies across time, making them highly suited for tasks that involve time-series data or sequential learning [41]. As depicted in Figure 4, the Long Short-Term Memory model, a variant of RNN, addresses the challenges posed by sequence complexity by structuring the network in a way that eliminates the need for excessive stacking of layers. Initially, the LSTM model utilizes a preprocessing step where logging data parameters, particularly those sensitive to lithology, are standardized using mean-variance normalization to mitigate dimensional effects. Concurrently, lithology data is encoded using a thermal coding technique, creating a digitized dataset. This digitized data is then fed into a hierarchical LSTM network for lithology identification. RNNs and LSTM networks are both used to process sequential data, but they differ in how they handle long-term dependencies. RNNs work by maintaining a memory of previous inputs, making them suitable for tasks like time-series forecasting. However, RNNs often struggle with a problem known as the vanishing gradient, where their ability to learn from long-range relationships decreases over time. LSTMs, a refined version of RNNs, address this by using gates—such as the forget, input, and output gates—to control the flow of information, allowing the network to selectively remember or discard data over longer sequences. This structure enables LSTMs to manage long-term dependencies more effectively, making them better suited for complex tasks involving sequential data, such as landscape planning and time-series analysis.

The original input data to the RNN shown in Figure 4 consists primarily of time-series data, including environmental and geographic parameters relevant to landscape planning. These data points may include factors such as soil composition, temperature, humidity, and precipitation, all of which change over time. This time-dependent data enables the RNN to model relationships between different time steps and predict future conditions or identify trends. The data is pre-processed to ensure it is suitable for training the RNN and generating accurate predictions based on sequential patterns.

Given the multifaceted optimization challenges in landscape design, the cross-entropy loss function is adopted to evaluate the learning outcomes of the network. Additionally, the Adam optimizer, which adapts the learning rate during training, significantly accelerates the convergence of the network compared to traditional stochastic gradient descent methods. As illustrated in Figure 5, this adaptive optimization strategy enhances the efficiency of the training process, leading to more rapid and stable learning outcomes. In Figure 5, the flow map is intended to illustrate the general structure of the model. However, it is important to note that the final LSTM layer plays a crucial role in the model’s prediction process. The output from the final LSTM layer is used to generate predictions, which are then compared to the known results (from the input data) to calculate the error. This error is used to update the model parameters during training, ensuring that the model learns from the differences between predicted and known results.

In classical rural landscape design, data analysis highlights four key variables—air humidity, soil composition, precipitation, and land area—that play a significant role in shaping design characteristics. Figure 6 illustrates how these variables respond to changes in environmental factors, with each influencing design attributes in different ways. Natural gamma-ray noise is particularly valuable in differentiating parameters like soil composition and land area. By revealing landscape features that may not be visible through other methods, gamma-ray data helps identify more distinct land areas, with higher values correlating to more easily recognizable features. While precipitation has a smaller impact on gamma-ray noise, it still provides useful information for classification. Additionally, acoustic time differences in air moisture and the photoelectric absorption cross-section index further contribute to the analysis of landscape features. Acoustic time differences provide insights into air moisture levels, which can influence design characteristics, while the photoelectric absorption cross-section index helps distinguish features related to soil composition and moisture. Together, these factors deepen the understanding of landscape conditions, enhancing the model’s predictive power. Figure 6 also shows how outliers, especially those caused by noise points, can distort extreme values. To mitigate this, the 90th and 10th percentiles are used to define the lithologic structural response range, improving data accuracy. This approach, integrating gamma-ray noise with other environmental variables, offers a comprehensive framework for landscape design analysis.

In Figure 6, AR refers to the “Acoustic Response” and GR refers to the “Gamma-Ray” values. The y-axis represents the measured response values derived from both the acoustic and gamma-ray data. These values provide insights into landscape features based on environmental parameters such as air moisture and soil composition. Outliers in the data are identified as points that significantly deviate from the rest of the dataset, often due to noise or measurement errors. These outliers are typically located at extreme values compared to the majority of the data points. To address these discrepancies, the 90th and 10th percentiles are used to define the upper and lower bounds for the lithologic structural response range, ensuring that the data representation remains accurate and consistent. This approach allows for a more precise analysis of the environmental influences on landscape design.

The number of epochs, which represents the total number of complete iterations over the training dataset, plays a critical role in the model’s performance. Too few epochs can result in underfitting, where the network fails to capture the essential patterns in the data, leading to poor predictions. On the other hand, too many epochs can lead to overfitting, where the model memorizes the training data but struggles to generalize to new, unseen data [42]. Cross-entropy loss functions are used to evaluate the network’s performance during training, with the loss decreasing significantly as the number of epochs increases, indicating faster learning. The cross-entropy loss function is commonly applied in classification tasks to measure the discrepancy between the true labels and the predicted probability distribution. It quantifies how far the predicted probabilities are from the actual labels, and is defined as:

$$L=-\sum _{i=1}^n{y}_i\mathit{log}\left(pi\right)$$

where $y_i$ represents the true label (0 or 1), and $pi$ is the predicted probability for each class. As the number of epochs exceeds 1000, the loss curve flattens, suggesting that the network has adequately learned the relevant features of the data. This process helps ensure that the model achieves an optimal balance between underfitting and overfitting.

