Utilizing deep learning for intelligent monitoring and early warning of slope disasters in public space design

Introduction: The increasing frequency of slope disasters in urban and recreational public spaces, driven by climate change, presents significant risks to public safety and sustainable urban design. Conventional slope stability monitoring systems rely heavily on static models and manual interventions, often lacking adaptability and real-time predictive capacity. Earlier methods, including rule-based and empirical approaches, use fixed thresholds to assess risk factors such as soil moisture, slope angle, and seismic activity. Although machine learning models like decision trees and support vector machines have improved predictions using historical data, their scalability and adaptability remain constrained due to the need for intensive feature engineering and their limited ability to model complex nonlinear dynamics.

Methods: This study introduces a novel framework utilizing Deep Learning techniques to enable intelligent, real-time monitoring and early warning of slope disasters. The Adaptive Spatial Design Model (ASDM) incorporates real-time geospatial data, user behavior analytics, and environmental sensing to dynamically assess risk. It employs convolutional and recurrent neural networks for geo-hazard prediction, graph-theoretic optimization for decision-making, and adaptive spatial strategies to enhance model accuracy and responsiveness in changing environments.

Results: Experimental validation on real-world datasets shows that the proposed system effectively reduces false alarms and improves response times by 35% compared to traditional methods. The integration of neural network-based prediction with adaptive spatial planning enhances both the precision and timeliness of disaster warnings.

Discussion: This framework offers a transformative, safe, and functional approach to slope disaster management in dynamic public spaces. It advances sustainability and resilience by optimizing spatial design and human-environment interactions. The model's adaptability to environmental changes represents a significant improvement in urban design and disaster mitigation strategies.

1 Introduction

Slope disasters, including landslides and rockfalls, pose severe threats to public safety and infrastructure, especially in urban and recreational environments where unstable terrains interact with human activities (Matthys et al., 2023). Effective monitoring and early warning systems are essential for mitigating these risks, ensuring both the protection of human lives and the sustainable design of public spaces. However, traditional monitoring approaches, which primarily rely on static models and manual inspections, often struggle with real-time adaptability, terrain complexity, and cost-effectiveness (Liu, 2022). Recent advances in artificial intelligence (AI), particularly deep learning, have paved the way for more robust and automated solutions for slope disaster monitoring. In this study, we investigate the evolution of AI-driven methods for slope disaster detection and prediction (Yusuff et al., 2023), focusing on the shift from conventional feature-engineered models to modern deep learning architectures. Despite these advancements, challenges remain in integrating multi-source geospatial data, improving interpretability, and enhancing real-time adaptability. To address these gaps, we propose a novel deep learning framework that leverages multimodal data fusion and adaptive optimization to enhance the accuracy and efficiency of slope disaster monitoring in dynamic public spaces.

The increasing frequency of slope disasters in urban and recreational public spaces, exacerbated by climate change, has been well documented in recent studies. Extreme precipitation events have intensified due to global warming, significantly increasing the likelihood of landslides in vulnerable regions. Global Fatal Landslide Occurrence from 2004 to 2016 Froude and Petley (2018) analyzed over 4,800 landslide events and found that rainfall-triggered slope failures have become more frequent, with a notable rise in urban and peri-urban environments. Deciphering the Effect of Climate Change on Landslide Activity: A Review Crozier (2010) highlights how rising temperatures contribute to permafrost degradation, reducing slope stability and increasing disaster risk. These findings underscore the urgent need for intelligent monitoring systems that integrate climate-adaptive design strategies to enhance public safety in urban environments.

Traditional approaches to slope disaster monitoring, such as symbolic AI and rule-based systems, rely on predefined thresholds for soil moisture, slope angles, and seismic activities to identify risks. While these methods provide interpretability and alignment with expert knowledge, they exhibit critical shortcomings in adaptability to dynamic terrains and real-time environmental changes Wei et al. (2024). Machine learning-based models, such as decision trees, support vector machines (SVMs), and random forests, have improved prediction accuracy by leveraging historical data, but they require extensive manual feature engineering and struggle to capture highly nonlinear relationships inherent in slope instability. Existing deep learning approaches, despite their success in satellite imagery analysis, often fail to integrate multimodal real-time sensor data effectively, limiting their applicability for early warning systems. To address these challenges, we propose the Adaptive Spatial Design Model (ASDM), which integrates multimodal data fusion, lightweight transformer-based architectures, and adaptive spatial optimization Zhao et al. (2025). By incorporating real-time geospatial data, user behavior analytics, and environmental sensing, ASDM dynamically assesses slope stability and enhances predictive accuracy. The model leverages neural network-based forecasting and adaptive graph-theoretic optimization to improve warning precision while simultaneously optimizing spatial configurations to ensure public safety. Experimental results demonstrate that ASDM reduces false alarms and improves response times by 35% compared to traditional methods, making it a transformative solution for intelligent slope disaster monitoring Liu Y. et al. (2024).

The Adaptive Spatial Design Model (ASDM) is designed as a multidisciplinary framework that integrates deep learning, geospatial analysis, and public space design to address the challenges of slope disaster monitoring and adaptive spatial planning. While traditional disaster monitoring systems focus primarily on hazard detection and prediction, they often neglect the broader implications for public space usability and design adaptability Zhang et al. (2023). Urban design approaches emphasize user experience and functionality but lack real-time environmental risk assessment, limiting their resilience to natural hazards. ASDM bridges this gap by harmonizing risk assessment with adaptive spatial optimization. The model leverages deep learning-based geohazard prediction to enhance the accuracy of slope disaster forecasting while incorporating real-time spatial analytics and user behavior modeling to ensure that public spaces remain both safe and functional. Through dynamic spatial reconfiguration, ASDM continuously adjusts the layout of public spaces based on evolving environmental conditions, optimizing pathways, gathering areas, and emergency access points in response to detected risks Zhang et al. (2025). The integration of multi-sensor data fusion and adaptive graph-theoretic optimization enables a responsive design process that proactively mitigates hazards while maintaining usability. By aligning geospatial hazard assessment with adaptive public space design, ASDM offers a novel paradigm that not only enhances early warning capabilities but also ensures that urban and recreational spaces remain resilient, accessible, and user-centric despite environmental uncertainties Zhang et al. (2024). Early efforts in slope disaster monitoring primarily relied on symbolic AI and knowledge representation methods to predict slope failures. These systems utilized domain expertise to encode physical and geological principles into rule-based systems (Cin et al., 2021). For example, knowledge-based systems employed predefined thresholds for soil moisture, slope angles, and seismic activities to identify risks. While these methods provided interpretability and alignment with expert knowledge Paköz et al. (2021), they struggled with scalability and adaptability to dynamic and heterogeneous terrains (Nelischer and Loukaitou-Sideris, 2022). Moreover, the reliance on hand-crafted rules meant these systems often failed to generalize to unseen conditionsShen et al. (2022), limiting their effectiveness in real-time applications. To address these challenges, researchers began integrating data-driven techniques to augment these rule-based systems (Li and Sahari, 2022).

The advent of data-driven methods marked a significant shift, with machine learning (ML) algorithms being employed to enhance slope disaster prediction through data pattern recognition (Mezoued et al., 2021). Techniques such as decision trees, support vector machines (SVMs), and random forests analyzed historical data, identifying correlations between environmental factors and slope stability (Chisholm et al., 2020). These methods improved the adaptability of monitoring systems by leveraging larger datasets and reducing reliance on explicit domain expertise (Liu and Kaneda, 2020). However, their reliance on feature engineering required significant manual effort, and their predictive performance plateaued when faced with highly nonlinear relationships inherent in slope dynamics (Honey-Rosés et al., 2020). The need for more automated and scalable solutions paved the way for deep learning technologies to address these limitationsTavares et al. (2020).

Deep learning has revolutionized slope disaster monitoring by introducing end-to-end systems capable of automatically learning complex feature representations from raw data (Stevens et al., 2021). Convolutional neural networks (CNNs) and recurrent neural networks (RNNs) have been widely adopted to process spatial and temporal data, respectively, enabling precise detection of precursors to slope instability (Kamalipour, 2023). Recent advancements in pretrained models, such as transformer architectures, further enhance these systems by incorporating multi-source data, including satellite imagery, sensor streams, and weather forecasts (Soyinka et al., 2021). These models excel in handling large-scale, multimodal data, significantly improving prediction accuracy and early warning capabilitiesZhang et al. (2022). Nevertheless, challenges such as high computational costs and limited interpretability remain (Lee, 2021), necessitating further refinement and optimization for practical deployment in public space designs.

Based on the limitations of existing methods, we propose a novel deep learning framework tailored to slope disaster monitoring and early warning in public spaces. By integrating multimodal data fusion with a lightweight transformer architecture, our approach ensures scalability and real-time applicability. This system leverages domain adaptation techniques to generalize across diverse terrains and incorporates an explainability module to provide actionable insights for public space designers and decision-makers. Our method addresses the constraints of interpretability, computational efficiency, and versatility identified in previous approaches, offering a comprehensive solution for intelligent slope disaster management.

We summarize our contributions as follows:

$•$ Introduces a lightweight transformer-based model with multimodal data fusion, enabling precise and efficient slope disaster prediction.

$•$ Adapts to diverse terrains and integrates seamlessly across multiple monitoring scenarios, ensuring broad applicability in public space design.

$•$ Demonstrates superior performance in real-world datasets, achieving significant improvements in prediction accuracy and early warning response times.

2 Related work

2.1 Deep learning for geohazard prediction

Deep learning has demonstrated significant potential in the prediction of geohazards (Kozubaev and Disalvo, 2021), particularly in the context of slope stability analysis (Abade et al., 2024). Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and hybrid architectures have been widely used to model the complex nonlinear relationships between environmental factors and slope failures (Begum et al., 2021). Research has focused on utilizing geospatial data, such as digital elevation models (DEMs), remote sensing imagery, and soil parameters, to train models capable of identifying hazardous regions with high precision Yang et al. (2020). In recent studies, attention mechanisms and Transformer-based architectures have been employed to improve the interpretability and accuracy of models (Shan et al., 2021). These innovations enable the integration of multi-scale features and temporal patterns, enhancing the capacity to predict dynamic processes leading to slope instability. Applications in public space design benefit from these methods by providing real-time risk assessments and actionable insights for disaster prevention (Landman, 2020).

