FragMangro: A cross-domain zero-shot model for monitoring fragmented mangrove ecosystems

Monitoring fragmented mangrove ecosystems presents significant challenges due to their sparse distribution and the limitations of traditional detection methods, which often suffer from poor convergence and high rates of false positives and false negatives. To address these issues, we propose FragMangro, a cross-domain zero-shot detection model designed to enhance the accuracy and generalization of mangrove detection. FragMangro integrates parallel multi-scale convolution for progressive dimensional upscaling, improving spectral feature extraction. Additionally, it employs a dynamic learning rate and iterative control algorithm, which adaptively adjusts learning rates and training epochs based on loss monitoring to prevent suboptimal convergence. We conducted extensive experiments in three major mangrove reserves in Guangxi, China. The results demonstrate that FragMangro achieves over 98% accuracy, an average F1 Score exceeding 85%, and the Intersection over Union(IoU) surpassing 74%, significantly outperforming conventional methods such as support vector machines (SVM), random forests (RF), and k-nearest neighbors (KNN). Specifically, in the Beilun estuary, FragMangro exhibits a 0.96% improvement in accuracy over RF, a 17.24% higher F1 Score than SVM, and an 18.08% higher IoU than KNN. Furthermore, the model’s adaptive learning rate and dynamic adjustment strategy enhance convergence efficiency and feature representation. FragMangro serves as an effective tool for ecological monitoring and mangrove conservation, with broader implications for environmental protection and sustainable ecosystem management.