Abstract
Coastal aquaculture ponds represent a significant contributor to economic growth and food provision, underscoring the necessity of precise spatial mapping to support sustainable resource management. Current extraction methods often rely on single-source data and are easily confused by spectral heterogeneity in complex coastal environments, leading to blurred boundaries and misclassification. To overcome these challenges, this study proposes an innovative hybrid model that combines multi-source feature stacking with a hierarchical decision-tree architecture for coarse extraction, followed by an ensemble-learning framework for fine-scale classification. Implemented on the Google Earth Engine cloud platform, the model integrates Sentinel-1 and Sentinel-2 data to leverage complementary spectral, microwave, and terrain features. Applied to the Zhoushan Archipelago in China, the approach produced a high-resolution distribution map of aquaculture ponds with clear boundaries and accurate geolocation. Compared with conventional approaches such as random forest (RF), classification regression trees (CART) and support vector machines (SVM), the proposed model achieved an overall accuracy of 87.34%, improving by 2.55% to 5.39%. The model also achieved a Kappa Coefficient of 73.82% and an F1 score of 89.46%, demonstrating its effectiveness and reliability for automated coastal aquaculture pond extraction in complex coastal environments.