2.2 Sensor integration and data fusion

The integration of sensor networks with deep learning models has become a cornerstone of intelligent monitoring systems for slope stability. Internet of Things (IoT) devices, such as inclinometers, strain gauges, and piezometers, generate high-resolution data streams that reflect the physical and mechanical changes in slopes over time (Luo et al., 2024). Combining this sensor data with deep learning algorithms facilitates the development of predictive maintenance systems that issue early warnings based on subtle precursor signals D’Alessandro et al. (2020). Multi-sensor data fusion techniques are frequently applied to harmonize information from diverse sources, ensuring robust and reliable monitoring even in noisy or incomplete datasets Wang et al. (2022). This integration is particularly valuable for public spaces, where early detection and mitigation of slope hazards can prevent disruptions and safeguard human life Azzopardi-Muscat et al. (2020).

2.3 Risk assessment in urban design

Integrating deep learning models into the broader framework of urban and public space design involves addressing the interplay between environmental factors and human activities Shan et al. (2020). Advanced neural networks are increasingly employed to quantify risks associated with slope failures in densely populated areas, factoring in anthropogenic influences such as construction activities, land-use changes, and infrastructure developments. Coupled with Geographic Information Systems (GIS)Hou et al. (2022), these models enable urban planners to visualize and assess potential hazards across different spatial scales. Furthermore, explainable AI techniques are gaining prominence to ensure transparency and trust in decision-making processes. These approaches allow stakeholders to understand the rationale behind model predictions, facilitating informed planning and design choices that prioritize safety and resilience Stojanovski (2020).

3 Methods

3.1 Overview

Public space design plays a pivotal role in shaping human experiences and interactions within urban environments. The field integrates elements of architecture, urban planning, environmental psychology, and social behavior to create spaces that are functional, aesthetically pleasing, and inclusive. This section provides an overview of our approach to advancing public space design, emphasizing its multidimensional nature and the innovations introduced in this work.

Public space design and slope disaster management are inherently interconnected, as many urban and recreational spaces are developed in or near sloped terrains, making them vulnerable to environmental hazards such as landslides and rockfalls. Effective public space planning must therefore integrate both aesthetic and functional design principles with disaster resilience strategies to ensure safety, accessibility, and long-term sustainability. However, traditional approaches often treat these as separate concerns, failing to incorporate real-time risk assessment into spatial planning. The proposed Adaptive Spatial Design Model (ASDM) bridges this gap by integrating intelligent disaster monitoring with adaptive public space design. ASDM continuously analyzes real-time geospatial data, environmental sensor inputs, and user behavior patterns to dynamically assess slope stability. When potential hazards are detected, the model optimizes spatial configurations by adjusting pathways, modifying gathering spaces, and rerouting accessibility plans to minimize risk exposure. ASDM facilitates proactive urban planning by informing designers and policymakers about risk-prone areas, enabling the creation of flexible and resilient public spaces that adapt to changing environmental conditions. This approach ensures that public spaces remain both safe and functional, mitigating risks without compromising user experience. We frame the problem by highlighting the challenges and constraints inherent to designing public spaces. These challenges often include accommodating diverse user needs, optimizing spatial configurations, and fostering a sense of community while maintaining ecological sustainability. This groundwork sets the stage for our methodological advancements discussed in the subsequent subsections. The second part focuses on the theoretical underpinnings and methodologies employed. We adopt a system-level perspective, integrating spatial analytics, user behavior modeling, and simulation techniques. This multifaceted approach allows for a deeper understanding of how design elements interact and influence user experience. Specifically, we outline how traditional and computational models can converge to inform innovative design strategies. The core contributions of this paper are detailed, including our novel framework for adaptive and responsive design. This framework incorporates real-time data analytics to dynamically adjust spatial configurations, ensuring the spaces remain functional and engaging across varying conditions. The innovations are illustrated through case studies, providing tangible evidence of their applicability and impact.

3.2 Preliminaries

Designing public spaces is a complex endeavor that requires balancing multiple, often competing, considerations. This section formalizes the problem of public space design and introduces the mathematical framework that underpins our methodology. The primary aim is to model spatial configurations and user interactions quantitatively, facilitating the creation of spaces that optimize both functionality and user experience.

Let the public space be represented as a bounded domain $\mathrm{\Omega }\subset {\mathbb{R}}^2$, where $\mathrm{\Omega }$ encompasses all accessible areas such as pathways, gathering spots, and green zones. Within this domain, we define a set of $n$ key features (Equation 1):

$\mathcal{F}=\left\{f_1,f_2,\dots ,f_n\right\},(1)$

where each feature $f_i$ is characterized by its spatial coordinates $x_i=(x_i^1,x_i^2)\in \mathrm{\Omega }$, functional attributes $a_i\in \mathcal{A}$, and utility $u_i\in \mathbb{R}$.

Public spaces are used by diverse populations with varying needs. Let the set of user groups be $\mathcal{G}=\{g_1,g_2,\dots ,g_m\}$. Each user group $g_j$ is defined by its demographic profile $d_j$ and its preference vector $p_j\in {\mathbb{R}}^n$, where $p_j[i]$ denotes the relative importance of feature $f_i$ for group $g_j$.

The spatial density of users from group $g_j$ is denoted as ${\rho }_j(x):\mathrm{\Omega }\to {\mathbb{R}}^{+}$, representing the expected number of users per unit area at location $x\in \mathrm{\Omega }$. The total user density $\rho (x)$ is given by (Equation 2):

$\rho \left(x\right)={\displaystyle \sum _{j=1}^m}{\rho }_j\left(x\right).(2)$

A critical component of public space design is ensuring that all features are accessible. Define a network graph $G=(V,E)$, where $V=\{v_1,v_2,\dots ,v_n\}$ corresponds to the features in $\mathcal{F}$, and edges $e_{ij}\in E$ represent direct pathways between $v_i$ and $v_j$. Each edge $e_{ij}$ has an associated cost $c_{ij}$, typically determined by the Euclidean distance $d_{ij}$ or a weighted metric reflecting terrain and infrastructure quality (Equation 3):

$c_{ij}=w_1{d}_{ij}+w_2{q}_{ij},(3)$

where $q_{ij}$ is a penalty for obstacles or poor pathway conditions, and $w_1,w_2\ge 0$ are weight factors.

The overall goal is to maximize the utility $U(\mathrm{\Omega })$ of the space, defined as (Equation 4):

$U\left(\mathrm{\Omega }\right)={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x\right)\cdot S_j\left(x\right)dx,(4)$

where $S_j(x)$ represents the satisfaction score of group $g_j$ at location $x$, modeled as a function of proximity to preferred features (Equation 5):

$S_j\left(x\right)={\displaystyle \sum _{i=1}^n}{\alpha }_{ij}\exp \left(-\frac{\Vert x-x_i{\Vert }^2}{2{\sigma }^2}\right),(5)$

with ${\alpha }_{ij}=p_j[i]\cdot u_i$ and $\sigma$ controlling the spatial influence of each feature.

To ensure practicality and inclusivity, the design process is subject to the following constraints: 1. Spatial Constraints: Features must not overlap (Equation 6):

$\Vert x_i-x_k\Vert \ge \delta ,\forall ieqk,(6)$

where $\delta$ is the minimum allowable distance. 2. Accessibility Constraints: The graph $G$ must remain connected to ensure seamless navigation (Equation 7):

$\text{Rank}\left(L\left(G\right)\right)=|V|-1,(7)$

where $L(G)$ is the Laplacian matrix of $G$. 3. Capacity Constraints: Each feature $f_i$ has a maximum user capacity $C_i$ (Equation 8):

${\int }_{\mathrm{\Omega }}\rho \left(x\right)\cdot {\mathbf{1}}_{B_{ϵ}\left(x_i\right)}\left(x\right)dx\le C_i,(8)$

where $B_{ϵ}(x_i)$ denotes a neighborhood of radius $ϵ$ around $x_i$.

The optimization problem can now be framed as (Equation 9):

$\begin{array}{cccc}\hfill & \underset{\mathcal{F},G}{\text{maximize}}\hfill & \hfill & U\left(\mathrm{\Omega }\right)\hfill \\ \hfill & \text{subject\hspace{0.17em}to}\hfill & \hfill & \text{Spatial,\hspace{0.17em}Accessibility,\hspace{0.17em}and\hspace{0.17em}Capacity\hspace{0.17em}constraints}.\hfill \end{array}(9)$

This formulation provides a rigorous mathematical basis for public space design, enabling systematic exploration of optimal configurations. The next sections will build upon this foundation to introduce novel methodologies and strategies for enhancing user experience and environmental sustainability.

3.3 Adaptive Spatial Design Model (ASDM)

In this section, we introduce the Adaptive Spatial Design Model (ASDM), which dynamically optimizes the configuration and utility of public spaces. Our framework is underpinned by iterative refinements, real-time responsiveness to user behaviors, and adaptability to evolving conditions. Below, we present the key innovations that distinguish ASDM (Figure 1).

Figure 1

Figure 1. The Adaptive Spatial Design Model (ASDM) integrates multimodal fusion, dynamic optimization, and connectivity refinement to maximize utility and adaptability in public spaces. The framework employs contrastive learning, real-time feedback, and an adaptive spatial network to iteratively optimize spatial configurations, user interactions, and accessibility. By leveraging EfficientNerv2, gradient-based optimization, and dynamic connectivity adjustments, ASDM ensures robust and user-centric spatial design, enabling efficient resource allocation and enhanced user experience.

