DECDNet: A dual encoder change detection network for monitoring mangrove gain and loss using Sentinel-2 data

Mangrove forests are increasingly threatened by human activities such as aquaculture, agriculture, urban development, and illegal logging. Monitoring these dynamic changes requires accurate and efficient methods. However, traditional change detection approaches typically involve multi-step processes which can be time-consuming and prone to errors. Most existing deep learning models combined with remote sensing have shown great potential for environmental monitoring but are limited to binary classification (change and no change), making it difficult to capture specific land cover transitions such as mangrove gain or loss. To address these limitations, this study introduces DECDNet (Dual Encoder Change Detection Network), a novel deep learning model specifically designed for detecting and mapping mangrove gain and loss using Sentinel-2 imagery. The model utilizes a dual encoder-decoder structure that extracts spatial features from two time points and compares them using a subtraction layer. DECDNet was trained on Sentinel-2 data from 2015 and 2020, incorporating spectral indices to enhance discrimination. As a result, DECDNet achieved superior performance, with an IoU of 0.87, F1 score of 0.93, precision of 0.94, and recall of 0.92. In comparison, the standard deep learning models U-Net and FCN produced IoU values of 0.84 and 0.84, F1 scores of 0.91 and 0.91, precision values of 0.92 and 0.93, and recall values of 0.90 and 0.89, respectively. The generalization capability of DECDNet was further confirmed on a separate 2020-2023 dataset. The model detected 204.22 ha of mangrove loss and 747.09 ha of gain (2015-2020), and 463.48 ha of loss with 48.36 ha of gain (2020-2023) in the Wunbaik Reserved Mangrove Forest. These findings highlight practical implementation of DECDNet as a robust and scalable tool for mangrove monitoring and management.