SiamMask-ICDNet: A Hybrid Algorithm of Siamese Neural Network and Mask R-CNN for Island Building Change Detection

Timely detection of construction changes on islands is crucial for environmental conservation, resource management, and sustainable development. Current methodologies typically employ multi-temporal remote sensing imagery combined with artificial visual interpretation, image classification algorithms, and threshold-based techniques to analyze morphological, textural, and spectral variations. However, conventional approaches often suffer from computational inefficiency, weak feature representation, and high false-alarm rates due to oversimplified classification strategies and inadequate feature extraction in complex island environments, rendering them unsuitable for high-precision applications. To address these challenges, this study introduces SiamMask-ICDNet, a novel framework integrating the Siamese neural networks with the Mask R-CNN architecture. The framework enhances feature discriminability through a triple-branch Siamese encoder and incorporates three key innovations: (1) a Focal Loss-based adaptive weighting mechanism to mitigate class imbalance, (2) structural simplification of Mask R-CNN by removing redundant classification and regression branches, retaining only the mask prediction branch for efficient binary change mask generation, and (3) a multi-scale feature fusion module to improve edge preservation in coastal construction areas. Extensive experiments demonstrate that SiamMask-ICDNet achieves state-of-the-art performance, achieving a recall of 0.9779, IoU of 0.9668, F1-score of 0.9830, and precision of 0.9881—surpassing existing methods by $\mathbf{7. 5 - 2 4. 9 \%}$ across metrics. Further analysis confirms its robustness across diverse environmental conditions and improved computational efficiency. These results signify a significant improvement over existing techniques, effectively addressing issues of low accuracy and high false-alarm rates caused by environmental complexity and image quality inconsistencies in island monitoring. This framework provides a reliable approach for large-scale, high-precision change detection in ecologically sensitive coastal areas.