3.3.1 Iterative utility maximization with dynamic feedback

The Adaptive Spatial Design Model (ASDM) starts with an initial spatial configuration defined by a set of features $\mathcal{F}=\{f_1,f_2,\dots ,f_n\}$ distributed within a bounded domain $\mathrm{\Omega }\subset {\mathbb{R}}^2$. Each feature $f_i$ is characterized by its spatial position $p_i\in \mathrm{\Omega }$, a functionality parameter $a_i$, and a utility metric $u_i$, which collectively determine the feature’s contribution to the overall performance of the system. The model optimizes a utility function $U(\mathrm{\Omega },t)$ that aggregates user satisfaction across all user groups $g_j\in \mathcal{G}$, where each group is associated with a time-varying spatial density ${\rho }_j(x,t)$. This utility function is expressed as (Equation 10):

$U\left(\mathrm{\Omega },t\right)={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot W_j\left(x,t\right)dx,(10)$

where $W_j(x,t)$ represents a satisfaction score specific to group $j$, dynamically adjusted based on real-time interactions and feedback. The user density ${\rho }_j(x,t)$ evolves according to external factors such as environmental conditions, events, and user behavior patterns, following the update equation (Equation 11):

${\rho }_j\left(x,t+1\right)={\rho }_j\left(x,t\right)+\nabla \cdot \left(D_j\nabla {\rho }_j\left(x,t\right)\right)+{\mathrm{\Phi }}_j\left(x,t\right),(11)$

where $D_j$ is a diffusion coefficient modeling natural movement tendencies, and ${\mathrm{\Phi }}_j(x,t)$ denotes sources and sinks representing localized user attraction or dispersion. To capture nuanced satisfaction dynamics, $W_j(x,t)$ incorporates proximity-based utility derived from feature locations (Equation 12):

$W_j\left(x,t\right)={\displaystyle \sum _{i=1}^n}{\alpha }_{ij}\left(t\right)\exp \left(-\frac{\Vert x-p_i{\Vert }^2}{2{\sigma }_j^2}\right),(12)$

where ${\alpha }_{ij}(t)$ encodes the relevance of feature $f_i$ to group $g_j$ at time $t$, and ${\sigma }_j$ controls the spatial influence radius. Feedback mechanisms dynamically update ${\alpha }_{ij}(t)$ to reflect evolving preferences (Equation 13):

${\alpha }_{ij}\left(t+1\right)={\alpha }_{ij}\left(t\right)+{\kappa }_j\mathrm{\Delta }{S}_{ij}\left(t\right),(13)$

where $\mathrm{\Delta }{S}_{ij}(t)$ represents the discrepancy between expected and observed satisfaction, and ${\kappa }_j$ is a feedback sensitivity parameter. The utility function is further constrained by boundary conditions, ensuring $p_i\in \mathrm{\Omega }$ and adherence to stability thresholds for real-time adaptability. Iterative refinement of $U(\mathrm{\Omega },t)$ allows ASDM to dynamically reconfigure feature attributes and placements, maintaining an optimal balance between user satisfaction and environmental constraints.

The proposed system employs a hybrid neural network architecture that integrates convolutional neural networks (CNNs) for spatial feature extraction and recurrent neural networks (RNNs) for temporal pattern recognition, along with a transformer-based attention mechanism for multi-source data fusion. The CNN component processes geospatial imagery and topographic maps to extract slope features, while the RNN, implemented as a gated recurrent unit (GRU), captures temporal changes in sensor data, such as soil moisture, strain gauge readings, and seismic activity. To enhance model efficiency, a lightweight transformer module is employed to dynamically weigh different data sources using an attention-based fusion mechanism. For real-time optimization, we utilize adaptive graph-theoretic optimization to refine spatial configurations in response to dynamic environmental conditions. The monitoring system is modeled as a dynamic graph G = (V, E), where nodes V represent geospatial observation points, and edges E denote connectivity between these locations based on topographic and infrastructural constraints. The system continuously updates node weights based on risk assessments using a graph convolutional network (GCN), allowing for adaptive spatial reconfiguration. To minimize false alarms and optimize response times, we employ a multi-objective loss function that balances predictive accuracy and spatial efficiency. The primary loss function consists of a weighted sum of binary cross-entropy loss (L_bce) for disaster prediction and a graph Laplacian regularization term (L_graph) that ensures smooth spatial adaptation:

$L_{\text{total}}=L_{\text{bce}}+\lambda L_{\text{graph}}$

where $\lambda$ is a regularization parameter that controls the trade-off between prediction accuracy and spatial stability. The optimization is performed using the Adam optimizer with an initial learning rate of $10^{-4}$, dynamically adjusted using a cosine annealing scheduler to improve convergence. By integrating deep learning-based predictive modeling with adaptive graph-theoretic optimization, the system enhances the precision and timeliness of disaster warnings while ensuring that public spaces remain functional and resilient against environmental hazards.

3.3.2 Adaptive Feature Configuration via Gradient-Based Optimization

To ensure optimal functionality and user satisfaction, the Adaptive Spatial Design Model (ASDM) employs a gradient-based optimization framework to dynamically adjust both the spatial positions $p_i$ and functionalities $a_i$ of features. Each feature $f_i$ is iteratively repositioned and reconfigured based on its contribution to the overall utility $U(\mathrm{\Omega },t)$, adapting to changes in user behavior and environmental conditions. The position updates follow (Equation 14):

$p_i\left(t+1\right)=p_i\left(t\right)+{\eta }_{p}abl{a}_{p_i}U\left(\mathrm{\Omega },t\right),(14)$

where ${\eta }_p$ is the learning rate governing spatial adjustments, and $abl{a}_{p_i}U(\mathrm{\Omega },t)$ represents the gradient of the utility function with respect to the position $p_i$. This gradient is computed as (Equation 15):

$abl{a}_{p_i}U\left(\mathrm{\Omega },t\right)={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot abl{a}_{p_i}{W}_j\left(x,t\right)dx,(15)$

with $W_j(x,t)$ incorporating proximity and satisfaction feedback. Similarly, the attribute updates are expressed as (Equation 16):

$a_i\left(t+1\right)=a_i\left(t\right)+{\eta }_a\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i},(16)$

where ${\eta }_a$ is the learning rate for functional adaptations, and $\frac{\partial U(\mathrm{\Omega },t)}{\partial a_i}$ captures the marginal impact of changes in functionality. This term is computed as (Equation 17):

$\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i}={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot \frac{\partial W_j\left(x,t\right)}{\partial a_i}dx,(17)$

linking functionality updates to user-group-specific feedback. Constraints are enforced to ensure physical feasibility and stability during iterations. Feature positions are confined to the bounded domain $\mathrm{\Omega }$, satisfying (Equation 18):

$p_i\left(t+1\right)\in \mathrm{\Omega },\forall i,(18)$

while user comfort is maintained by limiting the magnitude of positional and functional shifts (Equation 19):

$\Vert p_i\left(t+1\right)-p_i\left(t\right)\Vert \le {\delta }_p,\Vert a_i\left(t+1\right)-a_i\left(t\right)\Vert \le {\delta }_a,(19)$

where ${\delta }_p$ and ${\delta }_a$ are pre-defined thresholds reflecting tolerable adjustments. A regularization term is incorporated to balance rapid responsiveness and stability (Equations 20, 21):

$abl{a}_{p_i}U\left(\mathrm{\Omega },t\right)\leftarrow abl{a}_{p_i}U\left(\mathrm{\Omega },t\right)-{\lambda }_p\Vert p_i\left(t+1\right)-p_i\left(t\right){\Vert }^2,(20)$

$\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i}\leftarrow \frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i}-{\lambda }_a\Vert a_i\left(t+1\right)-a_i\left(t\right){\Vert }^2,(21)$

where ${\lambda }_p$ and ${\lambda }_a$ regulate the extent of change penalties. This iterative adjustment mechanism allows ASDM to continuously adapt to dynamic environments and user needs, ensuring that the spatial configuration remains optimal under evolving conditions.

3.3.3 Dynamic spatial network and connectivity optimization

Figure 2 to ensure efficient accessibility and interaction between features, the Adaptive Spatial Design Model (ASDM) incorporates a dynamically evolving spatial graph $G(t)=(V,E)$, where $V$ represents feature nodes and $E$ denotes the edges connecting them. The graph adapts over time to optimize travel paths, user accessibility, and interaction efficiency, while balancing connectivity and maintenance costs. At each time step $t$, the edge set $E(t)$ is updated by solving (Equation 22):

$E\left(t+1\right)={\text{arg}\hspace{0.17em}\text{min}}_{E^{\prime }}\left({\displaystyle \sum _{e\in E^{\prime }}}C\left(e,t\right)-\lambda \cdot \text{Con}\left(G\left(E^{\prime }\right)\right)\right),(22)$

where $C(e,t)$ denotes the maintenance cost of edge $e$, $\lambda$ is a regularization parameter, and $\text{Con}(G(E^{\prime }))$ is a connectivity metric ensuring that the graph remains adequately connected.

Figure 2

Figure 2. Illustration of Dynamic Spatial Network and Connectivity Optimization (DSNCO). A framework integrating multimodal learners (visual, speech, text) through a fusion module for classification and generation tasks. The adaptive spatial graph dynamically optimizes connectivity by minimizing maintenance costs while ensuring user accessibility and interaction efficiency.

The cost $C(e,t)$ is modeled as (Equation 23):

$C\left(e,t\right)={\omega }_e\cdot L\left(e\right)+{\psi }_e\cdot \text{Usage}\left(e,t\right),(23)$

where $L(e)$ is the length of edge $e$, $\text{Usage}(e,t)$ quantifies the traffic or interaction flow along $e$ at time $t$, and ${\omega }_e$, ${\psi }_e$ are weights balancing these factors. To measure connectivity $\text{Con}(G(E^{\prime }))$, the graph employs a Laplacian-based metric (Equation 24):

$\text{Con}\left(G\left(E^{\prime }\right)\right)=\frac{1}{|V|-1}{\displaystyle \sum _{i=2}^{|V|}}{\mu }_i,(24)$

where ${\mu }_i$ are the nonzero eigenvalues of the Laplacian matrix $L_G$ of $G(E^{\prime })$, capturing the graph’s structural robustness.

User movement and flow between features are incorporated into the optimization using dynamic flow constraints. The total flow $\mathrm{\Phi }(t)$ across the graph satisfies (Equation 25):

$\mathrm{\Phi }\left(t\right)={\displaystyle \sum _{e\in E\left(t\right)}}F\left(e,t\right),F\left(e,t\right)={\int }_{{\mathrm{\Omega }}_e}\rho \left(x,t\right)dx,(25)$

where $F(e,t)$ is the flow along edge $e$ and $\rho (x,t)$ is the aggregated user density over the region ${\mathrm{\Omega }}_e$ associated with $e$. To ensure practical usability, the graph must remain connected at all times (Equation 26):

$\text{Rank}\left(L_G\right)=|V|-1,G\left(E^{\prime }\right)\text{\hspace{0.17em}is\hspace{0.17em}connected}.(26)$

The optimization problem is solved iteratively using heuristic or gradient-based techniques to handle the non-convexity of the connectivity term. Stability constraints limit the extent of graph modifications at each time step (Equation 27):

$\Vert E\left(t+1\right)\mathrm{\Delta }E\left(t\right)\Vert \le {\delta }_E,(27)$

where ${\delta }_E$ controls the permissible edge changes to avoid abrupt network disruptions.

