Multisource Flood Risk Assessment for Shenyang Townships Using Dynamic Integration of Analytic Network Process and Autoencoder Weights

Abstract

Township areas are particularly susceptible to flood hazards due to the distribution of water systems and the impact of intense rainfall, which often leads to insufficient flood defense capacity. Consequently, there is an urgent need for scientific flood risk assessment at this scale. This study addresses key limitations in existing research, such as limited characterization of indicator associations, inadequate integration of deep learning with traditional methods, insufficient consideration of township-scale and agricultural characteristics, and restricted use of multisource data. Historical inundation areas from 2014 to 2023 were extracted using MODIS imagery, while key parameters such as the proportion of cultivated land were retrieved from multisource remote sensing data, including Landsat. These data, combined with hydrological and socio-economic information, supported the construction of a township-scale flood risk assessment indicator system. A combined weighting framework, integrating the analytic network process (ANP) and autoencoder, was developed. ANP was employed to quantify network associations among indicators, addressing the limitations of hierarchical structures, while the autoencoder extracted multisource data features to enhance the reliability of objective weights. Dynamic weighting was applied to integrate subjective and objective information within the assessment process. A four-dimensional coupling system was established to construct the assessment model, categorizing risk into five levels. Results demonstrate that the proposed framework outperforms mainstream methods in both balanced risk classification and consistency with historical disaster records: moderate- and high-risk areas encompass 86.84% of historical disaster points, and the recognition rate for historical inundation areas reaches 95.63%. The spatial risk pattern in the study area is characterized by 'river-concentrated aggregation and significant county-level differentiation,' with Xinmin and Liaozhong dominated by moderate to high risk, Kangping exhibiting lower risk, and Faku displaying an interlaced distribution of high and low risk. The findings offer a refined approach to flood risk assessment for townships in Shenyang and provide remote sensing-driven decision support for environmental risk prevention and sustainable township planning.