Batch size, which refers to the number of samples processed before updating the network’s weights, also affects the efficiency of training. Due to the large volume of data required for deep learning tasks, small batch sizes are typically used. This approach ensures that each iteration of training is computationally manageable and facilitates more efficient learning. However, very small batches can introduce noise into the learning process, while excessively large batches require more epochs to converge. In this study, the optimal batch size was found to be 32, as it yielded the highest accuracy in rock recognition, with accuracy declining as the batch size increased beyond this threshold. The time-step, which refers to the number of past data points used to predict the current time step, is another key parameter in the training process. A small time-step may result in insufficient data for accurate predictions, as the network would only use information from nearby time steps, ignoring the broader temporal context. On the other hand, excessively large time-steps can introduce irrelevant data, increasing training time without improving performance. Through careful evaluation, it was found that a time-step value of 4 yielded the best results, with accuracy stabilizing around 96% as time-step values increased further. This optimal time-step value, as demonstrated in Figure 7a, ensures that the network efficiently captures the sequential relationships within the data while maintaining robust prediction accuracy.

In Figure 7b, while the cross-entropy loss does not show a sharp decrease, the decision to use 1000 epochs as the optimal number is based on the trend observed in the loss curve. After 1000 epochs, the curve levels off, indicating that the model has effectively learned the key features of the data. This plateau suggests that additional training would not lead to significant improvements in performance. As a result, 1000 epochs were chosen as the optimal number, balancing training efficiency and model effectiveness.

4. Discussion

4.1. Principle of Ecologically Based Sustainability

The ecological principle advocates for alignment with the natural laws governing ecological change, encompassing the fundamental laws of ecological development, balance, and engineering, as well as comprehensive optimization of the natural environment [43]. This principle is rooted in a “scientific and reasonable arrangement according to local conditions”. In the context of urban landscape design, this translates into a holistic approach where the interactive functions of installation art are considered within the broader landscape environment, striving for cohesion and unity in the landscape’s stylistic design. The spatial placement of urban landscape installations must prioritize user accessibility and convenience, ensuring integration with the surrounding natural environment and aligning with the main activity zones of the city, such as streets and squares. The functional design of urban installations should be tailored to the diverse needs of the public, enabling more effective utilization of urban public spaces. This can be achieved by leveraging various design techniques that provide conducive spaces for both necessary and discretionary public activities. The infusion of ecological thinking into landscape design has significantly expanded the conceptual and practical frameworks of urban planning. Landscape design has transcended its traditional boundaries and now intersects with a variety of disciplines and broader environmental concerns. Ecological design extends beyond the simple planting of trees and grasses to include critical factors such as air quality management, energy utilization, water resource collection and reuse, and waste disposal—each of which profoundly influences the overall ecological integrity of urban landscapes. Sustainable urban construction, therefore, revolves around creating a harmonious relationship between human populations, the environment, and future generations. In the context of classical rural landscape design and renovation in China, it is imperative to treat the environment as the fundamental aspect of the design process. Economic development and natural preservation must coexist, ensuring that both ecological and cultural resources—viewed as irreplaceable assets—are not depleted. The modern approach to urban installation design should emphasize the protection of nature, prioritizing the creation of ecological, culturally enriched, and environmentally sustainable urban spaces. By fostering a more responsible interaction with the natural world, landscape design can minimize environmental degradation and contribute to the sustainable use of resources.

4.2. Principle of Human-Centered Design

The progression of modern urban settlements and lifestyles, driven by technological advances and shifting societal needs, reflects a broader understanding of human values and the social constructs that shape them. From the industrial revolution’s rapid technological expansion to today’s global emphasis on human-centered sustainable development, there has been a profound reevaluation of human identity and our role in the environment. This shift is indicative of societal advancement toward higher developmental ideals. In contemporary society, fostering harmonious urban environments that cater to human well-being, while simultaneously promoting the dissemination of high-quality information, has become a dominant trend within the knowledge economy era. In this context, installation art within urban landscapes serves as an essential element of public space design. Such spaces must account for diverse user needs, considering factors like accessibility for healthcare professionals, disabled individuals, and people of varying age, economic, and social backgrounds. Urban landscape installations, which serve as open and often multifunctional spaces, should thus be adaptable to the needs of various user groups, facilitating interaction, recreation, and aesthetic appreciation. Modern urban design must reflect the multiplicity of human nature—addressing both the basic survival needs, such as food, shelter, and safety, and the higher-order desires for beauty and social connection. Furthermore, it is crucial that urban landscapes provide spaces conducive to social interaction, recreation, and the fulfillment of psychological needs. The “people-oriented” design principle prioritizes understanding human psychological needs and creating spaces that are diverse, personalized, and reflective of regional identities. This approach ensures that urban landscapes cater to the various physical, emotional, and social needs of the populace, thereby fostering a more inclusive, dynamic public space.