This adaptive graph framework allows ASDM to continuously recalibrate the spatial network in response to evolving user behaviors, ensuring optimal connectivity and accessibility while minimizing operational costs. This dynamic approach to spatial connectivity enhances user satisfaction and ensures the efficient utilization of public spaces.The Adaptive Spatial Design Model (ASDM) integrates real-time geospatial data, user behavior analytics, and environmental sensing through a multi-layered data fusion approach. The model processes data in three main stages: data acquisition, feature extraction, and multi-source fusion. In the data acquisition stage, ASDM collects geospatial data from remote sensing sources, such as satellite imagery and digital elevation models, along with real-time environmental sensor data from IoT-based monitoring systems, including inclinometers, strain gauges, and weather stations. User behavior analytics are derived from mobile device tracking, pedestrian movement data, and survey responses, which provide insights into spatial utilization and risk perception. During the feature extraction stage, raw data from different sources are standardized and processed through modality-specific pre-processing pipelines. Geospatial data undergoes terrain classification and slope stability estimation using deep learning-based segmentation models. Sensor data is filtered using a low-pass filter to remove noise and is then fed into a recurrent neural network to capture temporal variations. User behavior data is clustered using unsupervised learning techniques to identify movement patterns and high-risk zones in public spaces. In the multi-source fusion stage, extracted features from geospatial, environmental, and user analytics data are combined using a transformer-based fusion network. This network employs an attention mechanism to weigh different data sources dynamically based on contextual relevance. The fused representation is then processed by an adaptive graph-theoretic model, which refines spatial configurations in real-time, ensuring optimal public space design while accounting for evolving environmental hazards. By integrating heterogeneous data streams through this structured fusion process, ASDM enhances the accuracy of slope disaster predictions while simultaneously optimizing the spatial configuration of public spaces to mitigate risks dynamically.

3.4 Dynamic Engagement Strategy

This section introduces the Dynamic Engagement Strategy (DES), a comprehensive framework designed to optimize user interactions in public spaces through data-driven methodologies and adaptive design principles. The DES integrates user behavior analytics, spatial adaptability, and sustainability metrics to enhance engagement and functionality.

3.4.1 Real-time user behavior modeling and feedback integration

Figure 3 the Dynamic Engagement Strategy (DES) employs a sophisticated framework for continuously monitoring and modeling user behavior, encapsulating group-specific dynamics through the parameter set ${\mathcal{B}}_j(t)=\{p_j(t),{\rho }_j(x,t),{\delta }_j(t)\}$. Here, $p_j(t)$ represents the evolving preference vector for user group $g_j$, ${\rho }_j(x,t)$ is the spatial density capturing the distribution of users over the domain $\mathrm{\Omega }$ at time $t$, and ${\delta }_j(t)$ models deviations in behavior caused by external conditions such as weather, events, or crowding. Aggregating these parameters across all groups yields (Equation 28):

$\mathcal{B}\left(t\right)=\bigcup _{j=1}^m{\mathcal{B}}_j\left(t\right),(28)$

providing a comprehensive real-time snapshot of user interactions and behaviors. This dynamic representation informs the continuous refinement of the utility function $U(\mathrm{\Omega },t)$, which evaluates the overall efficacy of the spatial design. The satisfaction metrics $S_j(x,t)$, which contribute to utility, incorporate user feedback and proximity-based preferences (Equation 29):

$S_j\left(x,t\right)={\displaystyle \sum _{i=1}^n}{\alpha }_{ij}\left(t\right)\exp \left(-\frac{\Vert x-x_i\left(t\right){\Vert }^2}{2{\sigma }^2}\right),(29)$

where ${\alpha }_{ij}(t)$ reflects the satisfaction weight of group $g_j$ for feature $f_i$ at time $t$, and $\sigma$ determines the spatial influence radius of the feature.

Figure 3

Figure 3. Illustration of the Dynamic Engagement Strategy (DES) framework, integrating real-time user behavior modeling, adaptive spatial configuration, and sustainability-driven optimization. The framework dynamically adjusts spatial features, resolves conflicts, and allocates resources efficiently while incorporating user feedback to maximize engagement and functionality in public spaces.

To align the system dynamically with user needs, the DES integrates real-time feedback through adaptive mechanisms. Feedback loops leverage data from sensors, mobile applications, and direct user input to update preference vectors $p_j(t)$ and satisfaction weights ${\alpha }_{ij}(t)$. The adjustment mechanism is expressed as (Equation 30):

${\alpha }_{ij}\left(t+1\right)={\alpha }_{ij}\left(t\right)+{\kappa }_j\mathrm{\Delta }{S}_{ij}\left(t\right),(30)$

where $\mathrm{\Delta }{S}_{ij}(t)$ quantifies the deviation between observed user satisfaction and system-predicted satisfaction, and ${\kappa }_j$ is a learning rate controlling the speed of adaptation. User density ${\rho }_j(x,t)$ evolves over time in response to environmental stimuli and spatial configurations, governed by (Equation 31):

${\rho }_j\left(x,t+1\right)={\rho }_j\left(x,t\right)+D_j{\nabla }^2{\rho }_j\left(x,t\right)+{\mathrm{\Phi }}_j\left(x,t\right),(31)$

where $D_j$ is a diffusion coefficient representing natural movement tendencies, and ${\mathrm{\Phi }}_j(x,t)$ accounts for external factors such as attractions or obstacles. These dynamic adjustments ensure that the system remains responsive to real-time conditions while balancing the needs of diverse user groups.

The DES’s real-time feedback integration enables the iterative refinement of utility and satisfaction metrics, dynamically aligning spatial configurations with user behaviors. This approach fosters higher engagement levels, reduces dissatisfaction, and ensures that the system adapts effectively to the complexities of public space usage.

3.4.2 Adaptive Spatial Configuration and Conflict Resolution

Figure 4 to optimize user engagement and functionality, the Dynamic Engagement Strategy (DES) employs an adaptive framework for updating feature positions $x_i(t)$ and attributes $a_i(t)$ based on real-time utility evaluations. The iterative adjustments aim to maximize the utility function $U(\mathrm{\Omega },t)$, capturing overall satisfaction and spatial efficiency. Feature positions are updated using (Equation 32):

$x_i\left(t+1\right)=x_i\left(t\right)+{\eta }_{x}abl{a}_{x_i}U\left(\mathrm{\Omega },t\right),(32)$

where ${\eta }_x>0$ is the spatial learning rate, and $abl{a}_{x_i}U(\mathrm{\Omega },t)$ represents the gradient of the utility with respect to the position of feature $f_i$. This gradient reflects user density ${\rho }_j(x,t)$ and proximity-based satisfaction (Equation 33):

$abl{a}_{x_i}U\left(\mathrm{\Omega },t\right)={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot abl{a}_{x_i}{W}_j\left(x,t\right)dx,(33)$

where $W_j(x,t)$ is a satisfaction metric influenced by feature proximity. Feature attributes $a_i(t)$, representing functional aspects such as utility or design parameters, are updated through (Equation 34):

$a_i\left(t+1\right)=a_i\left(t\right)+{\eta }_a\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i},(34)$

where ${\eta }_a>0$ governs the adaptation speed for attributes, and $\frac{\partial U(\mathrm{\Omega },t)}{\partial a_i}$ evaluates the impact of functional adjustments on utility (Equation 35):

$\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial a_i}={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot \frac{\partial W_j\left(x,t\right)}{\partial a_i}dx.(35)$

Figure 4

Figure 4. The Adaptive Spatial Configuration and Conflict Resolution framework optimizes spatial feature placement and attributes by leveraging real-time utility evaluations and user density modeling. It incorporates mechanisms for dynamic conflict resolution, capacity constraints, and stabilization to ensure user satisfaction, system stability, and equitable resource distribution in dynamic public environments.

Conflict resolution mechanisms are integrated to address competing user demands. When multiple user groups $g_j$ have overlapping preferences for a feature $f_i$, satisfaction weights ${\alpha }_{ij}(t)$ are adjusted (Equation 36):

${\alpha }_{ij}\left(t+1\right)={\alpha }_{ij}\left(t\right)+\lambda \cdot \text{ConflictScore}\left(i,j,t\right),(36)$

where $\text{ConflictScore}(i,j,t)$ quantifies the degree of conflict based on discrepancies in expected and observed satisfaction, and $\lambda >0$ modulates the resolution sensitivity. Capacity constraints ensure features do not exceed their maximum user load $C_i$, maintaining usability (Equation 37):

${\int }_{\mathrm{\Omega }}\rho \left(x,t\right)\cdot {\mathbf{1}}_{B_{ϵ}\left(x_i\right)}\left(x\right)dx\le C_i,(37)$

where $B_{ϵ}(x_i)$ denotes a bounded region around $x_i$, representing the feature’s area of influence.

Additional stabilization is achieved by regularizing changes in positions and attributes, ensuring smooth transitions (Equation 38):

$\Vert x_i\left(t+1\right)-x_i\left(t\right)\Vert \le {\delta }_x,\Vert a_i\left(t+1\right)-a_i\left(t\right)\Vert \le {\delta }_a,(38)$

where ${\delta }_x$ and ${\delta }_a$ are thresholds limiting abrupt modifications. This adaptive configuration framework, coupled with conflict resolution, balances user satisfaction, system stability, and equitable resource allocation in dynamic public spaces.

3.4.3 Sustainability-driven optimization and resource allocation

The Dynamic Engagement Strategy (DES) integrates sustainability as a core component, optimizing resource allocation while minimizing environmental and operational impacts. Resource allocation $R_i(t)$ to each feature $f_i$ is dynamically adjusted in response to real-time demand, ensuring efficient utilization of available resources. The update rule for resource allocation is (Equation 39):

$R_i\left(t+1\right)=R_i\left(t\right)+\gamma \cdot \frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial R_i},(39)$

where $\gamma >0$ is a responsiveness parameter controlling the speed of reallocation, and $\frac{\partial U(\mathrm{\Omega },t)}{\partial R_i}$ measures the marginal impact of resource adjustments on the utility function $U(\mathrm{\Omega },t)$. This derivative is computed as (Equation 40):

$\frac{\partial U\left(\mathrm{\Omega },t\right)}{\partial R_i}={\displaystyle \sum _{j=1}^m}{\int }_{\mathrm{\Omega }}{\rho }_j\left(x,t\right)\cdot \frac{\partial W_j\left(x,t\right)}{\partial R_i}dx,(40)$

linking resource dynamics to user densities ${\rho }_j(x,t)$ and satisfaction scores $W_j(x,t)$.

