Coastal Urban Ecological Security Pattern Identification Integrating Land Subsidence Factors: A Deep Learning-Based Case Study of Zhuhai City

Rapid urbanization and geological disasters pose significant challenges to regional ecological security. Although Ecological Security Pattern (ESP) construction is important for ecosystem stability and sustainable development, traditional approaches rarely incorporate vertical geological factors, such as land subsidence. This study proposes a framework that integrates land subsidence into ESP construction through machine learning and multi-source data fusion methods. Using Zhuhai City as a case study, we analyzed 30 environmental variables, including historical land subsidence data, topography, soil distribution, land use, climatic factors, and human activity indicators. The methodology consisted of four main steps: (1) correlation and principal component analyses to identify key factors and reduce dimensionality; (2) development of a multilayer perceptron (MLP) deep learning model with three fully connected hidden layers using ReLU activation functions and dropout regularization to predict ecological pattern types; (3) comparison of four fusion methods (weighted average, nonlinear sigmoid transformation, information entropy, and principal component analysis) to integrate prediction results; and (4) spatial analysis of the relationship between land subsidence and ecological security patterns using chi-square tests and spatial overlay analysis. Results showed that the MLP model achieved an average prediction accuracy of 84.5% with an F1-score of 0.844, demonstrating the feasibility of deep learning approaches in ESP construction. The principal component analysis showed that the first four principal components cumulatively explained 71.4% of the total variance, with the first two components explaining 27.1% and 19.8%. The first principal component was dominated by climatic factors, whereas the second primarily reflected the topographic and geological vulnerability characteristics. Spatial analysis revealed significant spatial heterogeneity in the impact of land subsidence on the ESP, with moderate historical subsidence (8-41 mm/year) showing more notable effects (x²= 57.008, P<0.001). Land subsidence in the 8-16.5 mm/year range showed particularly significant differences in the corridor areas compared to the non-subsidence zones (P = 5.7e-05). Source and construction areas exhibited higher proportions of mild subsidence (7.14% and 9.84%, respectively), which should be prioritized for monitoring and management. Different fusion methods showed varying effectiveness. Principal component analysis and information entropy performed better in identifying construction and corridor areas, whereas nonlinear fusion showed advantages in source area identification. This study makes three key contributions: (1) it establishes a novel methodological framework for incorporating vertical geological factors into ESP construction, addressing a significant gap in traditional approaches; (2) it quantitatively reveals the spatial heterogeneity of land subsidence impacts on different functional ecological zones, providing evidence-based guidance for targeted management; and (3) it demonstrates the effectiveness of deep learning and multisource data fusion techniques in complex ecological-geological system modeling. These findings provide methodological support for developing an ecological security pattern centered on coastal wetlands and estuarine systems in Zhuhai City and suggest potential approaches for coordinating ecological protection, disaster prevention, and urban development under land subsidence conditions. Future research should focus on utilizing high-resolution spatiotemporal data, refining algorithms, and developing mechanisms to translate research findings into practical urban planning and ecological management policies.