Improved mangrove refined extraction method based on remote sensing intelligent extraction model

Abstract

Mangrove ecosystems are blue carbon ecosystems with important ecological and environmental significance. Refined monitoring of mangroves is a prerequisite for their management and protection, and remote sensing technology is an indispensable tool for timely and accurate monitoring of mangroves. However, the present methods for extracting mangroves from high-resolution images suffer from problems such as inaccurate boundaries, missing targets, and low accuracy for sparse mangroves. To address these challenges, a novel mangrove extraction method is employed, integrating DeepLabV3+ with the Convolutional Block Attention Module (CBAM). A multi-spectral dataset of mangroves has been established, using a long-time series of multi-source high-resolution images covering many provinces. This comprehensive dataset is subsequently applied to both the baseline model and an improved model for comparative assessment. In comparison with the baseline network, the improved network demonstrates superior performance in mangrove segmentation, exhibiting heightened accuracy, particularly in challenging areas such as the intricate edges and sparsely vegetated regions. The improved model is applied to GF-1 satellite images in the northern Beibu Gulf, yielding segmentation accuracy exceeding 95% when validated against ground truth measurement data. Our proposed methodology significantly contributes to the efficacious management and protection of mangrove ecosystems.