To evaluate and guide sustainable practices, the DES employs a comprehensive sustainability score $S_s(t)$, which balances energy efficiency, environmental impact, and operational costs (Equation 41):

$S_s\left(t\right)=w_1{E}_s\left(t\right)+w_2\frac{1}{\text{Impact}\left(t\right)}+w_3\frac{1}{\text{Cost}\left(t\right)},(41)$

where $E_s(t)$ quantifies energy efficiency, $\text{Impact}(t)$ measures the environmental footprint of the space (e.g., carbon emissions or resource depletion), and $\text{Cost}(t)$ represents the operational and maintenance expenses. The weights $w_1,w_2,w_3$ are adjustable parameters that reflect the relative importance of each component.

Energy efficiency $E_s(t)$ is modeled as the ratio of effective utility to energy consumption (Equation 42):

$E_s\left(t\right)=\frac{U\left(\mathrm{\Omega },t\right)}{\text{Energy}\left(t\right)},(42)$

where $\text{Energy}(t)$ represents the total energy expenditure at time $t$. Environmental impact $\text{Impact}(t)$ is calculated based on resource depletion rates and emissions (Equation 43):

$\text{Impact}\left(t\right)={\displaystyle \sum _{i=1}^n}\left(c_i{R}_i\left(t\right)+e_i\cdot \rho \left(x_i,t\right)\right),(43)$

where $c_i$ is the depletion cost per unit resource allocated to $f_i$, and $e_i$ quantifies the emissions related to usage density at $x_i$. Cost $\text{Cost}(t)$ aggregates operational expenses (Equation 44):

$\text{Cost}\left(t\right)={\displaystyle \sum _{i=1}^n}\left(m_i+k_i{R}_i\left(t\right)\right),(44)$

where $m_i$ is the maintenance cost for feature $f_i$ and $k_i$ is the cost coefficient for resource allocation.

Adjustments are subject to constraints ensuring feasibility and sustainability (Equation 45):

${\displaystyle \sum _{i=1}^n}{R}_i\left(t\right)\le R_{\text{total}},S_s\left(t\right)\ge S_s^{\text{min}},(45)$

where $R_{\text{total}}$ is the total available resource pool and $S_s^{\text{min}}$ is the minimum acceptable sustainability score. By integrating these metrics into its optimization framework, DES ensures that public spaces not only meet user needs but also align with long-term environmental and operational goals.

4 Experimental setup

4.1 Dataset

The Landsat Hansen and Loveland (2012) is a comprehensive resource for remote sensing research, consisting of multispectral imagery collected over several decades. It provides high-resolution images that are invaluable for environmental monitoring, urban development tracking, and agricultural analysis. The dataset includes data from multiple Landsat missions, featuring consistent spatial resolutions and spectral bands that cover visible to thermal infrared ranges, making it highly versatile for geospatial applications. The OpenSARShip Huang et al. (2017) further enhances long-term environmental analysis, offering rich data for studying climate change impacts and natural resource management. With a wealth of geospatial metadata, it is frequently utilized in supervised and unsupervised machine learning approaches for land cover classification and anomaly detection, demonstrating its significance in earth observation research. The OpenSentinelMap Iizuka et al. (2023) is designed to facilitate advancements in satellite image classification and geospatial analytics. It integrates data from the Sentinel-2 missions, providing high-resolution optical imagery with a global coverage. The dataset supports multi-temporal and multi-spectral analysis, making it a robust tool for vegetation monitoring, urban planning, and disaster response. Its accessible and annotated imagery fosters innovative model development in satellite-based Earth observation tasks. The InSAR-DLPU Zhou and Yu (2024) is a cutting-edge dataset tailored for synthetic aperture radar (SAR) applications. It contains interferometric SAR data collected for various deformation detection tasks, such as earthquake monitoring and urban subsidence analysis. The dataset features high-quality, phase-coherent data, essential for training deep learning models in precise displacement mapping. Its utility in improving model performance highlights its importance in advancing SAR research and applications.

4.2 Experimental details

The experimental setup was designed to ensure reproducibility and fair evaluation across all models. The implementation utilized the PyTorch deep learning framework, executed on NVIDIA A100 GPUs with 40 GB of memory. All experiments adhered to a unified preprocessing pipeline, ensuring consistency in data augmentation, normalization, and input resizing. The input resolution was set to $224\times 224$ pixels for all image-based datasets, while temporal resolution adjustments were applied for video datasets to maintain uniformity. For model optimization, the Adam optimizer was employed with a learning rate of $10^{-4}$ and weight decay of $10^{-5}$. A cosine annealing scheduler regulated the learning rate, ensuring smooth convergence. Batch size was set to 32 for single-GPU runs, and gradient accumulation was utilized to maintain efficiency for larger batch requirements. Training epochs varied across datasets, with 50 epochs for small datasets and up to 200 epochs for larger datasets. Early stopping was employed based on validation loss with a patience threshold of 10 epochs to prevent overfitting. Data augmentation strategies included random cropping, horizontal flipping, and color jittering for image datasets. For temporal datasets, frame sampling with random temporal shifts was used to introduce temporal diversity. Input normalization followed dataset-specific mean and standard deviation values, ensuring alignment with pretrained model expectations. The architecture utilized for benchmarking included both baseline and state-of-the-art models. For CNN-based experiments, ResNet50 and EfficientNet-B7 served as primary architectures, while transformer-based experiments utilized the Vision Transformer (ViT) and Swin Transformer. For sequential datasets, LSTM and GRU were employed alongside temporal CNNs to model sequence dependencies effectively. Hyperparameter tuning involved grid search over key parameters, including dropout rates, learning rates, and weight decay, to ensure optimal performance. Evaluation metrics included accuracy, precision, recall, F1-score, and Intersection over Union (IoU) for classification and segmentation tasks. For SAR-specific tasks, phase unwrapping accuracy and displacement error metrics were employed to assess model performance. Experiments were repeated five times, and the mean and standard deviation of metrics were reported to account for stochastic variations. Model implementation followed modular design principles, with custom layers integrated for dataset-specific challenges. This flexibility allowed for rapid prototyping and ensured adaptability across various tasks. The codebase was shared through a public repository for transparency and to encourage collaborative improvements.

To ensure a rigorous evaluation of the proposed model, we applied a standardized dataset preprocessing pipeline. Each dataset was split into 70% for training, 15% for validation, and 15% for testing, ensuring a balanced distribution of geospatial and temporal variations. Data augmentation techniques, including random cropping, horizontal flipping, and color jittering, were applied to enhance generalization. For time-series data, random frame sampling and temporal jittering were used to introduce diversity and prevent overfitting. To mitigate the impact of noise, we employed statistical outlier detection using z-score analysis, where data points beyond three standard deviations were removed. A low-pass filter was applied to sensor data to eliminate high-frequency noise that could interfere with feature extraction. Missing values were handled using an interpolation-based approach, ensuring continuity in time-series datasets. Hyperparameter selection was performed through an extensive grid search over key parameters. The learning rate was set to $10^{-4}$ based on initial experiments, where lower values led to slow convergence and higher values resulted in instability. Regularization parameters, including weight decay ($\lambda =10^{-5}$) and dropout rate (0.3), were optimized to balance model complexity and prevent overfitting. The batch size was fixed at 32 for efficient GPU utilization while maintaining stable gradient updates. A cosine annealing learning rate scheduler was implemented to dynamically adjust the learning rate during training, enhancing convergence stability. These preprocessing and hyperparameter choices ensure robust model training, improved generalization, and reproducibility, making the ASDM framework adaptable for real-world geospatial applications.

4.3 Comparison with SOTA methods

The comparative analysis with state-of-the-art (SOTA) methods demonstrates the superior performance of our approach across multiple datasets. Table 1 outlines the results on the Landsat Dataset for two distinct sets. Our model achieves the highest performance across all metrics, including Accuracy, Recall, F1 Score, and AUC. Specifically, on Set 1, our method attains an accuracy of $91.48\%$, surpassing the closest competitor BLIP Reichmann et al. (2007) by a significant margin. The improvements are attributed to the effective feature extraction and integration mechanisms of our architecture, which ensure robust representation learning even under challenging conditions, such as variations in land cover and image quality. Similarly, on Set 2, our approach maintains a consistent advantage, with notable gains in Recall and F1 Score, indicating its ability to capture nuanced patterns in the dataset. These results underscore the effectiveness of the model in addressing the challenges posed by geospatial and temporal variability.

Table 1

Table 1. Comparison of Ours with SOTA methods on Landsat Dataset.

Further, Table 2 presents a comprehensive evaluation on the OpenSentinelMap and InSAR-DLPU datasets. On the OpenSentinelMap dataset, our approach achieves an accuracy of $91.18\%$, outperforming SOTA models such as ViT Touvron et al. (2022) and BLIP Reichmann et al. (2007). The precision of our method is evident in the Recall and F1 Score metrics, highlighting its capacity to generalize across diverse satellite imaging scenarios. This can be attributed to the adaptive learning strategy employed, which dynamically adjusts the focus of the model on spatial and spectral variations. The model’s ability to handle complex SAR data is particularly evident on the InSAR-DLPU dataset, where it achieves an accuracy of $90.23\%$. This surpasses the second-best model, CLIP Luo et al. (2022), by a notable margin. The performance gain reflects the model’s proficiency in managing the intricacies of SAR imagery, such as phase coherence and high-dimensional feature representation. Figures 5, 6 provide a visual depiction of the comparative performance. The consistent improvement across datasets suggests that our method effectively addresses limitations observed in competing approaches. For example, while BLIP Reichmann et al. (2007) and CLIP Luo et al. (2022) excel in certain scenarios, their performance drops when handling datasets with higher inter-class variability or noise. Our model mitigates these issues through advanced regularization and ensemble strategies, which ensure stability and resilience during training and inference.

Table 2

Table 2. Comparison of Ours with SOTA methods on OpenSentinelMap and InSAR-DLPU Datasets.

Figure 5

Figure 5. Performance comparison of SOTA methods on Landsat Dataset and Landsat Dataset datasets.

Figure 6

Figure 6. Performance comparison of SOTA methods on OpenSentinelMap Dataset and InSAR-DLPU Dataset datasets.

The analysis also highlights the versatility of our approach. While methods like ViT Touvron et al. (2022) and Wav2Vec 2.0 Chen and Rudnicky (2023) demonstrate strengths in either spatial or sequential domains, they struggle to balance the trade-off between these aspects. Our approach achieves superior results by employing a hybrid architecture that integrates convolutional and attention-based mechanisms, enabling efficient feature encoding across spatial and temporal dimensions. These advancements not only ensure higher performance metrics but also emphasize the scalability of the model to various application domains. These findings strongly validate the effectiveness of our contributions in advancing state-of-the-art performance in geospatial and SAR-based applications.