Balancing ecological sustainability with human-centered design is crucial for creating urban spaces that are both environmentally harmonious and supportive of human well-being. Ecological sustainability emphasizes the importance of preserving the natural environment, ensuring that urban landscapes contribute to the long-term health of ecosystems. In contrast, human-centered design focuses on meeting the needs of individuals, prioritizing accessibility, inclusivity, and overall well-being. To achieve this balance, principles have been incorporated that prioritize both environmental conservation and human interaction. By creating green spaces that encourage community engagement while considering their ecological impact, urban landscapes can be both aesthetically pleasing and functional. This approach ensures that urban designs support not only the ecological health of the environment but also meet the diverse needs of the population, ultimately creating spaces that are both sustainable and responsive to human needs.

4.3. The Role of Color in Urban Landscape Design

Color is an immensely powerful element in design, influencing both aesthetic appreciation and emotional responses. It can amplify the visual impact of a design, enhance its inherent aesthetic value, and profoundly affect human perception. Color’s influence on visual perception is among its most notable characteristics, with different colors eliciting varying psychological and emotional reactions. The interplay between color and human experience is not only shaped by its immediate visual impact but also by deeper cultural and psychological associations. Colors are often categorized by their inherent qualities: luminosity, hue, and saturation, which together determine the emotional and visual effects they impart. For instance, warm colors such as red and orange are associated with energy, warmth, and stimulation, whereas cool colors like blue and violet evoke calmness, serenity, and coolness. These color attributes influence human behavior and perception, with warm colors typically stimulating action and excitement, and cool colors promoting relaxation and contemplation. Studies have shown that individuals tend to share common reactions to specific colors, underlining the importance of color choice in shaping the sensory and emotional atmosphere of urban landscapes. In landscape design, the strategic use of color not only enhances visual aesthetics but also contributes to creating environments that support psychological well-being, providing spaces that reflect the emotional and sensory needs of urban inhabitants.

4.4. Synthesis of Experience and Insights

Traditional backpropagation neural networks, characterized by their static layer structures, are limited in their ability to handle variable-sized sequential data due to the fixed output size generated through a fixed number of operations. In contrast, recurrent neural networks address this limitation by facilitating the transmission of information across time dimensions through recursive feedback loops within the network’s hidden layers. This dynamic structure allows RNNs to effectively capture the temporal dependencies inherent in sequential data, making them more suited for tasks that involve time-series or chronological data.

In both traditional and modern landscape design practices, the integration of natural materials and regional cultural histories is essential for creating cohesive and meaningful public spaces [44]. In contemporary urban landscape design, the focus extends beyond mere visual beauty to include the symbolism and functional integration of natural environments and architectural elements. By drawing upon natural materials and regional artistic traditions, designers can create spaces that resonate with both aesthetic and cultural significance. The use of installation art within landscape design serves not only as a tool for environmental enhancement but also as a medium for conveying social messages, fostering interaction, and deepening the connection between people and their environments. In the modern era, urban landscapes are evolving from static, homogeneous designs to dynamic, multifaceted spaces that embrace diversity, creativity, and sustainability. The integration of installation art into these spaces injects a new vitality into environmental design, imbuing the urban landscape with greater meaning, purpose, and engagement. As society progresses and the field of landscape design diversifies, urban spaces are increasingly seen as multifaceted cultural and environmental installations that invite public interaction, reflection, and aesthetic enjoyment. Thus, the role of installation art in urban landscape design extends beyond aesthetic enhancement, becoming a pivotal element in creating interactive, communicative, and environmentally conscious urban environments.

5. Conclusions

This study explores the intersection of ecological sustainability, human-centered design, and the integration of advanced computational methods in urban landscape planning, with a particular focus on the use of Recurrent Neural Networks and Long Short-Term Memory models for lithology identification in rural landscapes. The findings demonstrate that ecological principles, when applied to urban landscape design, not only preserve the natural environment but also promote the creation of spaces that foster community engagement, social interaction, and environmental harmony. Through the adoption of ecological design strategies, such as optimizing energy use, water conservation, and waste management, landscape design has evolved from a purely aesthetic practice to a more holistic, sustainable approach that addresses broader environmental and social concerns.

Furthermore, the study highlights the importance of human-centered design principles in creating urban spaces that are inclusive, accessible, and responsive to the diverse needs of the public. The use of installation art within urban landscapes offers new opportunities for public interaction, allowing these spaces to become dynamic environments that cater to a wide range of social, cultural, and psychological needs. By leveraging technologies such as RNNs and LSTMs, urban landscape design can be enhanced through data-driven insights, which allow for the more accurate and efficient identification of landscape features and the prediction of ecological patterns, contributing to better urban planning and development.

Ultimately, this research underscores the need for an integrated approach to urban landscape design that considers both ecological sustainability and human well-being. The successful fusion of these elements can lead to the development of urban spaces that are not only visually appealing but also ecologically responsible and socially inclusive, ensuring a more sustainable and harmonious coexistence between people and their environment.