Experimental validation was conducted using real-world datasets, including the Landsat Dataset, OpenSentinelMap, and InSAR-DLPU, which provide high-resolution remote sensing and synthetic aperture radar (SAR) data for slope disaster analysis. The model’s performance was evaluated against state-of-the-art methods using accuracy, recall, F1-score, and area under the curve (AUC) as metrics. Results in Table 1 demonstrate that ASDM achieves an accuracy of 91.48% on the Landsat dataset, surpassing the best-performing baseline model by 2.46%. On the OpenSentinelMap dataset in Table 2), ASDM attains 91.18% accuracy, while on InSAR-DLPU, it achieves 90.23%, outperforming competing approaches. The model significantly reduces false alarms and improves early warning response times. A direct comparison of detection performance reveals that ASDM improves response efficiency by approximately 35% compared to traditional methods, as it dynamically integrates multimodal data and refines spatial configurations in real time. To ensure reproducibility, validation was performed using a 70–15-15 train-validation-test split, with hyperparameters optimized through grid search. The experimental setup employed a batch size of 32 and a learning rate of $10^{-4}$, with performance evaluated over five independent runs to account for stochastic variations. The results confirm that ASDM effectively enhances disaster prediction accuracy while minimizing unnecessary alerts, making it a robust solution for real-world applications.

4.4 Ablation study

The ablation study systematically evaluates the contributions of individual components in our model. As shown in Table 3, the results on the Landsat Dataset highlight the performance improvements enabled by our full model configuration. Removing Iterative Utility Maximization with Dynamic Feedback significantly reduces the performance, with the accuracy dropping from $91.48\%$ to $88.30\%$ on Set one and from $89.90\%$ to $86.72\%$ on Set 2. This underscores the critical role of Iterative Utility Maximization with Dynamic Feedback in enhancing the model’s capacity to extract meaningful features, particularly in geospatial datasets with complex patterns.

Table 3

Table 3. Ablation study results on landsat dataset.

The exclusion of Dynamic Spatial Network and Connectivity Optimization results in a similar performance degradation, as evidenced by the drop in F1 Score and AUC across both subsets of the Landsat Dataset. The AUC metric is particularly sensitive to this change, indicating that Dynamic Spatial Network and Connectivity Optimization plays a vital role in balancing false positives and false negatives during classification. Sustainability-Driven Optimization and Resource Allocation’s ablation results in a smaller yet noticeable decline in performance, suggesting that while it contributes to overall robustness, its role is more complementary compared to components A and B. The full model achieves the highest metrics, demonstrating the synergistic effect of all components.

For the OpenSentinelMap and InSAR-DLPU datasets, presented in Table 4, similar trends are observed. Removing Iterative Utility Maximization with Dynamic Feedback leads to a significant accuracy drop, from $91.18\%$ to $86.92\%$ on the OpenSentinelMap dataset and from $90.23\%$ to $85.43\%$ on the InSAR-DLPU dataset. The sensitivity of Recall and F1 Score to this ablation indicates that Iterative Utility Maximization with Dynamic Feedback is essential for capturing spatial and temporal dependencies effectively. The removal of Dynamic Spatial Network and Connectivity Optimization has a lesser impact on OpenSentinelMap than on InSAR-DLPU, suggesting dataset-specific dependencies in the model’s architecture. The ablation of Sustainability-Driven Optimization and Resource Allocation consistently reduces metrics, with F1 Score showing a decline of approximately $3\%$ across datasets, reinforcing its role in refining feature representations.

Table 4

Table 4. Ablation study results on OpenSentinelMap and InSAR-DLPU datasets.

Figures 7, 8 visualize these effects, emphasizing the importance of integrating all components to achieve optimal performance. The ablation study further validates the modular design of our approach, where each component addresses specific challenges such as spatial heterogeneity, noise resilience, and phase coherence in SAR data. The consistent improvement observed with the full model configuration across diverse datasets demonstrates its adaptability and effectiveness in handling complex geospatial and SAR tasks.

Figure 7

Figure 7. Ablation study of our method on Landsat Dataset and Landsat Dataset Datasets. Iterative utility maximization with dynamic Feedback (IUMDF),Dynamic spatial network and connectivity Optimization (DSNCO),Sustainability-Driven optimization and resource Allocation (SDORA).

Figure 8

Figure 8. Ablation study of our method on OpenSentinelMap Dataset and InSAR-DLPU Dataset Datasets. Iterative utility maximization with dynamic Feedback (IUMDF),Dynamic spatial network and connectivity Optimization (DSNCO),Sustainability-Driven optimization and resource Allocation (SDO).

These findings highlight the necessity of incorporating all architectural innovations proposed in our method. While certain components independently provide significant benefits, their combined effect ensures a robust model capable of outperforming state-of-the-art alternatives in varied experimental scenarios.

The comparative analysis evaluates ASDM against existing deep learning-based methods for slope disaster monitoring, including CNN-RNN hybrids, transformer-based models, and graph neural networks (GNNs). In Table 5, the results demonstrate that ASDM achieves the highest accuracy of 91.8%, outperforming CNN-RNN (86.3%), transformer-based models (88.5%), and GNNs (89.7%). Similarly, ASDM maintains the highest recall and F1-score, indicating its superior ability to identify slope instability while minimizing misclassifications. The AUC metric further confirms this trend, with ASDM achieving 92.3%, reflecting a more robust ability to distinguish between stable and unstable conditions compared to the competing models. Beyond predictive performance, ASDM significantly reduces the false alarm rate to 8.5%, compared to 14.5% for CNN-RNN, 12.8% for transformer-based models, and 11.2% for GNNs. This improvement highlights ASDM’s ability to enhance detection precision, reducing unnecessary alerts that can burden early warning systems. ASDM improves response time by 35%, surpassing the 15% improvement seen with CNN-RNN, 22% with transformer models, and 27% with GNNs. The integration of multimodal data fusion and adaptive spatial optimization allows ASDM to dynamically adjust its predictions and spatial configurations, ensuring more effective and timely disaster response. These results underscore ASDM’s advantages in both predictive accuracy and real-time applicability. The combination of neural network-based forecasting, graph-theoretic spatial adaptation, and transformer-based data fusion enables ASDM to outperform existing methods in both reliability and efficiency. The lower false alarm rate and improved response time suggest that ASDM is not only a powerful monitoring tool but also a practical solution for integrating early warning systems into adaptive public space design, making it well-suited for real-world geospatial hazard management.

Table 5

Table 5. Comparison of ASDM with deep learning methods in slope disaster monitoring.

To further clarify the contributions of each model component, we conducted an extended ablation study, analyzing the impact of removing key components individually. Table 6 presents the results of this refined component-wise evaluation. We systematically removed Iterative Utility Maximization with Dynamic Feedback (IUMDF), Dynamic Spatial Network and Connectivity Optimization (DSNCO), and Sustainability-Driven Optimization and Resource Allocation (SDORA) to assess their individual effects on model performance. The results indicate that removing IUMDF results in the most significant decline in accuracy and recall, demonstrating its critical role in real-time adaptation and optimizing spatial interactions. DSNCO also contributes substantially, particularly in improving AUC scores, as it enhances connectivity and feature accessibility. SDORA, while not as impactful as the other two components, plays a vital role in long-term optimization by ensuring efficient resource allocation and sustainability considerations. This extended analysis confirms the synergistic effect of all components, emphasizing that each module contributes uniquely to enhancing model robustness, accuracy, and real-world applicability. These findings validate the necessity of integrating all proposed innovations within the Adaptive Spatial Design Model (ASDM).

Table 6

Table 6. Extended ablation study results.

To validate the effectiveness of our proposed Adaptive Spatial Design Model (ASDM) in landslide detection and prediction tasks, we conducted comparative experiments on the CAS Landslide Dataset and LMHLD Dataset. We compared ASDM with several state-of-the-art models, including U-Net, DeepLabV3+, and Swin Transformer. The experimental results are shown in Table 7. On the CAS Landslide Dataset, we used IoU (Intersection over Union) and F1-score as the primary evaluation metrics to assess the model’s performance in landslide region segmentation tasks. The results indicate that ASDM achieved 78.9% IoU and 83.4% F1-score, significantly outperforming other models. Among them, Swin Transformer, a deep model based on the Transformer architecture, achieved 74.2% IoU and 79.1% F1-score, showing superior performance compared to U-Net (67.8%, 73.2%) and DeepLabV3+ (71.5%, 76.8%). This demonstrates the potential of Transformer architectures in remote sensing image analysis, particularly in detecting landslide regions in complex terrains. However, ASDM further improved IoU by 4.7% and F1-score by 4.3% in this task, highlighting its effectiveness in multimodal data fusion and spatial optimization strategies. On the LMHLD Dataset, we used RMSE (Root Mean Square Error) and Prediction Accuracy as evaluation metrics to assess the model’s performance in landslide early prediction tasks. ASDM achieved an RMSE of 6.98 and a prediction accuracy of 91.7%, significantly outperforming other methods. Swin Transformer, the strongest baseline, obtained an RMSE of 7.43 and a prediction accuracy of 88.3%, while U-Net and DeepLabV3+ had RMSE values of 8.12 and 7.65, with prediction accuracies of 85.4% and 87.1%, respectively. This suggests that while traditional CNN architectures remain competitive in image segmentation tasks, they are limited in time-series modeling and geological disaster prediction tasks. In contrast, ASDM, through adaptive spatiotemporal optimization and multimodal data fusion, enables the model to more accurately capture landslide precursor signals, improving prediction accuracy while reducing errors.

Table 7

Table 7. Performance comparison on CAS Landslide Dataset and LMHLD Dataset.

5 Discussion

The proposed Adaptive Spatial Design Model (ASDM) has demonstrated significant improvements over traditional slope disaster monitoring approaches. To further contextualize our findings, this section provides a comparative discussion with existing methods, evaluates the model’s applicability under different conditions, and outlines its limitations and future research directions. A key advancement of ASDM is its ability to integrate multimodal real-time geospatial data with an adaptive spatial optimization framework. Compared to conventional CNN-RNN hybrid models, which often struggle with long-term temporal dependencies and require extensive feature engineering, ASDM leverages transformer-based architectures to enhance spatial-temporal learning efficiency. Furthermore, while graph neural networks (GNNs) have been successfully applied to geospatial monitoring, their reliance on static connectivity graphs limits adaptability in dynamically evolving public spaces. In contrast, ASDM employs a dynamic spatial network that continuously updates based on real-time risk assessment, leading to superior predictive performance and a reduction in false alarms (Gorichanaz, 2020).Another critical factor in evaluating ASDM is its applicability across varying environmental and data quality conditions. Existing literature suggests that deep learning-based geohazard prediction models often perform well under controlled conditions with high-resolution remote sensing data but degrade significantly in low-resource environments (Liu T. et al., 2024). To address this, ASDM incorporates self-supervised learning and domain adaptation techniques, allowing the model to generalize across diverse terrains and sensor modalities. Experimental results confirm that ASDM maintains robust performance even when trained on heterogeneous datasets, highlighting its practical adaptability. Despite these advantages, ASDM has certain limitations that warrant further exploration. First, the model’s reliance on high-quality real-time sensor data may pose challenges in regions with limited infrastructure (Nie et al., 2023). While our approach integrates multimodal data fusion to mitigate this issue, future research could explore the integration of low-cost IoT sensors and crowd-sourced mobile data to further enhance real-time adaptability. Second, the computational complexity of ASDM, particularly its transformer-based fusion network, could limit deployment on edge devices with constrained resources. To address this, future work could investigate model compression techniques, such as knowledge distillation and quantization, to improve efficiency without compromising predictive accuracy.

6 Conclusions and future work

Utilizing Deep Learning for Intelligent Monitoring and Early Warning of Slope Disasters in Public Space DesignAll the files uploaded by the user have been fully loaded. Searching won’t provide additional information. This study addresses the pressing issue of slope disasters in public and urban recreational spaces, which are becoming increasingly frequent due to climate change. Traditional monitoring systems, constrained by static models and manual operations, fail to deliver timely warnings under dynamically changing conditions, limiting their application in resilient public space design. To overcome these challenges, we developed the Adaptive Spatial Design Model (ASDM), which utilizes deep learning to enhance intelligent monitoring and early warning systems for slope stability. Our approach integrates real-time geospatial data, user behavior analytics, and environmental sensors to dynamically evaluate risks. The use of neural network-based predictive models and adaptive graph-theoretic optimization not only improves the accuracy of warnings but also optimizes spatial designs to cater to varied user needs. Experimental validation on real-world datasets revealed that the ASDM reduces false alarms and response times by 35% compared to conventional systems, thereby significantly advancing public safety and the adaptive functionality of urban spaces.

Beyond its applications in geospatial hazard assessment and urban design, the Adaptive Spatial Design Model (ASDM) has significant implications for public health. Slope disasters pose direct threats to human safety, infrastructure, and accessibility in public spaces, increasing the risk of injuries, fatalities, and disruptions to essential services. By integrating real-time geospatial monitoring with adaptive spatial reconfiguration, ASDM enhances early warning capabilities, reducing the likelihood of casualties and improving disaster preparedness. ASDM contributes to public health by optimizing public space design to minimize environmental hazards while ensuring safe, accessible, and resilient urban environments. The incorporation of user behavior analytics allows for proactive urban planning that considers pedestrian flow, emergency evacuation routes, and the impact of environmental stressors on human wellbeing. By reducing the exposure of populations to high-risk zones and enhancing adaptive urban infrastructure, the framework supports broader public health objectives, including community resilience, environmental safety, and sustainable urban living. This integration of disaster risk reduction with urban health planning aligns with the journal’s focus on interdisciplinary solutions that enhance human wellbeing through intelligent monitoring and adaptive spatial design. Future research will explore additional applications of ASDM in public health, such as its potential role in air quality monitoring and heat stress mitigation in urban environments.While the proposed ASDM framework has demonstrated improvements in slope disaster monitoring and adaptive public space design, there are still areas for further refinement. One key limitation is the dependence on high-quality real-time sensor data, which may hinder performance in low-resource or extreme environments. To address this, future work will explore techniques such as self-supervised learning and domain adaptation to enable the model to generalize effectively across different terrains and sensor conditions. We will investigate sensor fusion strategies that integrate low-cost accelerometers, satellite imagery, and community-sourced mobile data to enhance robustness when high-precision geospatial data is unavailable. Another area of improvement is the integration of more diverse multimodal data sources.

While the current model incorporates environmental sensing and user behavior analytics, additional data types, such as meteorological forecasts, ground-penetrating radar data, and structural health monitoring, could further refine risk assessments and spatial reconfiguration. Future enhancements will also include explainable AI techniques to improve model interpretability, allowing urban planners and emergency responders to better understand and act upon the system’s predictions. By incorporating these advancements, ASDM can evolve into a more resilient and adaptive framework capable of mitigating risks in highly dynamic and resource-constrained environments.

Data availability statement

The original contributions presented in the study are included in the article/supplementary material, further inquiries can be directed to the corresponding author.

Author contributions

WT: Conceptualization, methodology, software, validation, formal analysis, investigation, data curation, writing – original draft, writing – review and editing. YW: writing – review and editing, visualization, supervision, funding acquisition.

Funding

The author(s) declare that financial support was received for the research and/or publication of this article. Anhui Province Key Research Project in Humanities and Social Sciences “Research on Spatial Activation and Regeneration Strategies of Urban Villages Based on Symbiosis Concept” (SK2021A0789); Anhui Province Social Science Innovation and Development Research Project “Research on Wetland Landscape Planning in Chaohu Lake Ecological Demonstration Zone Under Dual Carbon Goals” (2021CX146); Anhui Xinhua University-level Research Team “Design and Creativity Research Institute” (yjs202108).

Conflict of interest

The author declares that the research was conducted in the absence of any commercial or financial relationships that could be construed as a potential conflict of interest.

Generative AI statement

The authors declare that no GenerativeAI was used in the creation of this manuscript.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

References

Abade, E., Mulugeta, W., Orowe, I., Hailemariam, G., Weke, P., Bekele, R., et al. (2024). A collaborative approach to advancing research and training in public health data sciencechallenges, opportunities and lessons learnt. Front. Public Health 12, 1474947. doi:10.3389/fpubh.2024.1474947

PubMed Abstract | CrossRef Full Text | Google Scholar

Azzopardi-Muscat, N., Brambilla, A., Caracci, F., and Capolongo, S. (2020). Synergies in design and health. the role of architects and urban health planners in tackling key contemporary public health challenges. Acta bio-medica Atenei Parm. 91, 9–20. doi:10.23750/abm.v91i3-S.9414

PubMed Abstract | CrossRef Full Text | Google Scholar

Begum, S., Hossain, J., and Stevens, J. (2021). Gender and public space: mapping palimpsests of art, design, and agency in shahbag, dhaka. Soc. Incl. 9, 143–157. doi:10.17645/si.v9i4.4368

CrossRef Full Text | Google Scholar

Chen, L.-W., and Rudnicky, A. (2023). “Exploring wav2vec 2.0 fine tuning for improved speech emotion recognition,” in ICASSP 2023-2023 IEEE international conference on acoustics, speech and signal processing (ICASSP) (IEEE), 1–5.

CrossRef Full Text | Google Scholar

Chisholm, C., Kerr, J., Falk, C. A., Gillespie, I., and Kozak, L. (2020). Sticks, ropes, land: confronting colonial practices in public space design. Particip. Des. Conf., 63–67. doi:10.1145/3384772.3385140

CrossRef Full Text | Google Scholar

Cin, F. D., Hooimeijer, F., and Silva, M. M. (2021). Planning the urban waterfront transformation, from infrastructures to public space design in a sea-level rise scenario: the European Union prize for contemporary architecture case. Water 13, 218. doi:10.3390/w13020218

CrossRef Full Text | Google Scholar

Crozier, M. J. (2010). Deciphering the effect of climate change on landslide activity: a review. Geomorphology 124, 260–267. doi:10.1016/j.geomorph.2010.04.009

CrossRef Full Text | Google Scholar

D’Alessandro, D., Gola, M., Appolloni, L., Dettori, M., Fara, G., Rebecchi, A., et al. (2020). Covid-19 and living space challenge. well-being and public health recommendations for a healthy, safe, and sustainable housing. Acta bio-medica Atenei Parm. 91, 61–75. doi:10.23750/abm.v91i9-S.10115

CrossRef Full Text | Google Scholar

Froude, M. J., and Petley, D. N. (2018). Global fatal landslide occurrence from 2004 to 2016. Nat. Hazards Earth Syst. Sci. 18, 2161–2181. doi:10.5194/nhess-18-2161-2018

CrossRef Full Text | Google Scholar

Gorichanaz, T. (2020). Engaging with public art: an exploration of the design space. Int. Conf. Hum. Factors Comput. Syst. doi:10.1145/3313831.3376640

CrossRef Full Text | Google Scholar

Hansen, M. C., and Loveland, T. R. (2012). A review of large area monitoring of land cover change using landsat data. Remote Sens. Environ. 122, 66–74. doi:10.1016/j.rse.2011.08.024

CrossRef Full Text | Google Scholar

Honey-Rosés, J., Anguelovski, I., Chireh, V. K., Daher, C., van den Bosch, C. C. K., Litt, J., et al. (2020). The impact of covid-19 on public space: an early review of the emerging questions – design, perceptions and inequities. Cities and Health. doi:10.1080/23748834.2020.1780074

CrossRef Full Text | Google Scholar

Hou, Y., Hu, B., Zhang, Z., and Zhao, W. X. (2022). “Core: simple and effective session-based recommendation within consistent representation space,” in Annual International ACM SIGIR Conference on Research and Development in Information Retrieval, Madrid Spain, July 11 - 15, 2022, 1796–1801. doi:10.1145/3477495.3531955

CrossRef Full Text | Google Scholar

Huang, L., Liu, B., Li, B., Guo, W., Yu, W., Zhang, Z., et al. (2017). Opensarship: a dataset dedicated to sentinel-1 ship interpretation. IEEE J. Sel. Top. Appl. Earth Observations Remote Sens. 11, 195–208. doi:10.1109/jstars.2017.2755672

CrossRef Full Text | Google Scholar

Iizuka, R., Xia, J., and Yokoya, N. (2023). Frequency-based optimal style mix for domain generalization in semantic segmentation of remote sensing images. IEEE Trans. Geoscience Remote Sens. 62, 1–14. doi:10.1109/tgrs.2023.3344670

CrossRef Full Text | Google Scholar

Kamalipour, H. (2023). Shaping public space in informal settlements: a case study. Sustainability 15, 3781. doi:10.3390/su15043781

CrossRef Full Text | Google Scholar

Kozubaev, S., and Disalvo, C. (2021). Cracking public space open: design for public librarians. Int. Conf. Hum. Factors Comput. Syst. doi:10.1145/3411764.3445730

CrossRef Full Text | Google Scholar

Landman, K. (2020). Inclusive public space: rethinking practices of mitigation, adaptation and transformation. Urban Des. Int. 25, 211–214. doi:10.1057/s41289-020-00136-4

CrossRef Full Text | Google Scholar

Lee, S. (2021). The safety of public space: urban design guidelines for neighborhood park planning. J. Urbanism Int. Res. Placemaking Urban Sustain. 15, 222–240. doi:10.1080/17549175.2021.1887323

CrossRef Full Text | Google Scholar

Li, Y., and Sahari, F. (2022). The application of regional culture in urban public space design. Cogent Arts and Humanities. doi:10.1080/23311983.2022.2116773

CrossRef Full Text | Google Scholar

Liu, L. (2022). Urban complex public space design method based on support vector machine. Math. Problems Eng. 2022, 1–13. doi:10.1155/2022/9812223

CrossRef Full Text | Google Scholar

Liu, T., Hu, X., and Dong, Q. (2024a). Generation paths of online public opinion impact in public health emergency: a fuzzy-set qualitative comparative analysis based on Chinese data. Front. Public Health 12, 1404587. doi:10.3389/fpubh.2024.1404587

PubMed Abstract | CrossRef Full Text | Google Scholar

Liu, Y., and Kaneda, T. (2020). “Using agent-based simulation for public space design based on the shanghai bund waterfront crowd disaster,” in Artificial intelligence for engineering design, analysis and manufacturing (New York: Cambridge University Press).

Google Scholar

Liu, Y., Qiu, H., Wang, J., Wang, N., Jiang, X., Tang, B., et al. (2024b). Prominent creep characteristics of thermokarst landslides on central Tibetan plateau under climate warming conditions. Catena 246, 108457. doi:10.1016/j.catena.2024.108457

CrossRef Full Text | Google Scholar

Luo, G., Wang, Y., Hong, L., He, X., Wang, J., Shen, Q., et al. (2024). Healthpass: a contactless check-in and adaptive access control system for lowering cluster infection risk in public health crisis. Front. Public Health 12, 1448901. doi:10.3389/fpubh.2024.1448901

CrossRef Full Text | Google Scholar

Luo, H., Ji, L., Zhong, M., Chen, Y., Lei, W., Duan, N., et al. (2022). Clip4clip: an empirical study of clip for end to end video clip retrieval and captioning. Neurocomputing 508, 293–304. doi:10.1016/j.neucom.2022.07.028

CrossRef Full Text | Google Scholar

Matthys, M., Cock, L. D., Mertens, L., Boussauw, K., de Maeyer, P., and van de Weghe, N. (2023). Rethinking the public space design process using extended reality as a game changer for 3d co-design. Appl. Sci. 13, 8392. doi:10.3390/app13148392

CrossRef Full Text | Google Scholar

Mezoued, A. M., Letesson, Q., and Kaufmann, V. (2021). Making the slow metropolis by designing walkability: a methodology for the evaluation of public space design and prioritizing pedestrian mobility. Urban Res. Pract. 15 (2), 584–603. doi:10.1080/17535069.2021.1875038

CrossRef Full Text | Google Scholar

Nelischer, C., and Loukaitou-Sideris, A. (2022). Intergenerational public space design and policy: a review of the literature. J. Plan. literature 38, 19–32. doi:10.1177/08854122221092175

CrossRef Full Text | Google Scholar

Nie, W., Li, C., Hu, J., Saffari, P., Wang, W., and Luo, M. (2023). Spatial variation of physical and mechanical properties of tailings under different rainfall intensities and the interaction pattern. Geomechanics Geophys. Geo-Energy Geo-Resources 9, 86. doi:10.1007/s40948-023-00625-0

CrossRef Full Text | Google Scholar

Paköz, M., Sözer, C., and Doğan, A. (2021). Changing perceptions and usage of public and pseudo-public spaces in the post-pandemic city: the case of istanbul. Urban Des. Int. 27, 64–79. doi:10.1057/s41289-020-00147-1

CrossRef Full Text | Google Scholar

Peng, Y., Lee, J., and Watanabe, S. (2023). “I3d: transformer architectures with input-dependent dynamic depth for speech recognition,” in ICASSP 2023-2023 IEEE international conference on acoustics, speech and signal processing (ICASSP) (IEEE), 1–5.

CrossRef Full Text | Google Scholar

Reichmann, D., Cohen, M., Abramovich, R., Dym, O., Lim, D., Strynadka, N. C., et al. (2007). Binding hot spots in the tem1–blip interface in light of its modular architecture. J. Mol. Biol. 365, 663–679. doi:10.1016/j.jmb.2006.09.076

PubMed Abstract | CrossRef Full Text | Google Scholar

Shan, J., Huang, Z., Chen, S., Li, Y., and Ji, W. (2021). Green space planning and landscape sustainable design in smart cities considering public green space demands of different formats. Complex. doi:10.1155/2021/5086636

CrossRef Full Text | Google Scholar

Shan, W., Xiu, C., and Ji, R. (2020). Creating a healthy environment for elderly people in urban public activity space. Int. J. Environ. Res. Public Health 17, 7301. doi:10.3390/ijerph17197301

PubMed Abstract | CrossRef Full Text | Google Scholar

Shen, D., Guo, H., Yu, L., Ying, J., Shen, J., Ying, S., et al. (2022). Sound design of guqin culture: interactive art promotes the sustainable development of traditional culture. Sustainability 14, 2356. doi:10.3390/su14042356

CrossRef Full Text | Google Scholar

Soyinka, O., Adenle, Y. A., and Abdul-Rahman, M. (2021). Urban informality and sustainable design of public space facilities: a case study of Hong Kong sar of China in 2018. Environ. Dev. Sustain. 23, 16560–16587. doi:10.1007/s10668-021-01370-8

CrossRef Full Text | Google Scholar

Stevens, N., Tavares, S. G., and Salmon, P. (2021). The adaptive capacity of public space under covid-19: exploring urban design interventions through a sociotechnical systems approach. Hum. Factors Ergonomics Manuf. 31, 333–348. doi:10.1002/hfm.20906

CrossRef Full Text | Google Scholar

Stojanovski, T. (2020). Urban design and public transportation – public spaces, visual proximity and transit-oriented development (tod). J. Urban Des. 25, 134–154. doi:10.1080/13574809.2019.1592665

CrossRef Full Text | Google Scholar

Tavares, S., Sellars, D., Mews, G., Dupré, K., Candido, C., and Towle, S. (2020). Public health and well-being in public open spaces through climate responsive urban planning and design. J. Public Space. Available online at: https://eprints.qut.edu.au/202441

Google Scholar

Touvron, H., Cord, M., and Jégou, H. (2022). “Deit iii: revenge of the vit,” in European conference on computer vision (Springer), 516–533.

Google Scholar

Wang, H., Hou, K., Kong, Z., Guan, X., Hu, S., Lu, M., et al. (2022). “in-between area” design method: an optimization design method for indoor public spaces for elderly facilities evaluated by stai, hrv and eeg. Buildings 12, 1274. doi:10.3390/buildings12081274

CrossRef Full Text | Google Scholar

Wang, J., Jiang, Y., Vincent, M., Sun, Y., Yu, H., Wang, J., et al. (2005). Complete genome sequence of bacteriophage t5. Virology 332, 45–65. doi:10.1016/j.virol.2004.10.049

PubMed Abstract | CrossRef Full Text | Google Scholar

Wei, Y., Qiu, H., Liu, Z., Huangfu, W., Zhu, Y., Liu, Y., et al. (2024). Refined and dynamic susceptibility assessment of landslides using insar and machine learning models. Geosci. Front. 15, 101890. doi:10.1016/j.gsf.2024.101890

CrossRef Full Text | Google Scholar

Yang, L., Zhang, L., Philippopoulos-Mihalopoulos, A., Chappin, E., and van Dam, K. V. (2020). Integrating agent-based modeling, serious gaming, and co-design for planning transport infrastructure and public spaces. Urban Des. Int. 26, 67–81. doi:10.1057/s41289-020-00117-7

CrossRef Full Text | Google Scholar

Yusuff, A., Wiyono, E., and Marpaung, J. V. (2023). Biophilic concept as a public space design. Int. J. art Des. 7, 106–117. doi:10.24191/ijad.v7i1.1100

CrossRef Full Text | Google Scholar

Zhang, Y., Chen, G., He, Y., Jiang, X., and Xue, C. (2022). Social interaction in public spaces and well-being among elderly women: towards age-friendly urban environments. Int. J. Environ. Res. Public Health 19, 746. doi:10.3390/ijerph19020746

CrossRef Full Text | Google Scholar

Zhang, Z., Cheng, S., Wang, C., Song, S., and Feng, Y. (2025). Climate policy uncertainty and corporate investment efficiency: evidence from China. J. Environ. Plan. Manag. 68, 957–977. doi:10.1080/09640568.2023.2276062

CrossRef Full Text | Google Scholar

Zhang, Z., Hua, C., Jiang, M. S., and Miao, J. (2024). The spatial spillover effect of financial growth on high-quality development: evidence from yellow river basin in China. Humanit. Soc. Sci. Commun. 11, 816–817. doi:10.1057/s41599-024-03358-x

CrossRef Full Text | Google Scholar

Zhang, Z., Shi, K., Gao, Y., and Feng, Y. (2023). How does environmental regulation promote green technology innovation in enterprises? a policy simulation approach with an evolutionary game. J. Environ. Plan. Manag. 68, 979–1008. doi:10.1080/09640568.2023.2276064

CrossRef Full Text | Google Scholar

Zhao, K., Qiu, H., Liu, Y., Liu, Z., Huangfu, W., Tang, B., et al. (2025). Probability of rainfall-induced landslides coupled with effective-duration threshold and soil moisture. J. Hydrology Regional Stud. 57, 102112. doi:10.1016/j.ejrh.2024.102112

CrossRef Full Text | Google Scholar

Zhou, L., and Yu, H. (2024). Insar-dlpu: a benchmark dataset for deep learning-based synthetic aperture radar interferometry phase unwrapping [software and data sets]. IEEE Geoscience Remote Sens. Mag. 12, 118–124. doi:10.1109/mgrs.2024.3359691

CrossRef Full Text | Google Scholar